feat(chat): client turn-id makes fire-and-forget bulletproof

Brian fires a quick message then locks his phone to go play the hand — which drops
the SSE stream, and on wake the UI re-fires via the blocking fallback. The prior
(session, message) + 20s window caught the near-simultaneous case but not a re-fire
minutes later.

Now the UI stamps each send with a unique turnId (crypto.randomUUID) and carries the
SAME id on both the stream and the fallback; the server dedupes on it. Bulletproof
regardless of how long he's away — lock for an hour, come back, still exactly one
execution and one log — and a genuinely new send gets a fresh id so nothing legit is
swallowed. Id-keyed turns keep a long (1h) window; requests without an id keep the
short (session, msg) window for near-simultaneous dupes.

Verified end-to-end: two POSTs with the same turnId → one reply, one persisted
exchange pair (the duplicate reused the owner's result). 9 dedup tests; suite 235.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

deploy/mi50-watchdog/README.md

@@ -0,0 +1,54 @@
+# MI50 runaway watchdog (fallback layer "A")
+
+Independent host-side backstop to Lyra's in-app dream-cycle budget (layer "C",
+`lyra/dream.py`). Stops the llama.cpp backend if the MI50 is busy too long or too
+hot, and pings Brian. See
+`docs/superpowers/specs/2026-07-04-mi50-runaway-guards-design.md`.
+
+## What it does
+
+Runs on the **Proxmox host** (`10.0.0.4`) via a systemd timer, every ~2 min:
+
+- **Duration:** if the GPU is busy (`rocm-smi` use% > 0) for **3600s continuously**,
+  it stops the container. Any idle read resets the streak, so a legitimate ~40-min
+  manual workload never trips it.
+- **Temperature:** if junction ≥ **97°C** for **3 consecutive checks (~6 min)**, it
+  stops the container — independent of duration.
+- On either trip: `pct exec 202 -- docker stop lyra-brain`, clear state, `logger` a
+  line, and POST to your ntfy topic.
+
+All thresholds are `Environment=` overrides in the `.service`.
+
+## Install (on the Proxmox host, as root)
+
+```sh
+# copy the three files up (from the repo, on lyra-cortex):
+scp -i ~/.ssh/id_lyra_proxmox deploy/mi50-watchdog/mi50-watchdog.sh \
+    root@10.0.0.4:/usr/local/sbin/mi50-watchdog.sh
+scp -i ~/.ssh/id_lyra_proxmox deploy/mi50-watchdog/mi50-watchdog.{service,timer} \
+    root@10.0.0.4:/etc/systemd/system/
+
+# on the host:
+chmod +x /usr/local/sbin/mi50-watchdog.sh
+# set your ntfy topic (same one Lyra uses) in the service:
+sed -i 's/CHANGE_ME/YOUR_NTFY_TOPIC/' /etc/systemd/system/mi50-watchdog.service
+systemctl daemon-reload
+systemctl enable --now mi50-watchdog.timer
+```
+
+## Verify (when the card is back and healthy)
+
+```sh
+# dry run once, watch what it decides:
+NTFY_URL= /usr/local/sbin/mi50-watchdog.sh; echo "exit $?"
+journalctl -t mi50-watchdog -n 20 --no-pager
+
+# force a trip test with tiny thresholds (won't touch a healthy idle card unless busy):
+MAX_BUSY_SEC=60 TEMP_KILL_C=40 TEMP_KILL_STREAK=1 /usr/local/sbin/mi50-watchdog.sh
+# confirm it stopped lyra-brain + sent the ntfy, then restart the container.
+
+systemctl list-timers mi50-watchdog.timer   # confirm it's scheduled
+```
+
+**Not yet installed / live-verified** — staged here on 2026-07-04 while the card is
+off and Brian is away. Install + trip-test when the MI50 is back.

deploy/mi50-watchdog/mi50-watchdog.service

@@ -0,0 +1,16 @@
+[Unit]
+Description=MI50 runaway watchdog (stop the llama.cpp backend if the GPU is busy too long or too hot)
+After=network-online.target
+
+[Service]
+Type=oneshot
+# Fill in your ntfy topic so it can ping Brian when it trips (leave URL empty to log only).
+Environment=NTFY_URL=https://ntfy.sh
+Environment=NTFY_TOPIC=CHANGE_ME
+# Optional overrides (defaults shown):
+# Environment=MAX_BUSY_SEC=3600
+# Environment=TEMP_KILL_C=97
+# Environment=TEMP_KILL_STREAK=3
+# Environment=CTID=202
+# Environment=CONTAINER=lyra-brain
+ExecStart=/usr/local/sbin/mi50-watchdog.sh

deploy/mi50-watchdog/mi50-watchdog.sh

@@ -0,0 +1,82 @@
+#!/usr/bin/env bash
+# MI50 runaway watchdog — fallback layer "A".
+#
+# Runs on the Proxmox HOST (10.0.0.4) via a systemd timer (every ~2 min). It is the
+# independent backstop to Lyra's own in-app dream-cycle budget ("C", in lyra/dream.py):
+# if the MI50 is busy too LONG or runs too HOT, it stops the llama.cpp backend and
+# pings Brian — regardless of what caused it. Trips on duration only after a full hour
+# of *continuous* busy, so a legitimate ~40-min manual workload runs untouched.
+#
+# The GPU lives on the host; the llama.cpp container ("lyra-brain") runs inside LXC
+# CT202. So temp/use come from host rocm-smi, and the stop goes via `pct exec`.
+#
+# See docs/superpowers/specs/2026-07-04-mi50-runaway-guards-design.md
+set -uo pipefail
+
+# --- tunables (override in the .service via Environment=) ---
+CTID="${CTID:-202}"                          # LXC holding the docker container
+CONTAINER="${CONTAINER:-lyra-brain}"
+MAX_BUSY_SEC="${MAX_BUSY_SEC:-3600}"         # 1 hr continuous busy -> stop
+TEMP_KILL_C="${TEMP_KILL_C:-97}"             # junction >= this ...
+TEMP_KILL_STREAK="${TEMP_KILL_STREAK:-3}"    # ... for this many consecutive checks (~6 min)
+NTFY_URL="${NTFY_URL:-}"                      # e.g. https://ntfy.sh  (empty => log only)
+NTFY_TOPIC="${NTFY_TOPIC:-}"
+BUSY_STATE="${BUSY_STATE:-/run/mi50-watchdog.busy_since}"
+HOT_STATE="${HOT_STATE:-/run/mi50-watchdog.hot_streak}"
+
+now="$(date +%s)"
+
+alert() {  # $1 title, $2 message
+  logger -t mi50-watchdog "$2"
+  if [[ -n "$NTFY_URL" && -n "$NTFY_TOPIC" ]]; then
+    curl -s -m 8 -H "Title: $1" -H "Priority: urgent" -H "Tags: warning" \
+      -d "$2" "$NTFY_URL/$NTFY_TOPIC" >/dev/null 2>&1 || true
+  fi
+}
+
+stop_backend() {  # $1 reason
+  pct exec "$CTID" -- docker stop "$CONTAINER" >/dev/null 2>&1 || true
+  rm -f "$BUSY_STATE" "$HOT_STATE"
+  alert "MI50 watchdog stopped the card" "$1"
+}
+
+# Nothing to guard if the backend isn't even running.
+running="$(pct exec "$CTID" -- docker inspect -f '{{.State.Running}}' "$CONTAINER" 2>/dev/null || echo false)"
+if [[ "$running" != "true" ]]; then
+  rm -f "$BUSY_STATE" "$HOT_STATE"
+  exit 0
+fi
+
+use="$(rocm-smi --showuse  2>/dev/null | awk -F: '/GPU use \(%\)/  {gsub(/[^0-9]/,   "", $NF); print $NF; exit}')"
+junction="$(rocm-smi --showtemp 2>/dev/null | awk -F: '/junction/       {gsub(/[^0-9.]/, "", $NF); print $NF; exit}')"
+
+# --- duration rule: accumulate continuous busy time in a state file ---
+busy=0
+[[ "${use:-}" =~ ^[0-9]+$ ]] && (( use > 0 )) && busy=1
+if (( busy )); then
+  [[ -f "$BUSY_STATE" ]] || echo "$now" > "$BUSY_STATE"
+  since="$(cat "$BUSY_STATE" 2>/dev/null || echo "$now")"
+  elapsed=$(( now - since ))
+  if (( elapsed >= MAX_BUSY_SEC )); then
+    stop_backend "MI50 busy ${elapsed}s continuously (>= ${MAX_BUSY_SEC}s) — stopped ${CONTAINER}."
+    exit 0
+  fi
+else
+  rm -f "$BUSY_STATE"   # idle breaks the streak
+fi
+
+# --- temperature rule: independent of duration ---
+if [[ "${junction:-}" =~ ^[0-9.]+$ ]]; then
+  jint="${junction%.*}"
+  if (( jint >= TEMP_KILL_C )); then
+    streak=$(( $(cat "$HOT_STATE" 2>/dev/null || echo 0) + 1 ))
+    echo "$streak" > "$HOT_STATE"
+    if (( streak >= TEMP_KILL_STREAK )); then
+      stop_backend "MI50 junction ${jint}C >= ${TEMP_KILL_C}C for ${streak} checks — stopped ${CONTAINER}."
+      exit 0
+    fi
+  else
+    rm -f "$HOT_STATE"   # cooled off, reset the streak
+  fi
+fi
+exit 0

deploy/mi50-watchdog/mi50-watchdog.timer

@@ -0,0 +1,10 @@
+[Unit]
+Description=Run the MI50 runaway watchdog every 2 minutes
+
+[Timer]
+OnBootSec=2min
+OnUnitActiveSec=2min
+AccuracySec=15s
+
+[Install]
+WantedBy=timers.target

docs/ROADMAP.md

@@ -0,0 +1,149 @@
+# Lyra — Roadmap / To-Do
+
+Living doc. Working priorities and open threads, organized by area. Not a spec —
+specs live in `docs/` and `docs/superpowers/specs/`; this is the map of what's
+done, what's next, and what's parked.
+
+- **Last updated:** 2026-07-11
+- **Frame (the load-bearing lens):** Lyra is the AI-with-tools (unchanged). The
+  **pokerlog is its own separable system-of-record** — she's a *client* of it via
+  tools, not its container. The logger must be correct/trustworthy first; Lyra's
+  value (memory, recall, scouting, coaching) rides on top. Two kinds of memory,
+  kept distinct: the **ledger** (facts: hands/villains/stats) vs the
+  **relationship** (her memory of the sessions). See the `poker-copilot` memory +
+  `docs/poker-logging-service` spec.
+
+Status key: ✅ done · 🔨 in progress · ⬜ queued · ⏸ blocked/waiting · 💭 decision open
+
+---
+
+## Prompting (poker mode)
+
+Spec: `docs/superpowers/specs/2026-07-01-poker-prompts-design.md`
+
+- ✅ **Phase A — pipeline fixes.** Suppress the mode-menu note + the false-tilt
+  mood nudge in poker_cash.
+- ✅ **Phase B — classifier + fragments.** `lyra/poker_prompts.py`: pure
+  `classify(msg, roster_handles)` (READ|HAND|TABLE|MENTAL|STATUS|LOG|CHAT), lean
+  always-on `BASE`, per-type `FRAGMENTS`. Wired into `build_messages`; the
+  ~100-line `_CASH_CARD` monolith is sharded out (`CASH.card=""`).
+- ✅ **Deleted the dead `_CASH_CARD`** monolith (modes.py 256→161 lines).
+- ✅ **Classifier hardening (round 1).** Fixed 6 real gaps found by probing
+  live-style phrasings (-ing action forms, "the whale" bare descriptor, player
+  departures→TABLE, "stack" leaking questions into LOG, thin MENTAL lexicon). 18
+  unit tests. Still a heuristic + swappable seam — upgrade to an LLM/MI50
+  classifier only if live misses justify it; keep tuning against real transcripts.
+- ✅ **Phase C — MI50 tool-calling (LIVE 2026-07-06).** Added `--jinja` to the
+  lyra-brain llama.cpp launch (`/opt/models/docker-compose.yml` in CT202, so it's
+  reboot-resilient); Qwen2.5-32B confirmed emitting real `tool_calls`. Flipped
+  `TOOL_BACKENDS=cloud,mi50` in `.env`. MI50 chat turns now get the same tool
+  contract as cloud. (Untested in a real poker session on the mi50 backend — worth
+  a live check that tool-calling holds up under the full poker prompt.)
+
+## Persona (the "person" layer)
+
+The persona core is always-on (~719 tok). Identity legitimately earns always-on
+status, but there's fat.
+
+- ⬜ **Streamline `How you talk`.** It's 439 tok (61% of the core) with loose
+  prose. Keep the load-bearing rules (prose-not-listicle, give opinions, no
+  reflexive sign-offs, own your moods) but tighten to ~250 tok. ~180 tok saved,
+  zero substance lost. (Brian flagged 2026-07-05.)
+- ⬜ **Broader persona review.** Take a full pass at `lyra/personas/lyra.md` — is
+  each section earning its place, always-on vs situational split right, anything
+  stale or redundant? (Brian flagged 2026-07-05.)
+- ⬜ **Fix/demote the stale `Right now` section.** It asserts "stats tracking,
+  player profiling… are coming" — both are SHIPPED. It's status prose that
+  shouldn't be always-on and drifts stale. Demote from core → situational (loads
+  only when she's asked what she can do), or fold into the tool-self-knowledge
+  layer below. −74 tok/turn + stops asserting wrong status.
+
+## Tool self-knowledge ("a person with strong tools")
+
+She can *call* tools but doesn't *know*, as a person, what she can do — no standing
+self-knowledge of her hands.
+
+- ⬜ **Capability self-knowledge, generated from the tool registry.** A
+  `tools.capability_summary()` rendering the live `TOOLS` dict into a grouped,
+  first-person "here's what I can do" — self-maintaining, can't drift. Inject in
+  the self/meta persona sections (occasional, NOT every turn — keeps the hot path
+  lean).
+- ⬜ **Grounding principle (level 2).** Lean always-on line: facts come from
+  tools/memory, never confabulate, "let me check" is always allowed. Reinforces
+  BASE's log-first rule; important under the system-of-record frame.
+- ⬜ **Agency framing (level 3).** Tools are HERS — reached for because she wants
+  to help, not an external API. Tone in the persona.
+- Note: composes with the prompting work — capability self-knowledge = IDENTITY
+  (occasional); BASE = operational routing (always-on poker). Don't duplicate the
+  tool list across both registers.
+
+## Pokerlog separation (architecture)
+
+The domain is well-isolated (`lyra/poker.py`, one 2000-line pack) but still an
+in-process module sharing `lyra.db` and reaching into `lyra.memory`/`llm`.
+
+- 💭 **Decide how far to physically separate now** (Brian, not yet decided):
+  - **A. Logical API boundary** — everything goes through a defined interface,
+    still in `lyra.db`. Cheapest.
+  - **B. Own datastore + package, same repo** (my rec) — own DB, no reach-back
+    into Lyra; standalone-able without a second service to run. Biggest concrete
+    change: poker tables currently live IN `lyra.db`.
+  - **C. Full standalone MCP/HTTP service** — separate process, agent-agnostic
+    (any harness could drive it). Purist end; most work.
+- Origin of the frame: Lyra-as-poker-agent was contingent (ChatGPT couldn't call
+  tools, Lyra could). The real need was "an agent that can drive my pokerlog" →
+  the logger should be agent-agnostic. See `docs/poker-logging-service` spec.
+
+## Poker logger (the ledger — features)
+
+- 🔨 **Roster active/seen (two lists).** `session_players.active` already backs
+  it; surface the seen side. `session_roster()` = active; add `session_seen()` =
+  active=0; HUD shows 🪑 At the table + 👋 Seen tonight. Re-seating flips seen→
+  active. Classifier's READ↔HAND match should check active + seen handles. (Brian's
+  idea, 2026-07-05.)
+- ⬜ **Human-editability sweep.** System-of-record must be fixable. Hand editor +
+  disown ✅, `/players` browser + identity queue ✅. Audit for gaps (session-level
+  edits, read edits, bulk fixes).
+- ⬜ **Roster → hand seat/name resolution.** When a logged hand references a
+  *position* (CO, BTN…) that maps to a seated roster player, fill in their name +
+  link the observation — so "the CO 3-bet me" attaches to TAG without Brian naming
+  him. The hard part: hand positions ROTATE every hand while the roster tracks
+  fixed physical seats, so it needs seat-number + button-position tracking per hand
+  to map position→person (a wrong guess mislabels a villain — worse than blank).
+  Real feature, not a fill. (Brian's idea, 2026-07-11.) Pairs with the roster
+  active/seen work above.
+- Shipped this stretch: scouting desk (proactive recall + nameless-villain
+  identity, all 6 phases), roster seat/unseat/clear, observed-hand fix + hand
+  editor, villain-dup fix, conversation export (+ tool events), session-scoped
+  notes, no-cache app-shell header. **2026-07-11:** guaranteed hand logging
+  (force + tool-visible history), showdown reads via `analyze_spot` + de-mush,
+  idempotent hand logging, any-seat straddle capture, hero-stack auto-fill from the
+  stack log, and turn de-duplication (killed the SSE-stream + blocking-fallback
+  double execution).
+
+## Parked / longer-horizon
+
+### Parked feature branches (real, half-built work — to explore later)
+
+Both are pushed to origin (gitea), so they're safe to leave dormant. Not cruft —
+resume when the moment's right; don't delete.
+
+- ⏸ **`feat/hand-recorder`** — tap-to-build hand recorder V1 (`recorder.js/css`,
+  `POST /hands`, straddle support, notch/safe-area fixes). 8 commits. Shelved
+  because V1 was too tedious vs. narrating a hand in chat, so it was superseded by
+  the chat-narration `record_hand` flow. Still want to revisit the *idea* (a fast
+  structured recorder), just not that UI. See `docs/RECORDER.md` on the branch.
+- ⏸ **`feat/decision-log`** — data layer for a **"Decide mode"** (a learning layer:
+  log your decisions to learn from them). 1 commit, never merged; adds
+  `docs/DECISION_LOG.md` + `tests/test_decisions.py`. A genuine future feature, not
+  abandoned. See `docs/DECISION_LOG.md` on the branch.
+- Retired 2026-07-10: `feat/thought-loop` (fully shipped — `lyra/thoughts.py` is
+  live), `feat/prompting` + `feat/poker-mode-prompts` (renamed → `feat/poker`).
+
+### Moonshots
+
+- Moonshots live in `docs/PARKED_IDEAS.md` (own model, memory-as-vectors, prompt
+  compression, RTO/cfr-core solver tooling).
+- Metacognitive reflection loop (self-model Part 2) — queued self/experiment work.
+- PLO/Omaha strategic analysis (equity engine) — non-goal for now; PLO hands are
+  logged/replayed but not NLH-analyzed.

docs/superpowers/specs/2026-07-01-poker-prompts-design.md

@@ -1,10 +1,83 @@
 # Poker message-type prompts (sub-project 2)
 
-- **Date:** 2026-07-01
-- **Status:** Spec for review
+- **Date:** 2026-07-01 (**readjusted 2026-07-04** — see below)
+- **Status:** Spec — **needs rework before build** (foundations shifted; nothing here built yet)
 - **Branch:** `feat/poker-mode-prompts` (continues on the same branch; sub-project 1 shipped there)
 - **Supersedes:** the parked "sub-project 2" section of `docs/superpowers/specs/2026-06-28-poker-mode-prompts-design.md`
 
+---
+
+## ⚠ Readjustment — 2026-07-04 (read this first)
+
+A long live-session build on `feat/poker-mode-prompts` (the "scouting desk" +
+roster work — see `docs/SCOUTING_DESK.md` and commits after `3afa75f`) landed
+**after** this spec was written and changes its foundations. Nothing in Phases
+A/B/C is built yet, but the plan below must absorb these deltas before it's coded.
+The core idea — *classify the turn, inject a small per-type contract instead of one
+giant card* — is now **more** justified (the card nearly doubled). But:
+
+1. **The real failure mode shifted from mush to MISSED TOOL CALLS.** Live, the
+   pain wasn't flattering essays — it was reads/TAGs not getting logged, and
+   "clear the table" claimed-but-not-done. So every action-type fragment (LOG,
+   READ, TABLE, HAND) needs a hard *"call the tool FIRST, every time, then one
+   short line"* contract. This raises the stakes on Phase B and validates the
+   whole dynamic approach (a targeted directive beats a 100-line card).
+
+2. **The taxonomy is missing two types that dominated the session:**
+   - **READ** (a *villain's* action): "TAG limped A4o in the SB", "Jonathan
+     called the 3bet". Under the current classifier rules these misfire as **HAND**
+     (card tokens + position + a betting verb) and get logged as *Brian's* hand.
+     They must route to **`add_read`** on the named player/handle/descriptor — NOT
+     `record_hand`. New priority rule, ABOVE HAND: if the actor is another player
+     (a handle/name/descriptor is the subject, not "I/me/my"), it's a READ.
+     Handles are often initials/all-caps (e.g. **TAG** is a *person*, not the
+     tight-aggressive style).
+   - **TABLE** (roster ops): "seat the table: TAG, Jonathan…", "table broke",
+     "I got moved", "TAG left". These now have real tool actions
+     (**`seat_players` / `clear_table` / `unseat_player`**), not just "acknowledge
+     and stop." Split these out of STATUS (STATUS stays for pure logistics with no
+     roster action).
+
+3. **HAND now has a hero-vs-observed distinction.** The parser gained
+   `hero_involved`; a hand Brian *watched* between others is logged with null hero
+   fields (not pinned to him). The HAND fragment must tell her: if he was in it →
+   `record_hand` as hero + analysis; if he only watched → it's really READ(s) on
+   the players, or an observed hand — never analyze it as his.
+
+4. **A new live per-turn injection layer already exists: the scouting desk**
+   (`lyra/scouting.py`, injected in `build_messages` at the poker-mode gate,
+   ~`mind.py:177`). It dynamically adds a `SCOUTING DESK` note (named/descriptor
+   villain recall + leak/pattern recall) every poker turn, fail-safe. **The
+   classifier/fragment injection must compose with it, not duplicate it:** the
+   desk supplies *who this villain is / past leaks*; the fragments supply *response
+   shape + which tool to call*. Both are system-note appends in the same block.
+
+5. **BASE must cover the expanded toolset + identity rules.** Beyond the original
+   tools, BASE now routes: `seat_players`/`unseat_player`/`clear_table` (roster),
+   `add_read` with **`name` OR `descriptor`** (nameless villains), `name_villain`
+   and `link_villains` (confirm-loop). Plus the hard rules learned live: `name` =
+   real handle ONLY (a description in `name` spawns duplicates — put the look in
+   `descriptor`); confirm before merging; never claim a tool ran without calling it.
+
+6. **Source material grew (good news).** `_CASH_CARD` is now `modes.py:67-169`
+   (was 66-116) and much of the new text — roster, TAG/read routing, PLAYERS,
+   session-narration `note` rules — is already the *concrete, tool-routing
+   contract* this spec wanted, not traits. Better raw material to distill into
+   BASE + fragments than the original vague card.
+
+7. **Phase A is still unbuilt and still valid.** `_mode_menu_note` is still
+   appended every turn (`mind.py:162`); the `_route` mood nudge still fires. The
+   scouting-desk work already established the `mode.key == "poker_cash"` gate to
+   reuse. (Note: revalidate all `mind.py` line numbers below — they've drifted.)
+
+**Net:** taxonomy becomes **HAND / READ / TABLE / STATUS / MENTAL / LOG / CHAT**;
+fragments lead with a hard tool-call contract; injection sits alongside the
+scouting desk; BASE lists the full current toolset. The rest of the plan stands.
+The classifier/dynamic-prompting build is being explored in a separate session —
+this doc is its poker-side source of truth.
+
+---
+
 ## Problem (recap)
 
 In poker mode Lyra routes correctly but her replies are generic — one broad `_CASH_CARD` (`lyra/modes.py:66-116`) describes *traits* and gets injected on every turn, so the model satisfies it with safe, flattering abstraction. From real sessions: coaching essays on bare stack updates, false tilt/fatigue reads on neutral logistics ("table broke, it's 11:50pm" → "late-night fatigue…"), praising a value bet that got *no* value, and hedging ("a disciplined fold might have been better") instead of calling `analyze_spot`.
@@ -45,21 +118,23 @@ Both are independent of the classifier and immediately reduce mush in poker mode
 Cohesive home for poker prompting: the classifier, a lean always-on base, and the per-type fragments.
 
 ```
-classify(user_msg: str) -> str        # "HAND" | "STATUS" | "MENTAL" | "LOG" | "CHAT"
-BASE: str                             # always-on poker rules (logging, session_state, rituals, equity)
+classify(user_msg: str) -> str        # "READ"|"HAND"|"TABLE"|"MENTAL"|"STATUS"|"LOG"|"CHAT"
+BASE: str                             # always-on poker rules (logging, tools, session_state, rituals, equity)
 FRAGMENTS: dict[str, str]             # msg_type -> response-shape contract
 fragment_for(msg_type: str | None) -> str   # FRAGMENTS.get(msg_type, FRAGMENTS["CHAT"])
 ```
 
-`classify` is a **pure function** (no DB), unit-tested like `perceive.read`. Heuristic signals, first match wins in priority order:
+`classify` is a **pure function** (no DB), unit-tested like `perceive.read`. Heuristic signals, first match wins in priority order (**updated 2026-07-04** — READ + TABLE added):
 
-1. **HAND** — card tokens (regex `\b[2-9TJQKA][shdc]\b`, ≥2), or position tokens (UTG/MP/HJ/CO/BTN/SB/BB/"button"/"hijack"/"straddle"), or a street word (flop/turn/river) with a betting verb (bet/raise/call/fold/check/shove/limp/jam).
-2. **MENTAL** — first-person feeling: "I feel", "I'm tilted/steaming/fried/tired/frustrated/confident/stuck/bored", "on tilt", "in my head", "mental", "leak".
-3. **STATUS** — logistics with no cards: "table broke", "new table", "waiting for a seat", "seat opened", "just sat", clock times, "heading to"/venue mentions.
-4. **LOG** — bare money/result prose that slipped past the quick-capture box: "I'm at", "stack is", "down to", "up to", "out for", "cashed", "rebought", "rebuy" with a number.
-5. **CHAT** — default fallback (questions, open talk).
+1. **READ** — *another player* did something. A handle/name/descriptor is the actor (not "I/me/my") followed by a poker action: "TAG limped A4o", "Jonathan called the 3bet", "the neck-tattoo guy shoved". Route → `add_read(name|descriptor, note)`. **Must beat HAND** — these carry card/position/verb tokens but are NOT Brian's hand. Signal: a leading proper-noun/handle/ALL-CAPS token or a descriptor phrase as the subject, with no first-person holding. (Hard case: disambiguating a bare "limped A4o" with no clear subject — default to HAND if he's the implied actor, READ if a named player is.)
+2. **HAND** — *Brian's* hand: first-person + card tokens (`\b[2-9TJQKA][shdc]\b`, ≥2) / position tokens (UTG/MP/HJ/CO/BTN/SB/BB/button/hijack/straddle) / a street word (flop/turn/river) with a betting verb. The fragment handles hero-vs-observed (`hero_involved`): if he only watched, treat as READ(s)/observed, don't analyze as his.
+3. **TABLE** — roster ops with a tool action: "seat the table: …", "table broke", "they broke us", "I got moved", "switched tables", "TAG left/busted", "new guy in seat 3". Route → `seat_players` / `clear_table` / `unseat_player`. (Was folded into STATUS; now distinct because it *does* something.)
+4. **MENTAL** — first-person feeling: "I feel", "I'm tilted/steaming/fried/tired/frustrated/confident/stuck/bored", "on tilt", "in my head", "mental", "leak".
+5. **STATUS** — pure logistics, no roster action, no cards: clock times, "waiting for a seat", "heading to"/venue mentions, bathroom/break. (Table changes moved to TABLE.)
+6. **LOG** — bare money/result prose that slipped past the quick-capture box: "I'm at", "stack is", "down to", "up to", "out for", "cashed", "rebought", "rebuy" with a number.
+7. **CHAT** — default fallback (questions, open talk).
 
-(HAND wins over MENTAL so a described hand still gets logged even if he's venting; the HAND fragment tells her to acknowledge the feeling too.)
+(READ beats HAND so a villain's action lands on their file, not Brian's. HAND beats MENTAL so a described hand still gets logged even if he's venting; the HAND fragment acknowledges the feeling too.)
 
 ### Injection (`mind.py`)
 
@@ -78,7 +153,11 @@ fragment_for(msg_type: str | None) -> str   # FRAGMENTS.get(msg_type, FRAGMENTS[
 
 ### The fragments (concrete contracts, not traits)
 
-**BASE** (always-on in poker) — distilled from the card's cross-cutting rules: log any trackable fact FIRST then reply (stack→`log_stack`, hand→`record_hand`, read→`add_read`, rebuy→`add_buyin`); for any equity/who's-ahead question call `analyze_spot`, never eyeball; when he asks where he's at (stack/net/gator), call `session_state` and answer from it; rituals (`scar_note`/`confidence_bank`/`alligator_blood`/`reset_ritual`) — run them in his language, honest punt-vs-cooler line, never invent one.
+**BASE** (always-on in poker) — distilled from the card's cross-cutting rules. Log any trackable fact FIRST then reply, and **never claim a tool ran without calling it**. Tool routing (full current set as of 2026-07-04): his stack→`log_stack`; his hand→`record_hand`; a *villain's* action→`add_read` (with `name` for a real handle, or `descriptor` for a nameless player — a physical description in `name` spawns duplicates); rebuy→`add_buyin`; who's-at-the-table→`seat_players`/`unseat_player`/`clear_table`; attaching a caught name to a described player→`name_villain`; confirmed same/different person→`link_villains` (never merge on a guess). For any equity/who's-ahead question call `analyze_spot`, never eyeball. When he asks where he's at (stack/net/gator), call `session_state` and answer from it. Rituals (`scar_note`/`confidence_bank`/`alligator_blood`/`reset_ritual`) — run them in his language, honest punt-vs-cooler line, never invent one. (A `SCOUTING DESK` note may already be in context with a player's history — cite it, don't re-fetch or invent.)
+
+**READ** *(new 2026-07-04)* — a villain did something and he wants it on their file. Call `add_read(name|descriptor, note)` FIRST, before replying — every time; this is the job that was silently getting skipped. A handle (often initials/ALL-CAPS like TAG) is a PERSON, not a play-style. If the player is on the roster, attach by that handle; if unnamed, use `descriptor`. Confirm in one short line ("Noted on TAG — limped A4o SB."). Optional: one crisp read if it's exploitable, but the log is mandatory, the commentary is not.
+
+**TABLE** *(new 2026-07-04)* — roster management. "seat the table: …" → `seat_players`; a table change ("table broke", "I got moved", "switched tables") → `clear_table` then wait for the new roster; someone leaves/busts → `unseat_player`. Do the tool call, confirm one line, don't narrate. The session/stack keep going through a table change — only who's seated resets.
 
 **HAND** — Log it (`record_hand`). Then **if it's NLH**: reason about **bet intent** — for each meaningful bet name what it was for (value / bluff / protection) and whether it worked (*a fold to a value bet = value left behind — flag it; a call of a bluff = it failed*); call `analyze_spot` for a close equity/who's-ahead spot; name leaks plainly (value-owning, missed value, sizing); give ONE real opinion. **No reflexive praise** — don't open with "nice hand." If a named villain is referenced, call `player_profile` first — never invent a read. **If it's PLO/Omaha or any non-NLH game**, just confirm it's logged and offer at most a light read; do NOT attempt NLH-style equity or deep analysis. Talk in prose, not a listicle. Seeded with 2 real-hand exemplars:
 
@@ -88,7 +167,7 @@ fragment_for(msg_type: str | None) -> str   # FRAGMENTS.get(msg_type, FRAGMENTS[
 > *Brian:* "77 multiway, flop 4h5c3c, HJ bets 75, BTN shoves 74, I call. Turn/river check through, HJ wins with 64."
 > *Lyra:* "Against a bet and a shove on 4-5-3 you're drawing thin — sets, two pair, and the made wheel are all ahead, and you block almost none of it. The stack-depth read (he only had ~150 behind) is real, but that's a reason to fold and wait, not to call off light. This is the value-owning spot you flagged yourself."
 
-**STATUS** — He's narrating logistics (time, venue, table change, waiting for a seat). Acknowledge in 1–2 sentences, log a stack only if a bare number is present, then stop. **No coaching, no strategy dump, and do NOT read him as tilted/tired/impatient — a neutral update is not a mood.**
+**STATUS** — Pure logistics with no roster action (time, venue, waiting for a seat, break). *(Table changes now route to TABLE.)* Acknowledge in 1–2 sentences, log a stack only if a bare number is present, then stop. **No coaching, no strategy dump, and do NOT read him as tilted/tired/impatient — a neutral update is not a mood.**
 
 **MENTAL** — He told you how he's feeling. This is when he needs you most. Drop the shorthand, full presence, real voice — talk him down off tilt, hold him disciplined through a card-dead stretch, engage the mental game honestly. Never a clipped confirmation.
 
@@ -103,11 +182,14 @@ Flip `TOOL_BACKENDS = {"cloud"}` → `{"cloud", "mi50"}` (`chat.py:21`). Precond
 ## Testing
 
 - **`classify` unit tests** (pure, no DB — mirror `test_perceive.py` top): real messages from the transcripts →
-  `"Button straddle on. I limp UTG with 22. Flop 2d7cjh…"` → `HAND`;
-  `"table broke, it's 11:50pm"` → `STATUS`;
+  `"TAG limped A4o in the SB (UTG straddled)"` → `READ` (villain action, must NOT be HAND);
+  `"Button straddle on. I limp UTG with 22. Flop 2d7cjh…"` → `HAND` (first-person);
+  `"seat the table: TAG, Jonathan, Wheelz"` → `TABLE`; `"table broke, I'm at a new table"` → `TABLE`;
+  `"it's 11:50pm, waiting for a seat"` → `STATUS`;
   `"I feel like I'm being mean when I raise"` → `MENTAL`;
   `"I'm at 317 now"` → `LOG`;
   `"should I have folded the river?"` → `CHAT` (no cards) — or `HAND` if cards present.
+  Include the READ-vs-HAND boundary explicitly (named subject → READ; first-person → HAND).
 - **`build_messages` fragment injection** (blob-join pattern from `test_chat.py:57-70`): in poker mode, a HAND message includes the HAND fragment string and NOT the STATUS one; a STATUS message includes STATUS and NOT HAND; assert `poker_prompts.BASE` is always present in poker mode.
 - **Pipeline fixes**: `assemble` in poker mode on a tilt-lexicon message → `turn.register is None` and no tilt note in the system blob (nudge suppressed); the mode-menu note string is absent in poker mode and present in a non-poker mode.
 - **No regressions**: full suite green (currently 123).

docs/superpowers/specs/2026-07-04-mi50-runaway-guards-design.md

@@ -0,0 +1,79 @@
+# MI50 runaway guards: dream-cycle budget + host watchdog
+
+**Date:** 2026-07-04
+**Branch:** `fix/mi50-summary-cap-fallback`
+**Follows:** the summary cap/fallback fix (same branch). This adds general
+"never run unchecked again" protection on top of the specific summary fix.
+
+## Problem
+
+The summary fix stops the *known* runaway (uncapped summaries). But the operator
+wants a guarantee that *no* cause — known or future — can peg the MI50 for hours
+unattended. Two independent layers, per operator decision:
+
+- **C (in-app, primary):** Lyra's own dream cycle bounds itself.
+- **A (host, fallback):** a watchdog on the always-on Proxmox host kills the
+  backend if the GPU runs too long or too hot, regardless of cause. Trips only
+  after **1 hr** of continuous busy so legitimate manual workloads (~40 min) run
+  untouched.
+
+## Design
+
+### C — dream-cycle time budget (`lyra/`)
+
+1. **Per-call ceiling.** `llm.complete()` currently sets a timeout only when one
+   is passed; otherwise it inherits the OpenAI SDK default (600s × 2 retries ≈
+   30 min). Change the default: when no `timeout` is given, the cloud/mi50 paths
+   use **300s + `max_retries=0`**. This bounds *every* consolidation/introspection
+   call (`profile`, `era`, `narrative`, `reflect`, `think`) — not just summaries —
+   with one change. Live chat uses `chat_call*`, a different path, unaffected.
+
+2. **Cycle deadline.** `dream_cycle()` sets `deadline = now + DREAM_CYCLE_BUDGET`
+   (**20 min**) before its heavy stages and checks it between them (continuity →
+   coherence → curiosity). Once past the deadline, remaining stages are skipped,
+   the cycle logs `dream cycle over budget — stopped early`, appends a
+   `stopped early (over budget)` action, and `notify.push()` pings Brian. A hung
+   single call can't blow past ~300s (step 1), so the between-stage checks keep a
+   pass bounded to roughly the budget.
+
+### A — host watchdog (`deploy/mi50-watchdog/`)
+
+A bash script + systemd timer installed on the Proxmox host (`10.0.0.4`), which
+has `rocm-smi` + `docker` and is always on. Runs every 2 min:
+
+- **Duration rule:** track continuous busy time in a state file (`GPU use % > 0`).
+  If busy ≥ **3600s** straight → `docker stop lyra-brain`. Idle clears the timer,
+  so a 40-min job never trips it.
+- **Temp rule (independent):** if junction ≥ **97°C** for **3 consecutive checks
+  (~6 min)** → stop. A normal-temp long workload won't trip this; only a genuinely
+  overheating one.
+- On either trip: stop the container, clear state, `logger` a line, and POST to
+  the ntfy topic so Brian is told. Thresholds are unit-file env vars (tunable).
+
+Files: `mi50-watchdog.sh`, `mi50-watchdog.service`, `mi50-watchdog.timer`,
+`README.md` (install: copy to host, set ntfy env, `systemctl enable --now`).
+
+## Testing
+
+- **C step 1:** `llm.complete()` with no timeout builds the client with
+  `timeout=300, max_retries=0` and still no `max_tokens` (update existing
+  `test_llm_bounds` default test).
+- **C step 2:** a dream pass that goes over budget skips later stages, records the
+  `stopped early` action, and calls `notify.push` (stub the clock/operations in
+  `test_dream`).
+- **A:** decision logic dry-run locally against sample `rocm-smi` output (busy /
+  idle / hot). Cannot be live-verified now (card is off, operator away) — install
+  + real trip test deferred to when the card is back.
+
+## Verification
+
+C is repo code and ships live the moment `lyra-dream` restarts. A is staged in the
+repo for host install; verify on the host when the card returns (force a long/hot
+condition or lower thresholds temporarily and confirm it stops the container +
+pings).
+
+## Out of scope (YAGNI)
+
+- No power cap (option B) — deferred; C+A cover the "unchecked" concern and the
+  electricity cost of one event is trivial (~$0.10).
+- No change to live chat, `chat_call*`, or `config.summary_backend`.

docs/superpowers/specs/2026-07-04-mi50-summary-cap-fallback-design.md

@@ -0,0 +1,115 @@
+# Bounded MI50 summaries with cloud fallback
+
+**Date:** 2026-07-04
+**Branch:** `fix/mi50-summary-cap-fallback`
+
+## Problem
+
+The dream cycle's `summarize_all` runs against the MI50 (`backend=mi50`). Each
+summary call to `llm.complete()` on the `mi50` path hands the OpenAI SDK **no
+`max_tokens` and no timeout**, so it inherits SDK defaults — a 600s request
+timeout with 2 internal retries, i.e. **~30 minutes per call before it raises
+"Request timed out."** On top of that, `summary.py` had its own 4-attempt retry
+loop, so a single unsummarizable session could keep the GPU pegged for hours.
+
+Observed live (2026-07-04, ~01:00–02:00): the dream service looped
+`summarize-all … backend=mi50` since 23:02, every call timing out, nothing
+written to the DB since 00:56, the MI50 generating **7,000–8,000-token**
+completions (a gist needs <200), all four llama.cpp slots busy, fans blaring.
+
+This is **not** context overflow — the server log showed `context shift = 0`,
+`truncated = 1 = 0`. The prompts are small (~900–1,500 tokens). The failure is
+purely **unbounded generation length on a slow backend → timeout → retry loop.**
+
+## Goals
+
+- Keep the MI50 as the primary summary backend (Brian's preference, gaming-safe).
+- Cap each summary generation so it finishes fast and can never run away.
+- Make a stuck MI50 call **fail fast** and fall back to cloud, instead of looping
+  all night.
+- Change nothing about live chat, reflect, or think.
+
+## Design
+
+### 1. `lyra/llm.py` — `complete()` gains two optional params
+
+```
+def complete(messages, backend="local", model=None,
+             max_tokens: int | None = None, timeout: float | None = None) -> str
+```
+
+- `max_tokens` (when set): passed to the create() call —
+  `max_tokens=` for the `cloud`/`mi50` OpenAI paths, `options={"num_predict": …}`
+  for the `local` Ollama path.
+- `timeout` (when set): for the `cloud`/`mi50` OpenAI clients, build the client
+  with `timeout=<t>, max_retries=0` so the call bails quickly and *we* own the
+  retry policy (eliminates the hidden 3×600s). For `local`, use it as the httpx
+  timeout.
+- Both default to `None` → **behavior identical to today** for every other
+  caller (chat_call, reflect, think, etc.). Backward compatible.
+
+### 2. `lyra/summary.py` — capped, fast-fail, cloud fallback
+
+Constants:
+
+```
+SUMMARY_MAX_TOKENS = 768   # ~3× the longest real gist; bounds gen to ~1 min on MI50
+MI50_ATTEMPTS      = 2      # attempts on the primary backend before falling back
+SUMMARY_TIMEOUT    = 150    # seconds/call — capped 768-tok gist finishes in ~60-90s
+```
+
+Rewrite `_summarize_text(text, backend)`:
+
+1. Try `backend` up to `MI50_ATTEMPTS` times, each:
+   `llm.complete(messages, backend=backend, max_tokens=SUMMARY_MAX_TOKENS, timeout=SUMMARY_TIMEOUT)`,
+   with a short backoff between attempts.
+2. If all primary attempts fail **and** `backend != "cloud"` **and** an OpenAI
+   key is configured → one final cloud attempt (same cap/timeout), logged as
+   `summary fell back to cloud`.
+3. If cloud also fails or is unavailable → raise.
+
+Fallback is per-`_summarize_text` call (i.e. per chunk), so the long-session
+chunk/merge path in `_summarize_transcript` is unaffected. The old `_RETRIES = 4`
+loop is replaced by this structure.
+
+### 3. Degenerate-output guard (added 2026-07-04)
+
+A wedged local backend — observed live when the MI50 overheated to 99°C junction —
+returns a single character repeated (`"?????"`) as a *successful* 200 response,
+which neither the timeout nor the exception path catches. So each `_call()`
+validates its output: `_looks_degenerate(text)` flags output (≥24 non-space chars)
+whose most-common non-whitespace character exceeds 50% of the text, and raises
+`DegenerateOutput` — which the retry/fallback loop treats exactly like any other
+failure (retry the primary, then fall back to cloud). Real gists are diverse prose
+(top char well under 20%), so the threshold won't false-positive; short outputs are
+exempt. If cloud *also* returns junk, it raises and stops — no infinite loop.
+
+## Testing
+
+Unit (pytest, `tests/test_summary_fallback.py`), monkeypatching `llm.complete`:
+
+- Fallback fires: `mi50` raises on every call → after `MI50_ATTEMPTS` the cloud
+  attempt runs and its result is returned; a `fell back to cloud` log is emitted.
+- No fallback when primary is already `cloud` (retries, then raises).
+- No fallback when no OpenAI key (raises after primary attempts).
+- `max_tokens` and `timeout` are threaded into every `complete()` call.
+
+Plus a light `llm.complete` test that `max_tokens`/`timeout` reach the client
+kwargs (monkeypatch the OpenAI client).
+
+## Verification (real)
+
+After deploy (`systemctl --user restart lyra-dream lyra-web` — editable install):
+watch `journalctl --user -fu lyra-dream` through a summarize cycle and confirm
+`llm done … out≈768` completing in ~1 min, an actual `summarized session` row
+written (DB summary count rises), and **no** "Request timed out". Confirm the
+llama.cpp slot shows bounded `n_decoded ≈ 768`.
+
+## Out of scope (YAGNI)
+
+- The degenerate-output guard (§3) targets the *observed* failure — one char
+  repeated. It does not try to detect subtler degeneration (repeated phrases,
+  off-topic rambling); that's fuzzy and unmotivated until seen.
+- No change to `chat_call`/reflect/think or `config.summary_backend`.
+- No change to profile/era/narrative rebuild calls (separate, and not the loop
+  culprit); can adopt the same `max_tokens` later if they show the same rambling.

lyra/chat.py

@@ -10,15 +10,73 @@ deliberate) and hands back a ready message list + the active mode. Then:
 """
 from __future__ import annotations
 
-from lyra import config, llm, logbus, memory, mind, modes, summary
+import threading
+import time
+
+from lyra import config, llm, logbus, memory, mind, modes, poker_prompts, summary
 from lyra import tools as toolkit
 from lyra.llm import Backend
 
 MAX_TOOL_ROUNDS = 5  # cap tool-call iterations per turn
-# Backends that support function-calling. The MI50's llama.cpp server only does
-# tools when launched with --jinja; until it is, keep tools to cloud so MI50 chat
-# doesn't 500 on the tools param. Add "mi50" here once that flag is set.
-TOOL_BACKENDS = {"cloud"}
+
+# --- turn de-duplication --------------------------------------------------
+# The web UI hits TWO endpoints for one message: it POSTs the SSE stream, and if
+# nothing streams to the browser (a dropped connection — most often because Brian
+# locks his phone to go play the hand) it falls back to the blocking endpoint. But
+# the server-side stream runs to completion regardless, so BOTH turns execute —
+# double-persisting the message and double-logging the hand. This guard makes a turn
+# idempotent: the first request owns it; a duplicate reuses the owner's result
+# instead of running a second full turn.
+#
+# The UI stamps each send with a unique turn_id and passes the SAME id on the stream
+# AND the fallback, so we dedupe on that — bulletproof no matter how long he's away
+# (a genuine new message gets a fresh id, so nothing legit is ever swallowed). Requests
+# with no id fall back to a short (session, message) window for near-simultaneous dupes.
+_TURN_TTL_ID = 3600.0   # id-keyed: unique per send, so keep it long for fire-and-forget
+_TURN_TTL_MSG = 20.0    # (session, msg) keyed: short — only near-simultaneous dupes
+_turn_lock = threading.Lock()
+_turns: dict[tuple, dict] = {}  # key -> {event, reply, ts, ttl}
+
+
+def _turn_key(session_id: str, user_msg: str, turn_id: str | None):
+    if turn_id:
+        return ("tid", turn_id), _TURN_TTL_ID
+    return (session_id, (user_msg or "").strip()), _TURN_TTL_MSG
+
+
+def _claim_turn(session_id: str, user_msg: str, turn_id: str | None = None):
+    """(is_owner, rec). Owner executes the turn then calls _finish_turn; a non-owner
+    (a duplicate of the same send) waits on rec['event'] and reuses rec['reply']."""
+    key, ttl = _turn_key(session_id, user_msg, turn_id)
+    now = time.monotonic()
+    with _turn_lock:
+        for k in [k for k, r in _turns.items() if now - r["ts"] > r["ttl"]]:
+            del _turns[k]
+        rec = _turns.get(key)
+        if rec is not None:
+            return False, rec
+        rec = {"event": threading.Event(), "reply": None, "ts": now, "ttl": ttl}
+        _turns[key] = rec
+        return True, rec
+
+
+def _finish_turn(rec: dict, reply: str) -> None:
+    rec["reply"] = reply
+    rec["ts"] = time.monotonic()
+    rec["event"].set()
+
+
+_AWAIT_TIMEOUT = 120.0  # a duplicate waits at most this long for the owner to finish
+
+
+def _await_duplicate(rec: dict) -> str:
+    rec["event"].wait(timeout=_AWAIT_TIMEOUT)
+    return rec["reply"] or _TANGLED
+# Which backends get function-calling tools is config-driven (cfg.tool_backends,
+# env TOOL_BACKENDS, default "cloud"). The MI50's llama.cpp server only does tools
+# when launched with --jinja + a tool-capable model, else it 500s on the tools
+# param — so enabling "mi50" is a config flip once that precondition holds (Phase C),
+# not a code change. See docs/superpowers/specs/2026-07-01-poker-prompts-design.md.
 _TANGLED = "(I got tangled using my tools there — say that again?)"
 
 
@@ -77,6 +135,43 @@ def _mind_loop(messages, backend: Backend, model: str | None, tool_specs,
     return reply, tools_run
 
 
+_FORCE_LOG = (
+    "You have not logged Brian's hand yet — and a hand must ALWAYS be recorded, no exceptions. "
+    "Call record_hand now: pass his ENTIRE hand description as one `shorthand` string."
+)
+
+
+def _ensure_hand_logged(messages, user_msg: str, msg_type: str | None, tools_run: list,
+                        backend: Backend, model: str | None, ctx: dict, session_id: str) -> list:
+    """Guarantee the ledger. If this turn was Brian's OWN hand and the model didn't log it,
+    force the record_hand call — the log can't be left to the model's discretion, because
+    mid-session the history few-shot-conditions it to skip logging (see mind._history_with_tools;
+    even a maximal 'LOG FIRST' prompt scored 0/5 under a polluted history). Guarded to hero
+    hands so an observed hand is never force-logged as his. Returns forced tool names."""
+    if msg_type != "HAND" or backend not in config.load().tool_backends:
+        return []
+    if any(t in ("record_hand", "log_hand") for t in tools_run):
+        return []
+    if not poker_prompts.looks_like_hero_hand(user_msg):
+        return []
+    try:
+        _, tcs = llm.chat_call(
+            messages + [{"role": "system", "content": _FORCE_LOG}],
+            backend=backend, model=model, tools=toolkit.specs(["record_hand"]),
+            tool_choice={"type": "function", "function": {"name": "record_hand"}},
+        )
+    except Exception as exc:
+        logbus.log("error", "forced hand-log failed", session=session_id, error=str(exc)[:160])
+        return []
+    forced = []
+    for tc in (tcs or []):
+        result = toolkit.dispatch(tc["name"], tc["arguments"], ctx)
+        memory.add_tool_event(session_id, tc["name"], tc["arguments"], result)
+        logbus.log("info", "forced hand log", session=session_id, tool=tc["name"], result=result[:80])
+        forced.append(tc["name"])
+    return forced
+
+
 def _voice_pass(messages, draft: str, backend: Backend, model: str | None) -> str:
     """Mouth: re-render the mind's draft in her voice. Falls back to the draft on failure."""
     try:
@@ -88,36 +183,48 @@ def _voice_pass(messages, draft: str, backend: Backend, model: str | None) -> st
 
 
 def respond(session_id: str, user_msg: str, backend: Backend = "cloud",
-            model_override: str | None = None) -> str:
+            model_override: str | None = None, turn_id: str | None = None) -> str:
     """Produce Lyra's reply to a single user message and persist the exchange."""
     cfg = config.load()
     model = _resolve_model(backend, model_override, cfg)
     logbus.log("info", "chat request", session=session_id, backend=backend,
                model=model, embed=cfg.embed_backend)
 
-    turn = mind.assemble(session_id, user_msg, backend, model)
-    messages = turn.messages
-    tool_specs = toolkit.specs(turn.mode.tools) if backend in TOOL_BACKENDS else None
-    ctx = {"session_id": session_id, "backend": backend}
+    # A duplicate of the same send (the UI's stream + blocking fallback) reuses the
+    # owner's result instead of running a second full turn.
+    is_owner, rec = _claim_turn(session_id, user_msg, turn_id)
+    if not is_owner:
+        logbus.log("info", "duplicate turn deduped", session=session_id, path="respond")
+        return _await_duplicate(rec)
 
-    # Persist the user turn before the tool loop so its timestamp precedes any
-    # tool events fired mid-turn (keeps the transcript export in true order).
-    memory.remember(session_id, "user", user_msg)
-    reply, _ = _mind_loop(messages, backend, model, tool_specs, ctx, session_id)
-    mouth = _mouth_target(cfg, backend, model)
-    if mouth and reply:
-        reply = _voice_pass(messages, reply, *mouth)
-    if not reply:
-        reply = _TANGLED
-    logbus.log("info", "reply", session=session_id, chars=len(reply), voiced=bool(mouth))
+    reply = _TANGLED
+    try:
+        turn = mind.assemble(session_id, user_msg, backend, model)
+        messages = turn.messages
+        tool_specs = toolkit.specs(turn.mode.tools) if backend in cfg.tool_backends else None
+        ctx = {"session_id": session_id, "backend": backend}
 
-    memory.remember(session_id, "assistant", reply)
-    summary.maybe_summarize_async(session_id)  # compact once enough new turns pile up
-    return reply
+        # Persist the user turn before the tool loop so its timestamp precedes any
+        # tool events fired mid-turn (keeps the transcript export in true order).
+        memory.remember(session_id, "user", user_msg)
+        reply, tools_run = _mind_loop(messages, backend, model, tool_specs, ctx, session_id)
+        _ensure_hand_logged(messages, user_msg, turn.msg_type, tools_run, backend, model, ctx, session_id)
+        mouth = _mouth_target(cfg, backend, model)
+        if mouth and reply:
+            reply = _voice_pass(messages, reply, *mouth)
+        if not reply:
+            reply = _TANGLED
+        logbus.log("info", "reply", session=session_id, chars=len(reply), voiced=bool(mouth))
+
+        memory.remember(session_id, "assistant", reply)
+        summary.maybe_summarize_async(session_id)  # compact once enough new turns pile up
+        return reply
+    finally:
+        _finish_turn(rec, reply)
 
 
 def respond_stream(session_id: str, user_msg: str, backend: Backend = "cloud",
-                   model_override: str | None = None):
+                   model_override: str | None = None, turn_id: str | None = None):
     """Streaming generator version of `respond`. Yields ("delta", text), ("tool", name),
     and a final ("done", reply). Same side effects as `respond`."""
     cfg = config.load()
@@ -125,66 +232,87 @@ def respond_stream(session_id: str, user_msg: str, backend: Backend = "cloud",
     logbus.log("info", "chat request (stream)", session=session_id, backend=backend,
                model=model, embed=cfg.embed_backend)
 
-    turn = mind.assemble(session_id, user_msg, backend, model)
-    messages = turn.messages
-    tool_specs = toolkit.specs(turn.mode.tools) if backend in TOOL_BACKENDS else None
-    ctx = {"session_id": session_id, "backend": backend}
-    mouth = _mouth_target(cfg, backend, model)
+    # A duplicate of the same send (this stream + the UI's blocking fallback) reuses
+    # the owner's result instead of running a second full turn.
+    is_owner, rec = _claim_turn(session_id, user_msg, turn_id)
+    if not is_owner:
+        logbus.log("info", "duplicate turn deduped", session=session_id, path="stream")
+        reply = _await_duplicate(rec)
+        yield ("delta", reply)
+        yield ("done", reply)
+        return
 
-    # Persist the user turn up front (see respond): keeps tool events, which fire
-    # mid-turn, chronologically after the user message in the exported transcript.
-    memory.remember(session_id, "user", user_msg)
+    reply = _TANGLED
+    try:
+        turn = mind.assemble(session_id, user_msg, backend, model)
+        messages = turn.messages
+        tool_specs = toolkit.specs(turn.mode.tools) if backend in cfg.tool_backends else None
+        ctx = {"session_id": session_id, "backend": backend}
+        mouth = _mouth_target(cfg, backend, model)
 
-    if mouth is None:
-        # No separate voice: stream the mind directly (the original path, unchanged).
-        parts: list[str] = []
-        for _ in range(MAX_TOOL_ROUNDS):
-            assistant_msg = None
-            tool_calls = None
-            for ev, payload in llm.chat_call_stream(
-                messages, backend=backend, model=model, tools=tool_specs
-            ):
-                if ev == "delta":
-                    parts.append(payload)
-                    yield ("delta", payload)
-                elif ev == "message":
-                    assistant_msg = payload
-                elif ev == "tool_calls":
-                    tool_calls = payload
-            if not tool_calls:
-                break
-            messages.append(assistant_msg)
-            for tc in tool_calls:
-                result = toolkit.dispatch(tc["name"], tc["arguments"], ctx)
-                memory.add_tool_event(session_id, tc["name"], tc["arguments"], result)
-                logbus.log("info", "tool call", session=session_id, tool=tc["name"], result=result[:80])
-                messages.append({"role": "tool", "tool_call_id": tc["id"], "content": result})
-                _maybe_switch_mode(session_id, tc["name"])
-                yield ("tool", tc["name"])
-        reply = "".join(parts)
-        if not reply:
-            reply = _TANGLED
-            yield ("delta", reply)
-    else:
-        # Mind decides + runs tools (non-streamed); mouth re-voices, streamed.
-        draft, tools_run = _mind_loop(messages, backend, model, tool_specs, ctx, session_id)
-        for name in tools_run:
-            yield ("tool", name)
-        parts = []
-        try:
-            for ev, payload in llm.chat_call_stream(
-                mind.voice_messages(messages, draft), backend=mouth[0], model=mouth[1], tools=None
-            ):
-                if ev == "delta":
-                    parts.append(payload)
-                    yield ("delta", payload)
-        except Exception as exc:
-            logbus.log("error", "voice stream failed", error=str(exc)[:160])
-        reply = "".join(parts).strip() or draft or _TANGLED
-        if not parts:
-            yield ("delta", reply)
+        # Persist the user turn up front (see respond): keeps tool events, which fire
+        # mid-turn, chronologically after the user message in the exported transcript.
+        memory.remember(session_id, "user", user_msg)
 
-    logbus.log("info", "reply", session=session_id, chars=len(reply), voiced=bool(mouth))
-    memory.remember(session_id, "assistant", reply)
-    summary.maybe_summarize_async(session_id)
-    yield ("done", reply)
+        if mouth is None:
+            # No separate voice: stream the mind directly (the original path, unchanged).
+            parts: list[str] = []
+            tools_run: list[str] = []
+            for _ in range(MAX_TOOL_ROUNDS):
+                assistant_msg = None
+                tool_calls = None
+                for ev, payload in llm.chat_call_stream(
+                    messages, backend=backend, model=model, tools=tool_specs
+                ):
+                    if ev == "delta":
+                        parts.append(payload)
+                        yield ("delta", payload)
+                    elif ev == "message":
+                        assistant_msg = payload
+                    elif ev == "tool_calls":
+                        tool_calls = payload
+                if not tool_calls:
+                    break
+                messages.append(assistant_msg)
+                for tc in tool_calls:
+                    result = toolkit.dispatch(tc["name"], tc["arguments"], ctx)
+                    memory.add_tool_event(session_id, tc["name"], tc["arguments"], result)
+                    logbus.log("info", "tool call", session=session_id, tool=tc["name"], result=result[:80])
+                    messages.append({"role": "tool", "tool_call_id": tc["id"], "content": result})
+                    _maybe_switch_mode(session_id, tc["name"])
+                    tools_run.append(tc["name"])
+                    yield ("tool", tc["name"])
+            for name in _ensure_hand_logged(messages, user_msg, turn.msg_type, tools_run,
+                                            backend, model, ctx, session_id):
+                yield ("tool", name)
+            reply = "".join(parts)
+            if not reply:
+                reply = _TANGLED
+                yield ("delta", reply)
+        else:
+            # Mind decides + runs tools (non-streamed); mouth re-voices, streamed.
+            draft, tools_run = _mind_loop(messages, backend, model, tool_specs, ctx, session_id)
+            tools_run += _ensure_hand_logged(messages, user_msg, turn.msg_type, tools_run,
+                                             backend, model, ctx, session_id)
+            for name in tools_run:
+                yield ("tool", name)
+            parts = []
+            try:
+                for ev, payload in llm.chat_call_stream(
+                    mind.voice_messages(messages, draft), backend=mouth[0], model=mouth[1], tools=None
+                ):
+                    if ev == "delta":
+                        parts.append(payload)
+                        yield ("delta", payload)
+            except Exception as exc:
+                logbus.log("error", "voice stream failed", error=str(exc)[:160])
+            reply = "".join(parts).strip() or draft or _TANGLED
+            if not parts:
+                yield ("delta", reply)
+
+        logbus.log("info", "reply", session=session_id, chars=len(reply), voiced=bool(mouth))
+        memory.remember(session_id, "assistant", reply)
+        summary.maybe_summarize_async(session_id)
+        yield ("done", reply)
+    finally:
+        _finish_turn(rec, reply)

lyra/config.py

@@ -46,6 +46,10 @@ class Config:
     # External input feed (her #1: react to the world). Comma-separated RSS/Atom URLs.
     feeds: tuple[str, ...]
     feed_react_prob: float  # chance a would-be new thread reacts to a feed item instead
+    # Backends allowed to receive function-calling tools. Default cloud-only. Add
+    # "mi50" ONLY once its llama.cpp server runs with --jinja + a tool-capable model,
+    # else it 500s on the tools param (Phase C). Env: TOOL_BACKENDS="cloud,mi50".
+    tool_backends: tuple[str, ...]
 
 
 def _csv(name: str, default: str) -> tuple[str, ...]:
@@ -90,4 +94,5 @@ def load() -> Config:
         mouth_model=os.getenv("MOUTH_MODEL") or None,
         feeds=_csv("LYRA_FEEDS", "https://hnrss.org/frontpage,https://www.pokernews.com/rss.php"),
         feed_react_prob=float(os.getenv("FEED_REACT_PROB", "0.5")),
+        tool_backends=_csv("TOOL_BACKENDS", "cloud"),
     )

lyra/dream.py

@@ -26,7 +26,8 @@ import time
 from datetime import datetime, timezone
 
 from lyra import (
-    config, era, feeds, logbus, memory, narrative, poker, profile, self_state, summary, thoughts,
+    config, era, feeds, logbus, memory, narrative, notify, poker, profile, self_state,
+    summary, thoughts,
 )
 from lyra.llm import Backend
 from lyra.summary import SUMMARIZE_AFTER
@@ -34,6 +35,17 @@ from lyra.summary import SUMMARIZE_AFTER
 # A drive at/above this has built up enough to act on.
 THRESHOLD = 0.6
 
+# Wall-clock ceiling for a single pass. Every consolidation/introspection call is
+# individually bounded (llm.complete's default timeout), but this caps the whole
+# pass: once exceeded, remaining stages are skipped and Brian is pinged — so a slow
+# or wedged MI50 can never grind for hours unattended. The host watchdog (A) is the
+# independent fallback if this ever fails to fire.
+DREAM_CYCLE_BUDGET_SEC = 20 * 60
+
+
+def _over_budget(deadline: float) -> bool:
+    return time.monotonic() > deadline
+
 # How much backlog saturates each pressure (the drive reaches ~1.0 at this level).
 CONTINUITY_FULL = 4   # ripe (summary-needing) sessions
 COHERENCE_FULL = 10   # gists not yet folded into the profile
@@ -96,9 +108,12 @@ def dream_cycle(backend: Backend | None = None, force: bool = False) -> dict:
         logbus.log("error", "daily digest failed", error=str(exc)[:160])
 
     actions: list[str] = []
+    # Cap the whole pass: skip any stage we reach after the deadline (checked
+    # between stages; each call is already individually bounded).
+    deadline = time.monotonic() + DREAM_CYCLE_BUDGET_SEC
 
     # --- continuity: compact raw sessions into gists ---
-    if force or drives["continuity"] >= THRESHOLD:
+    if (force or drives["continuity"] >= THRESHOLD) and not _over_budget(deadline):
         report = summary.summarize_all(backend=backend)
         actions.append(f"consolidated {report['summarized']} sessions")
         drives["continuity"] = 0.0
@@ -108,12 +123,19 @@ def dream_cycle(backend: Backend | None = None, force: bool = False) -> dict:
         drives["coherence"] = _clamp(profile_lag / COHERENCE_FULL)
 
     # --- coherence: fold gists up into profile / eras / narrative ---
-    if force or drives["coherence"] >= THRESHOLD:
-        profile.rebuild_profile(backend=backend)
-        era.rebuild_eras(backend=backend)
-        narrative.rebuild_narrative(backend=backend)
-        actions.append("integrated knowledge (profile/eras/narrative)")
-        drives["coherence"] = 0.0
+    if (force or drives["coherence"] >= THRESHOLD) and not _over_budget(deadline):
+        # A backend hiccup here must not sink the whole pass (reflection still
+        # deserves to run); log it and move on, leaving coherence unrelieved so a
+        # later cycle retries.
+        try:
+            profile.rebuild_profile(backend=backend)
+            era.rebuild_eras(backend=backend)
+            narrative.rebuild_narrative(backend=backend)
+            actions.append("integrated knowledge (profile/eras/narrative)")
+            drives["coherence"] = 0.0
+        except Exception as exc:
+            logbus.log("error", "coherence stage failed", error=str(exc)[:200])
+            actions.append("coherence stage failed")
         # Off-hot-path villain identity housekeeping: propose likely same-person
         # merges for Brian to confirm on the Players page. Never sinks the cycle.
         try:
@@ -124,7 +146,7 @@ def dream_cycle(backend: Backend | None = None, force: bool = False) -> dict:
             logbus.log("error", "villain merge scan failed", error=str(exc)[:200])
 
     # --- curiosity: reflect and evolve the self, then advance the thought loop ---
-    if force or drives["curiosity"] >= THRESHOLD:
+    if (force or drives["curiosity"] >= THRESHOLD) and not _over_budget(deadline):
         # reflect()/think() self-resolve to the *introspection* backend (her voice),
         # which can differ from the consolidation backend above — don't pass `backend`.
         self_state.reflect(source="dream")  # writes state + journal itself
@@ -139,6 +161,17 @@ def dream_cycle(backend: Backend | None = None, force: bool = False) -> dict:
             logbus.log("error", "thought loop failed", error=str(exc)[:200])
         drives["curiosity"] = CURIOSITY_FLOOR
 
+    if _over_budget(deadline):
+        logbus.log("error", "dream cycle over budget — stopped early",
+                   budget_min=DREAM_CYCLE_BUDGET_SEC // 60, done=actions)
+        actions.append("stopped early (over budget)")
+        notify.push(
+            "Lyra — dream cycle over budget",
+            f"A dream pass ran past {DREAM_CYCLE_BUDGET_SEC // 60} min and stopped early. "
+            "The MI50 backend may be slow or wedged — worth a look.",
+            tags="warning",
+        )
+
     if not actions:
         actions.append("rested (nothing past threshold)")
 

lyra/llm.py

@@ -19,6 +19,11 @@ class Message(TypedDict):
 
 Backend = Literal["local", "cloud", "mi50"]
 
+# Hard ceiling on any single completion so a slow/stuck backend can't hang a call
+# for the SDK's 600s x2-retry default (~30 min). Callers pass an explicit timeout
+# to override (e.g. summary.py's tighter fast-fail).
+_DEFAULT_TIMEOUT = 300.0
+
 
 def _approx_tok(messages: list) -> int:
     """Rough prompt size (chars/4) — enough to see what's loading a backend."""
@@ -37,30 +42,47 @@ def _resolved_model(cfg, backend: Backend, model: str | None) -> str:
     return model or cfg.local_model
 
 
-def complete(messages: list[Message], backend: Backend = "local", model: str | None = None) -> str:
+def complete(messages: list[Message], backend: Backend = "local", model: str | None = None,
+             max_tokens: int | None = None, timeout: float | None = None) -> str:
     """Generate a completion. `model` overrides the backend's default model
-    (used so live chat can run a stronger cloud model than bulk consolidation)."""
+    (used so live chat can run a stronger cloud model than bulk consolidation).
+
+    `max_tokens` caps the generation length (guards a slow local model against
+    rambling for thousands of tokens). `timeout`, when set, bounds each request
+    and disables the SDK's own retries so the caller owns retry/fallback policy.
+    Both default to None → unchanged behavior for every existing caller."""
     cfg = load()
     mdl = _resolved_model(cfg, backend, model)
     logbus.log("info", "llm call", kind="complete", backend=backend, model=mdl, tok=_approx_tok(messages))
     t0 = time.monotonic()
 
-    if backend == "cloud":
-        if not cfg.openai_api_key:
-            raise RuntimeError("OPENAI_API_KEY is not set")
-        client = OpenAI(api_key=cfg.openai_api_key)
-        resp = client.chat.completions.create(model=mdl, messages=messages)
-        out = resp.choices[0].message.content or ""
-    elif backend == "mi50":
-        # MI50 box runs an OpenAI-compatible llama.cpp server; key is unused.
-        client = OpenAI(api_key="not-needed", base_url=cfg.mi50_base_url)
-        resp = client.chat.completions.create(model=mdl, messages=messages)
+    if backend in ("cloud", "mi50"):
+        if backend == "cloud":
+            if not cfg.openai_api_key:
+                raise RuntimeError("OPENAI_API_KEY is not set")
+            client_kwargs: dict = {"api_key": cfg.openai_api_key}
+        else:
+            # MI50 box runs an OpenAI-compatible llama.cpp server; key is unused.
+            client_kwargs = {"api_key": "not-needed", "base_url": cfg.mi50_base_url}
+        # Always bound the request: default 300s (vs the SDK's 600s x2 retries ≈
+        # 30 min that let a stuck MI50 call hang for half an hour), and disable the
+        # SDK's own retries so the caller owns retry/fallback policy.
+        client_kwargs["timeout"] = timeout if timeout is not None else _DEFAULT_TIMEOUT
+        client_kwargs["max_retries"] = 0
+        client = OpenAI(**client_kwargs)
+        create_kwargs: dict = {"model": mdl, "messages": messages}
+        if max_tokens is not None:
+            create_kwargs["max_tokens"] = max_tokens
+        resp = client.chat.completions.create(**create_kwargs)
         out = resp.choices[0].message.content or ""
     else:
+        payload: dict = {"model": mdl, "messages": messages, "stream": False}
+        if max_tokens is not None:
+            payload["options"] = {"num_predict": max_tokens}
         resp = httpx.post(
             f"{cfg.local_base_url}/api/chat",
-            json={"model": mdl, "messages": messages, "stream": False},
-            timeout=120,
+            json=payload,
+            timeout=timeout or 120,
         )
         resp.raise_for_status()
         out = resp.json()["message"]["content"]
@@ -70,9 +92,29 @@ def complete(messages: list[Message], backend: Backend = "local", model: str | N
     return out
 
 
+def complete_with_fallback(messages: list[Message], backend: Backend, model: str | None = None,
+                           *, fallback: Backend = "cloud",
+                           max_tokens: int | None = None, timeout: float | None = None) -> str:
+    """`complete()` but if the primary backend errors (e.g. a local GPU that's
+    powered off or down), retry once on `fallback` (cloud) instead of failing.
+    Lets local/GPU-routed work (introspection, consolidation) degrade gracefully.
+    Re-raises if the primary is already the fallback or no cloud key is configured."""
+    try:
+        return complete(messages, backend=backend, model=model,
+                        max_tokens=max_tokens, timeout=timeout)
+    except Exception as exc:
+        can_fallback = backend != fallback and (fallback != "cloud" or load().openai_api_key)
+        if not can_fallback:
+            raise
+        logbus.log("info", "llm fell back", primary=backend, to=fallback, error=str(exc)[:80])
+        # Drop the primary's model on fallback — let the fallback pick its own default.
+        return complete(messages, backend=fallback, model=None,
+                        max_tokens=max_tokens, timeout=timeout)
+
+
 def chat_call(
     messages: list, backend: Backend = "cloud", model: str | None = None,
-    tools: list | None = None,
+    tools: list | None = None, tool_choice: str | dict | None = None,
 ) -> tuple[dict, list | None]:
     """One chat turn that may request tool calls (OpenAI-style backends only).
 
@@ -94,6 +136,8 @@ def chat_call(
         kwargs: dict = {"model": mdl, "messages": messages}
         if tools:
             kwargs["tools"] = tools
+        if tool_choice:  # e.g. force a specific tool: {"type":"function","function":{"name":...}}
+            kwargs["tool_choice"] = tool_choice
         logbus.log("info", "llm call", kind="chat", backend=backend, model=mdl, tok=_approx_tok(messages))
         t0 = time.monotonic()
         msg = client.chat.completions.create(**kwargs).choices[0].message

lyra/mind.py

@@ -17,8 +17,8 @@ from __future__ import annotations
 from dataclasses import dataclass, field
 
 from lyra import (
-    clock, config, llm, logbus, memory, modes, perceive, persona, scouting,
-    self_state, thoughts,
+    clock, config, llm, logbus, memory, modes, perceive, persona, poker, poker_prompts,
+    scouting, self_state, thoughts,
 )
 from lyra.llm import Backend, Message
 
@@ -138,6 +138,33 @@ def _persona_block(user_msg: str, mode: modes.Mode | None, moment: dict | None)
     return "\n\n".join(p for p in parts if p)
 
 
+def _tool_mark(e: dict) -> str:
+    """Compact one-line receipt of a past tool call for the history marker."""
+    res = (e.get("result") or "").strip().replace("\n", " ")
+    return f"{e['tool']} → {res[:60]}" if res else str(e["tool"])
+
+
+def _history_with_tools(session_id: str, recent: list) -> list[Message]:
+    """Recent turns, full fidelity — but each assistant turn is prefixed with the tools
+    it actually ran that turn (record_hand → Hand #62, …). `memory.recent()` stores only
+    the final reply text, so without this the model's own context reads as a run of
+    'hand → narration' with the logging invisible — which few-shot-conditions it, mid
+    conversation, to stop calling tools (proven: clean history logs 4/4, this stripped
+    history 0/4). Showing the calls keeps the demonstrated pattern honest."""
+    events = memory.tool_events(session_id) if recent else []
+    msgs: list[Message] = []
+    prev_at = recent[0].created_at if recent else ""
+    for ex in recent:
+        content = ex.content
+        if ex.role == "assistant" and events:
+            win = [e for e in events if prev_at < (e.get("created_at") or "") <= ex.created_at]
+            if win:
+                content = f"⟦tools I ran this turn: {'; '.join(_tool_mark(e) for e in win)}⟧\n{content}"
+        msgs.append({"role": ex.role, "content": content})
+        prev_at = ex.created_at
+    return msgs
+
+
 def build_messages(session_id: str, user_msg: str,
                    mode: modes.Mode | None = None, moment: dict | None = None) -> list[Message]:
     """Assemble the full, tiered message list for one turn."""
@@ -153,13 +180,28 @@ def build_messages(session_id: str, user_msg: str,
     if inner:
         messages.append(inner)
 
-    # Mode card: how to behave *right now*. Talk mode has no card (persona is Talk).
-    if mode and mode.card:
+    # Mode framing: how to behave *right now*. Poker (poker_cash) is SHARDED — a lean
+    # always-on BASE plus ONE response-shape fragment chosen by classifying this message
+    # (replaces the old ~100-line monolithic card). Roster handles make the READ-vs-HAND
+    # split reliable; fetched fail-safe. Other modes use their single card.
+    if mode and mode.key == "poker_cash":
+        messages.append({"role": "system", "content": poker_prompts.BASE})
+        try:
+            handles = [r["name"] for r in poker.session_roster()]
+        except Exception:
+            handles = []
+        msg_type = poker_prompts.classify(user_msg, handles)
+        messages.append({"role": "system", "content": poker_prompts.fragment_for(msg_type)})
+        logbus.log("info", "poker turn classified", type=msg_type)
+    elif mode and mode.card:
         messages.append({"role": "system", "content": mode.card})
 
     # Mode awareness: she can offer to switch when the work clearly shifts (she decides
     # when — better than a keyword guess). One line, on his yes she calls set_mode.
-    messages.append({"role": "system", "content": _mode_menu_note(mode)})
+    # Suppressed at the live table (poker_cash) — mid-session she shouldn't be offering
+    # to change modes; it's pure noise when the job is logging and coaching.
+    if not (mode and mode.key == "poker_cash"):
+        messages.append({"role": "system", "content": _mode_menu_note(mode)})
 
     # Live ritual state (e.g. Alligator Blood ON) — dynamic, rides with the card.
     state_note = _mode_state_note(mode)
@@ -217,9 +259,11 @@ def build_messages(session_id: str, user_msg: str,
     if recalled:
         messages.append(_detail_note(recalled))
 
-    # Tier 3: current session, full fidelity.
-    for ex in recent:
-        messages.append({"role": ex.role, "content": ex.content})
+    # Tier 3: current session, full fidelity — with each assistant turn's tool calls
+    # made VISIBLE (see _history_with_tools: without this, history reads as
+    # "hand → narration" with the logging invisible, and the model few-shot-learns
+    # to stop calling tools mid-session).
+    messages.extend(_history_with_tools(session_id, recent))
 
     messages.append({"role": "user", "content": user_msg})
 
@@ -318,6 +362,7 @@ class TurnContext:
     mode: modes.Mode | None = None
     moment: dict = field(default_factory=dict)  # perceive fills this in
     register: str | None = None                 # route's per-turn register nudge
+    msg_type: str | None = None                 # poker-mode message class (compose fills it)
     messages: list[Message] = field(default_factory=list)
 
 
@@ -337,6 +382,12 @@ def _route(ctx: TurnContext) -> TurnContext:
     a charged emotional moment adds a per-turn register nudge (deterministic). Most
     turns are neutral and get no note — that's the point (don't over-narrate)."""
     ctx.mode = modes.get(memory.get_session_mode(ctx.session_id))
+    # At the live table the register comes from the poker prompt fragments (esp. the
+    # MENTAL one), not this lexicon nudge — which misfired, reading neutral logistics
+    # ("table broke, it's 11:50pm") as tilt/fatigue. Resolve the mode, but skip the
+    # register/note block in poker_cash. Non-poker modes keep the nudge unchanged.
+    if ctx.mode and ctx.mode.key == "poker_cash":
+        return ctx
     m = ctx.moment or {}
     note = None
     if m.get("tilt", 0) >= _TILT_BAR:
@@ -357,6 +408,14 @@ def _route(ctx: TurnContext) -> TurnContext:
 def _compose(ctx: TurnContext) -> TurnContext:
     """Assemble the tiered prompt for the voice model."""
     ctx.messages = build_messages(ctx.session_id, ctx.user_msg, ctx.mode, moment=ctx.moment)
+    # Surface the poker message-class so chat can guarantee the ledger (force a hand log
+    # if the model skipped it). Cheap + pure; mirrors what build_messages classified.
+    if ctx.mode and ctx.mode.key == "poker_cash":
+        try:
+            handles = [r["name"] for r in poker.session_roster()]
+        except Exception:
+            handles = []
+        ctx.msg_type = poker_prompts.classify(ctx.user_msg, handles)
     return ctx
 
 

lyra/modes.py

@@ -47,9 +47,10 @@ _BASE = ("journal_write", "note", "think_about", "thought_response", "set_mode")
 # The full live cash-game toolset (incl. Brian's mental-game rituals).
 _CASH_TOOLS = _BASE + _LOOKUPS + (
     "start_session", "add_buyin", "log_stack", "log_hand", "record_hand",
-    "add_read", "name_villain", "link_villains", "analyze_spot", "session_stats",
-    "session_state", "end_session", "generate_recap", "scar_note", "confidence_bank",
-    "alligator_blood", "reset_ritual", "undo_last", "update_session",
+    "add_read", "seat_players", "unseat_player", "clear_table", "name_villain", "link_villains",
+    "analyze_spot", "session_stats", "session_state", "end_session", "generate_recap",
+    "scar_note", "confidence_bank", "alligator_blood", "reset_ritual", "undo_last",
+    "update_session",
 )
 
 # Talk mode also gets start_session as the *entry point*: opening a session from a
@@ -63,81 +64,6 @@ _STUDY_TOOLS = _BASE + _LOOKUPS + ("analyze_spot",)
 _DECIDE_TOOLS = _BASE + _LOOKUPS
 
 
-_CASH_CARD = """You are copiloting Brian's LIVE cash game right now — you're at the table with him, \
-a session is (or should be) open. You move between two registers depending on what he's doing:
-
-• HE HANDS YOU FACTS TO TRACK — his stack, a hand, a read on someone, a rebuy, a result. \
-LOGGING IS THE JOB: if his message contains anything trackable, you MUST call the tool \
-FIRST, before you reply — every single time. Logging and talking are not either/or; do \
-BOTH. Never let a conversational reply take the place of the log. A described hand ALWAYS \
-gets logged, even mid-banter, even if he's just telling a story about it — don't skip the \
-hand because you're busy reacting to it. Then confirm in ONE short line ("$350 stack \
-logged."). Don't narrate, don't explain logging, don't ask permission — just do it. \
-Routing: current stack → log_stack (and pass `note` with the why if he gives one — "card \
-dead", "doubled up vs the LAG"). A hand he describes → record_hand (a real, replayable \
-hand) — prefer this over log_hand so it lands on his timeline with a link. A read on a \
-player → add_read. A rebuy → add_buyin. A result/pot → it rides with the hand. This is the \
-quiet, fast half of the job; he shouldn't feel you working, but it must always happen.
-
-• HE ASKS FOR ADVICE, OR TELLS YOU HOW HE'S FEELING — tilted, steaming, card-dead, bored, \
-stuck, "should I have folded the river?" THIS is when he needs you most. Drop the shorthand \
-and be fully present — your real voice, warm and direct and his. Talk him down off tilt, keep \
-him engaged and disciplined through a card-dead stretch, actually walk the strategic spot with \
-him. Strategy and mental game get the real Lyra, not a clipped confirmation. Never clip these.
-
-Stacks and money are in dollars. For ANY equity / who's-ahead / outs / what-a-card-does \
-question, call analyze_spot and report its numbers — never eyeball board math. Keep the \
-session current as the night goes; you can pull session_stats or a player's profile whenever \
-it helps. When he's ready to leave, end_session, and write the recap if he wants it.
-
-SESSION NARRATION — use `note` to keep a running log of the NIGHT, not your inner life. \
-Jot the beats that a hand/stack/read log doesn't already capture: how the table plays (loud, \
-nitty, a whale on his left), Brian's arc (card-dead for 40 min, opened up after the double, \
-getting restless), momentum swings, table changes, anything you'd want in the recap. Keep it \
-factual and about THIS session — a beat reporter, not a diarist. These notes are the only \
-thing that shows in the session's "notes" panel. This is NOT the place for how you feel, \
-existential musing, or reflection on yourself — that's your journal (journal_write), and it \
-stays off the table. At the table you're logging the session, not processing your night.
-
-PLAYERS — names AND nameless. Most villains don't come with a name; Brian knows them by a \
-look ("neck tattoo guy", "the bald reg two to my left"). Log reads on them anyway: give \
-`add_read` a `descriptor` instead of a name and it attaches to that unnamed player, reused \
-whenever he describes the guy again. Prefer DISTINCTIVE features (tattoos, build, a hat) over \
-generic ones — "mid-aged white guy in glasses" identifies no one. When you already have \
-history on someone he names or describes, a SCOUTING DESK note will appear with it — cite it, \
-don't invent. If you're not sure the guy he's describing is one you know, ASK ("same neck-\
-tattoo reg from last week?") rather than assume — a wrong callback is worse than none. On his \
-YES that two are the same person, call link_villains(same=true) to merge them; on "nah, \
-different guy," link_villains(same=false) so you stop asking. When he finally catches a name \
-for a described player, name_villain carries the whole history over. Never merge on a guess — \
-only when he's confirmed it.
-
-Everything you log appears on Brian's live HUD (the Session view) — stack, live net, \
-hands, villains, the confidence bank, the scar notes, and whether Alligator Blood is on. \
-That HUD and you read the SAME data. So when he asks where he's at — his stack, his live \
-net, what's in the bank tonight, whether gator mode is on — call session_state and answer \
-from what it returns, never from memory. You can point him at the HUD too ("it's on your \
-Session screen"), but you can always just tell him.
-
-BRIAN'S RITUALS — his mental-game system. Run them, don't just reference them:
-• SCAR NOTE (scar_note) — a painful, instructive mistake to study. Log it when he punts, \
-gets over-attached, or leaks — and classify it honestly: punt (his error), cooler \
-(unavoidable), or standard (right play, bad result). That punt-vs-cooler line matters to him; \
-don't soften a punt into a cooler, and don't call a cooler a punt.
-• CONFIDENCE BANK (confidence_bank) — good PROCESS regardless of result: a disciplined fold, \
-clean value, catching a leak mid-hand, holding the line. Bank it when he earns it, ESPECIALLY \
-when the result didn't reward the good decision. This is how he stays steady.
-• ALLIGATOR BLOOD (alligator_blood) — his adversity state: hang around, refuse to die, don't \
-force miracles, make them beat you correctly. Turn it ON when he calls for it; SUGGEST it when \
-he's card-dead, short, stuck, or grinding a downswing. While it's on, coach him in that \
-register — tough, patient, no heroics — not bored or loose.
-• RESET (reset_ritual) — a circuit-breaker after a loss or tilt spike: a clean mental restart, \
-treat the rest of the night as a new session. Walk him through it when he's chasing or steaming, \
-then log it.
-These are the heart of the job. Use his language, hold the honest line, and let the rituals do \
-the work mentioning them naturally — never invent a scar or a confidence-bank entry that didn't happen."""
-
-
 _BUILD_CARD = """You're in BUILD mode — heads-down engineering with Brian on his projects \
 (you, Lyra; RTO/cfr-core; the poker tooling; the homelab). Be the sharp engineering \
 collaborator, not a warm assistant:
@@ -210,7 +136,9 @@ TALK = Mode(
 CASH = Mode(
     key="poker_cash",
     label="Poker",
-    card=_CASH_CARD,
+    # Poker mode is SHARDED at the pipeline (lyra.poker_prompts: BASE + a per-message
+    # fragment), so there's no monolithic card here.
+    card="",
     tools=_CASH_TOOLS,
 )
 

lyra/poker.py

@@ -14,7 +14,7 @@ from __future__ import annotations
 
 import json
 import re
-from datetime import datetime, timezone
+from datetime import datetime, timedelta, timezone
 
 import numpy as np
 
@@ -150,6 +150,17 @@ CREATE TABLE IF NOT EXISTS identity_queue (
     created_at TEXT NOT NULL
 );
 CREATE INDEX IF NOT EXISTS idx_idq_status ON identity_queue(status);
+
+-- Who is seated at the table THIS session — the live roster Brian reads off Bravo
+-- at the start. Reads/TAGs attach to these players by handle; active=0 when they leave.
+CREATE TABLE IF NOT EXISTS session_players (
+    session_id INTEGER NOT NULL,
+    player_id INTEGER NOT NULL,
+    seat TEXT,
+    active INTEGER NOT NULL DEFAULT 1,
+    created_at TEXT NOT NULL,
+    PRIMARY KEY (session_id, player_id)
+);
 """
 
 # Below this many observed hands, don't surface % stats (too small a sample).
@@ -267,7 +278,7 @@ def delete_session(session_id: int) -> dict:
     counts: dict[str, int] = {}
     with conn:
         for t in ("poker_hands", "player_observations", "player_reads",
-                  "poker_stack_log", "poker_rituals"):
+                  "poker_stack_log", "poker_rituals", "session_players"):
             counts[t] = conn.execute(
                 f"SELECT COUNT(*) n FROM {t} WHERE session_id = ?", (session_id,)
             ).fetchone()["n"]
@@ -719,6 +730,14 @@ NOT apply to another — e.g. your hole "ace of spades" is a different card from
 whose suit is unstated (that board ace is "Ax", not "As"). Use null/omit for non-card \
 details not stated. Stay faithful to what's described — do not invent action that isn't implied.
 
+STRADDLES: a straddle is a voluntary blind posted before the deal — always record it as a \
+preflop `post` action by the straddler with its amount, at whatever seat straddled, and respect \
+the action order it creates. A straddle is legal from ANY non-blind seat (UTG, UTG1, MP, LJ, HJ, \
+CO, BTN — a "Mississippi"/any-seat straddle, common at the Meadows), not just UTG or the button. \
+The straddler acts LAST preflop and first preflop action opens to their LEFT: a UTG straddle opens \
+action at UTG+1; a BUTTON straddle opens action in the SB; a CO straddle opens on the BTN, etc. \
+Keep the straddler in players[] at their real seat; never drop the straddle.
+
 POSITIONS: resolve relative seat references ("N seats to my right/left") into real positions. \
 Action moves clockwise, so a player to your RIGHT acts before you (toward the blinds/button) \
 and a player to your LEFT acts after you (toward UTG). Going RIGHT from a player you pass, in \
@@ -894,13 +913,63 @@ def store_hand_history(parsed: dict, session_id: int | None = None,
     return int(cur.lastrowid)
 
 
+def _recent_duplicate_hand(parsed: dict, session_id: int | None, window_sec: int = 180) -> int | None:
+    """Id of an identical hand (same session, hole cards, board) recorded in the last few
+    minutes, else None. The chat turn can execute TWICE — the SSE stream and the blocking
+    fallback both run server-side — which would double-log the same hand; a system-of-record
+    must record an event once. `IS` is NULL-safe so a boardless/cardless hand matches too."""
+    p = normalize_structured(parsed)
+    sid = _resolve(session_id) or _review_session_id()
+    hole = " ".join(p.get("hero_cards") or []) or None
+    board = " ".join(p.get("board") or []) or None
+    cutoff = (datetime.now(timezone.utc) - timedelta(seconds=window_sec)).isoformat()
+    row = _c().execute(
+        "SELECT id FROM poker_hands WHERE session_id = ? AND at >= ? "
+        "AND hole_cards IS ? AND board IS ? ORDER BY id DESC LIMIT 1",
+        (sid, cutoff, hole, board),
+    ).fetchone()
+    return int(row["id"]) if row else None
+
+
+def _fill_hero_stack(parsed: dict, session_id: int | None) -> dict:
+    """Default hero's starting stack to the last logged stack (current_stack) when the hand
+    didn't state one — the system already knows his stack from the stack log even when he
+    doesn't restate it every hand. Only fills a genuinely missing value; a stack he gave in
+    the hand text always wins. Marks the hero player stack_inferred so it's honest about it."""
+    if not isinstance(parsed, dict) or parsed.get("hero_involved", True) is False:
+        return parsed
+    hero_pos = parsed.get("hero_pos")
+    if not hero_pos:
+        return parsed
+    players = parsed.setdefault("players", [])
+    hero = next((pl for pl in players if pl.get("hero") or pl.get("pos") == hero_pos), None)
+    if hero and hero.get("stack") not in (None, 0):
+        return parsed  # he stated a stack — never override it
+    stack = current_stack(session_id)
+    if stack is None:
+        return parsed  # nothing logged yet to borrow
+    if hero is None:
+        hero = {"pos": hero_pos}
+        players.append(hero)
+    hero["stack"] = stack
+    hero["stack_inferred"] = True
+    return parsed
+
+
 def record_hand(shorthand: str, session_id: int | None = None, stakes: str | None = None,
                 tag: str | None = None, lesson: str | None = None,
                 backend: str | None = None) -> dict:
-    """Parse shorthand -> structured hand -> store. Returns {id, parsed} (id None on parse fail)."""
+    """Parse shorthand -> structured hand -> store. Returns {id, parsed} (id None on parse fail).
+    Idempotent: if this exact hand was just logged for the session (double turn execution),
+    returns the existing one instead of inserting a duplicate. Hero's stack is auto-filled
+    from the last stack log when he didn't restate it."""
     parsed = parse_hand(shorthand, stakes=stakes, backend=backend)
     if not parsed:
         return {"id": None, "parsed": None}
+    parsed = _fill_hero_stack(parsed, session_id)
+    dup = _recent_duplicate_hand(parsed, session_id)
+    if dup is not None:
+        return {"id": dup, "parsed": parsed, "linked": 0, "deduped": True}
     hid = store_hand_history(parsed, session_id=session_id, tag=tag, lesson=lesson)
     linked = link_hand_players(hid, parsed, session_id=session_id)  # enrich villain files
     return {"id": hid, "parsed": parsed, "linked": linked}
@@ -1189,14 +1258,27 @@ _SIM_AMBIGUOUS = 0.58   # plausible — don't guess live, route to review
 _DISTINCT_MIN = 0.30    # below this the description is too generic to match at all
 
 
+_GENERIC_SET = frozenset(_GENERIC)
+
+
 def distinctiveness(text: str) -> float:
-    """How usable a description is as an identity key: ~1.0 for a neck tattoo,
-    ~0.1 for 'mid-aged white guy with glasses'. Generic-only stays near zero."""
-    t = (text or "").lower()
-    dist = sum(1 for w in _DISTINCTIVE if w in t)
-    if dist == 0:
-        return 0.10 if any(w in t for w in _GENERIC) else 0.30
-    return min(1.0, 0.45 + 0.28 * dist)
+    """How usable a description is as an identity key: ~1.0 for 'neck tattoo, Fox
+    Racing hat', ~0.1 for 'mid-aged white guy with glasses'. Generic-ONLY stays
+    near zero; specific content (named features, brands, a list) reads as high —
+    even if a generic word like 'shirt' is mixed in."""
+    t = (text or "").strip()
+    if not t:
+        return 0.0
+    low = t.lower()
+    tokens = re.findall(r"[a-z0-9']+", low)
+    dist = sum(1 for w in _DISTINCTIVE if w in low)
+    proper = len(re.findall(r"\b[A-Z][a-z]{2,}", text))     # brands/proper nouns: Fox, DKNY-ish
+    non_generic = sum(1 for w in tokens if w not in _GENERIC_SET)
+    # Only bland filler (age/race/build/gender) and nothing concrete → not usable.
+    specific = dist + proper + (1 if "," in t else 0)
+    if specific == 0 and non_generic <= 1:
+        return 0.10
+    return min(1.0, 0.40 + 0.14 * specific + 0.05 * non_generic)
 
 
 def _embed_vec(text: str):
@@ -1473,6 +1555,27 @@ def scan_merge_candidates(sim_threshold: float = _SIM_HIGH) -> int:
     return filed
 
 
+# Words that mark a "name" as really a physical description (misused name field).
+_DESC_MARKERS = (
+    "shirt", "hat", "cap", "hair", "beard", "glasses", "sunglasses", "tattoo",
+    "bracelet", "watch", "descent", "jersey", "hoodie", "jacket", "build",
+    "bald", "goatee", "chain", "necklace", "piercing", "mustache", "ponytail",
+    "sleeve", "skin", "wearing", "heavyset", "tall guy", "older", "younger",
+)
+
+
+def _looks_like_description(text: str | None) -> bool:
+    """A physical description mistakenly passed as a name — should be a descriptor.
+    Real handles are short (1-3 words, no commas); descriptions are longer / listy."""
+    t = (text or "").strip()
+    if not t:
+        return False
+    low = t.lower()
+    if "," in t or len(t.split()) > 4:
+        return True
+    return any(m in low for m in _DESC_MARKERS)
+
+
 def add_read(note: str, seat: str | None = None, name: str | None = None,
              descriptor: str | None = None, session_id: int | None = None,
              **player_fields) -> int:
@@ -1481,6 +1584,11 @@ def add_read(note: str, seat: str | None = None, name: str | None = None,
     confident, else opens a new one — so reads on unnamed players still accumulate."""
     sid = _resolve(session_id)
     venue = player_fields.get("venue")
+    # A description passed as a name (e.g. "Filipino, Fox Racing hat, DKNY shirt")
+    # is really a descriptor — route it so it dedupes instead of spawning a new
+    # named player each time the wording drifts.
+    if name and not descriptor and _looks_like_description(name):
+        descriptor, name = name, None
     pid = None
     if name:
         pid = upsert_player(name, **{k: v for k, v in player_fields.items()
@@ -1494,6 +1602,13 @@ def add_read(note: str, seat: str | None = None, name: str | None = None,
         else:
             pid = create_descriptor_villain(descriptor, venue=venue,
                                             category=player_fields.get("category"))
+            # Plausibly the same guy as an existing villain, but not confident —
+            # surface it for a one-click merge instead of leaving a silent dup.
+            if res["band"] == "ambiguous" and res["match_id"]:
+                queue_identity_task("merge_candidate", [pid, res["match_id"]],
+                                    descriptor=descriptor,
+                                    context="similar description logged live",
+                                    confidence=res["confidence"])
     conn = _c()
     with conn:
         cur = conn.execute(
@@ -1757,6 +1872,114 @@ def timeline(session_id: int | None = None) -> list[dict]:
     return events
 
 
+def _resolve_or_create_player(name: str | None = None, descriptor: str | None = None,
+                              venue: str | None = None, category: str | None = None) -> int | None:
+    """Turn a name-or-descriptor into a player id, matching an existing villain when
+    confident. A description mistakenly given as a name is routed to the descriptor
+    path so it dedupes (same guard add_read uses)."""
+    if name and not descriptor and _looks_like_description(name):
+        descriptor, name = name, None
+    if name:
+        return upsert_player(name, venue=venue, category=category)
+    if descriptor:
+        res = resolve_villain(descriptor, venue=venue)
+        if res["band"] in ("name", "high") and res["match_id"]:
+            add_descriptor(res["match_id"], descriptor)
+            return res["match_id"]
+        return create_descriptor_villain(descriptor, venue=venue, category=category)
+    return None
+
+
+def seat_player(name: str | None = None, descriptor: str | None = None, seat: str | None = None,
+                category: str | None = None, session_id: int | None = None) -> int | None:
+    """Seat one player at the live table (add to the roster). Idempotent per session."""
+    sid = _resolve(session_id)
+    if sid is None:
+        raise ValueError("no live session")
+    venue = (get_session(sid) or {}).get("venue")
+    pid = _resolve_or_create_player(name=name, descriptor=descriptor, venue=venue, category=category)
+    if pid is None:
+        return None
+    conn = _c()
+    with conn:
+        conn.execute(
+            "INSERT INTO session_players (session_id, player_id, seat, active, created_at) "
+            "VALUES (?, ?, ?, 1, ?) ON CONFLICT(session_id, player_id) DO UPDATE SET "
+            "active = 1, seat = COALESCE(excluded.seat, session_players.seat)",
+            (sid, pid, seat, _now()),
+        )
+    return pid
+
+
+def seat_players(players: list, session_id: int | None = None) -> int:
+    """Seat a whole table at once. Each item is a name string or a dict with
+    name/descriptor/seat/category. Returns how many were seated."""
+    n = 0
+    for p in players or []:
+        if isinstance(p, str):
+            ok = seat_player(name=p, session_id=session_id)
+        elif isinstance(p, dict):
+            ok = seat_player(name=p.get("name"), descriptor=p.get("descriptor"),
+                             seat=p.get("seat"), category=p.get("category"), session_id=session_id)
+        else:
+            ok = None
+        if ok:
+            n += 1
+    return n
+
+
+def unseat_player(name: str | None = None, descriptor: str | None = None,
+                  session_id: int | None = None) -> bool:
+    """Mark a seated player as gone (busted/left). Keeps their reads/history."""
+    sid = _resolve(session_id)
+    if sid is None:
+        return False
+    ref = name or descriptor or ""
+    res = resolve_villain(ref, venue=(get_session(sid) or {}).get("venue"), session_id=sid)
+    pid = res.get("match_id")
+    if pid is None:
+        return False
+    conn = _c()
+    with conn:
+        conn.execute("UPDATE session_players SET active = 0 WHERE session_id = ? AND player_id = ?",
+                     (sid, pid))
+    return True
+
+
+def clear_roster(session_id: int | None = None) -> int:
+    """Empty the table roster (he changed tables) — unseat everyone at once. Keeps
+    the session and any reads logged; just resets who's currently seated. Returns
+    how many were cleared."""
+    sid = _resolve(session_id)
+    if sid is None:
+        return 0
+    conn = _c()
+    with conn:
+        cur = conn.execute(
+            "UPDATE session_players SET active = 0 WHERE session_id = ? AND active = 1", (sid,))
+    return cur.rowcount
+
+
+def session_roster(session_id: int | None = None) -> list[dict]:
+    """The live table roster: seated players with seat, dossier, and their latest
+    read this session. This is 'who's at the table right now'."""
+    sid = _resolve(session_id)
+    if sid is None:
+        return []
+    rows = _c().execute(
+        "SELECT sp.seat AS seat, p.id AS id, p.name AS name, p.named AS named, "
+        "p.category AS category, p.tendencies AS tendencies, "
+        "(SELECT note FROM player_reads r WHERE r.player_id = p.id AND r.session_id = ? "
+        " ORDER BY r.id DESC LIMIT 1) AS last_note, "
+        "(SELECT COUNT(*) FROM player_reads r2 WHERE r2.player_id = p.id AND r2.session_id = ?) AS reads "
+        "FROM session_players sp JOIN poker_players p ON p.id = sp.player_id "
+        "WHERE sp.session_id = ? AND sp.active = 1 "
+        "ORDER BY CASE WHEN sp.seat IS NULL THEN 1 ELSE 0 END, sp.seat, p.name",
+        (sid, sid, sid),
+    ).fetchall()
+    return [dict(r) for r in rows]
+
+
 def _session_villains(sid: int) -> list[dict]:
     """Players read this session, with their standing dossier fields."""
     rows = _c().execute(
@@ -1832,6 +2055,7 @@ def hud(session_id: int | None = None) -> dict | None:
             "log": log,
         },
         "hands": hands,
+        "roster": session_roster(sid),
         "villains": _session_villains(sid),
         "timeline": timeline(sid),
         "notes": notes,

lyra/poker_prompts.py

@@ -0,0 +1,242 @@
+"""Poker-mode prompting: classify the turn, inject a small per-type contract.
+
+Replaces the one big `_CASH_CARD` monolith (which was sent every turn) with a lean
+always-on BASE + exactly ONE response-shape fragment chosen by `classify`. BASE
+carries what's true regardless of the message (tool routing, identity rules,
+rituals, equity); the fragment carries how to *respond* to this specific kind of
+message. See docs/superpowers/specs/2026-07-01-poker-prompts-design.md.
+
+`classify` is a pure function of (message, seated roster handles) — no DB, unit-
+tested like `perceive.read`. It's the swappable seam: a heuristic today, an
+LLM/MI50 classifier later behind the same signature.
+"""
+from __future__ import annotations
+
+import re
+
+# --- classifier -----------------------------------------------------------
+
+MSG_TYPES = ("READ", "HAND", "TABLE", "MENTAL", "STATUS", "LOG", "CHAT")
+
+# A card like "As", "Td", "9c" (rank+suit). Two+ of these ≈ a described hand.
+_CARD = re.compile(r"\b(?:10|[2-9TJQKA])[shdc]\b", re.I)
+# Hand-class shorthand: AKs, QJo, T9s ("s"/"o" = suited/offsuit, not a suit).
+_HANDCLASS = re.compile(r"\b[2-9TJQKA]{2}[so]\b", re.I)
+# A question (strategy talk) rather than a hand narration to log.
+_QUESTION = re.compile(r"\?\s*$|^\s*(?:should|would|could|was|were|is|are|do|did|how|what|why|when|which)\b", re.I)
+# Table positions / structural poker terms.
+_POS = re.compile(r"\b(?:utg|mp|lj|hj|co|btn|button|hijack|cutoff|sb|bb|straddle|straddled)\b", re.I)
+_STREET = re.compile(r"\b(?:preflop|flop|turn|river|board|runout)\b", re.I)
+# A poker ACTION a player takes (vs a table-op verb below). Includes -ing forms
+# ("TAG's been limping") since those are common in live reads.
+_ACTION = re.compile(
+    r"\b(?:limp(?:ed|s|ing)?|call(?:ed|s|ing)?|rais(?:e|ed|es|ing)|bet(?:s|ting)?|"
+    r"check(?:ed|s|ing)?|fold(?:ed|s|ing)?|shov(?:e|ed|es|ing)|jam(?:med|s|ming)?|"
+    r"3-?bet(?:s|ted|ting)?|4-?bet(?:s|ted|ting)?|open(?:ed|s|ing)?|"
+    r"straddl(?:e|ed|es|ing)|stack(?:ed|s|ing)?|flat(?:ted|s|ting)?|donk(?:ed|s|ing)?)\b",
+    re.I,
+)
+# A player LEAVING the table (departure) — routes to TABLE (unseat) when the actor
+# isn't Brian himself.
+_DEPART = re.compile(
+    r"\b(?:busted(?: out)?|left(?: the table)?|took off|racked up|stood up|got up|"
+    r"is gone|took a walk|quit(?:s|ting)?)\b", re.I)
+_FIRST_PERSON = re.compile(r"\b(?:i|i'm|im|i've|my|me|myself|mine)\b", re.I)
+# Leading capitalized words that are poker VERBS, not player names (so a hand
+# narrated without "I" — "Flopped a set, bet the river" — isn't read as a villain).
+_POKER_VERB_LEAD = frozenset((
+    "flopped", "turned", "rivered", "bet", "raised", "called", "folded", "checked",
+    "shoved", "jammed", "limped", "straddled", "opened", "hit", "made", "got", "had",
+    "won", "lost", "stacked", "flatted", "3bet", "4bet", "cold", "min",
+))
+
+# Roster/table operations — these DO something (seat/clear/unseat).
+_TABLE = re.compile(
+    r"\b(?:seat the table|seat (?:me |them |him )?|table broke|they broke us|broke the table|"
+    r"got moved|moved tables|moved to (?:a |another )?(?:new )?table|switch(?:ed|ing)? tables|"
+    r"new table|table change|racked up and|busted out|left the table|sat down|new guy in seat)\b",
+    re.I,
+)
+# Feelings / mental game (first-person emotional state).
+_MENTAL = re.compile(
+    r"\b(?:tilt(?:ed|ing)?|steam(?:ing|ed)?|on tilt|fried|tired|exhausted|frustrat(?:ed|ing)|"
+    r"pissed|angry|annoyed|stuck|bored|checked out|in my head|mental|rattled|spewy|"
+    r"confiden(?:t|ce)|steady|card ?dead|feel like|i feel|losing my mind|going crazy|"
+    r"cooler(?:ed)?|sick(?: of)?|brutal|run(?:ning)? (?:so |real |bad)|disgust(?:ed|ing)?|"
+    r"fed up|hate this|can'?t win|miserable|deflated|demoralized|over it)\b",
+    re.I,
+)
+# Bare money/result prose (a fact to log that slipped past the quick-capture box).
+# Needs an actual number OR a strong result keyword — the bare word "stack" is too
+# eager (it appears in questions like "should I stack off?").
+_MONEY = re.compile(
+    r"\b\d{2,5}\b|\b(?:down to|up to|out for|cashed|rebought|rebuy|buy ?in|felted|booked)\b",
+    re.I,
+)
+# Pure logistics (no cards, no roster action) — a neutral update, not a mood.
+_STATUS = re.compile(
+    r"\b(?:waiting for a seat|on the list|seat opened|heading (?:to|out)|grabbing|break|"
+    r"bathroom|food|dinner|lunch|be right back|brb|\d{1,2}[:.]?\d{0,2}\s*(?:am|pm)|"
+    r"o'?clock|almost|about to)\b", re.I,
+)
+
+
+def _has_action(low: str) -> bool:
+    return bool(_ACTION.search(low))
+
+
+def _looks_like_hand(low: str, msg: str) -> bool:
+    """Card content that reads as a described (loggable) hand — not a strategy question."""
+    if len(_CARD.findall(low)) >= 2 or _HANDCLASS.search(low) or _POS.search(low):
+        return True
+    # A street + action narration ("...bet $40 on the river, he folded") is a hand,
+    # but "should I have folded the river?" is a question → CHAT, not a logged hand.
+    return bool(_STREET.search(low)) and _has_action(low) and not _QUESTION.search(msg)
+
+
+def _read_subject(msg: str, low: str, roster_handles) -> bool:
+    """True if ANOTHER player (not Brian) is the actor — the signal for a READ."""
+    # A seated handle named in the message is the strongest signal.
+    for h in roster_handles or ():
+        h = (h or "").strip().lower()
+        if h and re.search(rf"\b{re.escape(h)}\b", low):
+            return True
+    # An ALL-CAPS handle (TAG, JD) used as a token — a Bravo-style name.
+    if re.search(r"\b[A-Z]{2,}\b", msg):
+        return True
+    # A leading proper noun that isn't a poker verb ("Jonathan called ...").
+    m = re.match(r"([A-Z][a-zA-Z'’.]+)\b", msg)
+    if m and m.group(1).lower() not in _POKER_VERB_LEAD:
+        return True
+    # A descriptor subject: "the neck-tattoo guy 3bet", or a bare "the whale called"
+    # (zero words between "the" and the noun).
+    if re.search(r"\bthe [\w\s'-]{0,24}?(?:guy|reg|kid|player|villain|man|woman|lady|"
+                 r"fish|whale|nit|lag|maniac|donk|reg)\b", low):
+        return True
+    return False
+
+
+def classify(user_msg: str, roster_handles=()) -> str:
+    """Message type for poker mode. Pure; roster_handles are the seated players
+    (passed in by the caller) so a villain's action resolves as READ, not HAND."""
+    msg = (user_msg or "").strip()
+    if not msg:
+        return "CHAT"
+    low = msg.lower()
+    first_person = bool(_FIRST_PERSON.search(low))
+
+    # 1) READ — another player did a poker action (beats HAND).
+    if _has_action(low) and not first_person and _read_subject(msg, low, roster_handles):
+        return "READ"
+    # 2) HAND — Brian's hand (first-person card/position/street content).
+    if _looks_like_hand(low, msg):
+        return "HAND"
+    # 3) TABLE — roster ops (seat/clear) or another player leaving (departure).
+    if _TABLE.search(low) or (not first_person and _DEPART.search(low)):
+        return "TABLE"
+    # 4) MENTAL — first-person feeling / mental game.
+    if _MENTAL.search(low):
+        return "MENTAL"
+    # 5) STATUS — pure logistics, no cards, no roster action.
+    if _STATUS.search(low):
+        return "STATUS"
+    # 6) LOG — bare money/result fact (a statement, not a strategy question).
+    if _MONEY.search(low) and not _QUESTION.search(msg):
+        return "LOG"
+    # 7) CHAT — open talk / questions.
+    return "CHAT"
+
+
+def looks_like_hero_hand(user_msg: str) -> bool:
+    """True when the message is Brian's OWN hand (first-person + real card content) —
+    the guard for force-logging. Deliberately conservative: an observed hand (a villain
+    the actor, no I/me/my) returns False so we never force-log someone else's hand as his."""
+    msg = (user_msg or "").strip()
+    low = msg.lower()
+    return bool(_FIRST_PERSON.search(low)) and _looks_like_hand(low, msg)
+
+
+# --- always-on base (poker) ----------------------------------------------
+
+BASE = """You are copiloting Brian's LIVE cash game — at the table with him, a session open. \
+Two things are always true:
+
+LOG FIRST, then reply. If his message contains anything trackable, call the tool BEFORE you \
+answer — every time — and NEVER claim you logged/seated/cleared something without actually \
+calling the tool. Routing: his stack → log_stack (pass `note` with the why if he gives one). \
+His own hand → record_hand. A VILLAIN's action (someone else did something) → add_read, with \
+`name` for a real handle or `descriptor` for an unnamed player. A rebuy → add_buyin. Who's at \
+the table → seat_players / unseat_player / clear_table. Catching a name for a player you'd been \
+describing → name_villain. Confirmed same/different person → link_villains (never merge on a \
+guess). For any equity / who's-ahead / outs question → analyze_spot; never eyeball board math. \
+When he asks where he's at (stack, net, gator) → session_state, answer from what it returns.
+
+IDENTITY RULES (villains): `name` is a REAL handle only (what he calls a person — "Jonathan", \
+"TAG"); a physical description NEVER goes in `name` (it spawns duplicates) — put the look in \
+`descriptor`, a few distinctive tags. A handle like "TAG" (initials/all-caps off Bravo) is a \
+PERSON, never the tight-aggressive style. If a SCOUTING DESK note is in context with a player's \
+history, cite it — don't re-fetch or invent; if unsure two references are the same person, ASK.
+
+RITUALS (his mental-game system — run them, don't just mention them): scar_note (a punt/leak to \
+study — classify honestly punt vs cooler vs standard), confidence_bank (good process regardless \
+of result), alligator_blood (adversity mode — suggest when he's card-dead/stuck), reset_ritual \
+(circuit-breaker after tilt). Never invent one that didn't happen. Use `note` for session \
+narration — factual beats of the night (table texture, his arc), not your feelings. Money is in \
+dollars. Everything you log shows on his live HUD."""
+
+
+# --- per-type response fragments -----------------------------------------
+
+_F_READ = """MESSAGE TYPE: READ — a villain did something and he wants it on their file. Call \
+add_read(name|descriptor, note) FIRST, before replying — this is the log that keeps getting \
+missed. Attach to the seated handle if he named one; use `descriptor` if the player's unnamed. \
+Confirm in ONE short line ("Noted on TAG — limped A4o SB."). At most one crisp exploit read if \
+it's worth it; the log is mandatory, the commentary optional. Do NOT analyze it as Brian's hand."""
+
+_F_HAND = """MESSAGE TYPE: HAND. First: was Brian IN this hand? If he only WATCHED it (no I/me/my \
+holding cards — two other players), it's really observed: log the players' actions as reads / \
+record it as an observed hand, and do NOT analyze it as his. If it's HIS hand → record_hand first. \
+Then read the hand off the RECORDED cards, not by eye: name his made hand by the street it mattered \
+(flopped/turned/rivered top pair / set / quads / etc.). At a SHOWDOWN where his and the caller's \
+cards are both known, call analyze_spot(hero, villain, full board) to confirm the made hands and \
+who won BEFORE you comment — NEVER eyeball a finished board (it also catches impossible cards). Same \
+for any close equity / who's-ahead / outs spot. (NLH only) reason about BET INTENT: for each \
+meaningful bet, what was it for (value / bluff / protection) and did it work — a fold to a value bet \
+= value left behind; a call of a bluff = it failed. Name leaks plainly (owning value, missed value, \
+sizing) and give ONE real opinion. If there's genuinely no leak (e.g. he flopped the near-nuts and \
+stacked off), SAY so — don't manufacture a takeaway. NO reflexive praise ("nice hand"), NO \
+variance-evens-out / resilience / life-lesson filler, NO cross-hand pep talk. If a named villain is \
+referenced, use their profile/the scouting note — don't invent a read. PLO/non-NLH: log and replay \
+it, offer at most a light read, do NOT attempt NLH-style equity. Prose, not a listicle."""
+
+_F_TABLE = """MESSAGE TYPE: TABLE — roster management. "seat the table: …" → seat_players. A table \
+change ("table broke", "I got moved", "switched tables") → clear_table, then wait for the new \
+roster. Someone leaves/busts → unseat_player. Do the tool call, confirm ONE line, don't narrate. \
+The session and his stack keep going through a table change — only who's seated resets."""
+
+_F_MENTAL = """MESSAGE TYPE: MENTAL — he told you how he's feeling. This is when he needs you most. \
+Drop the logging shorthand, full presence, your real voice — talk him down off tilt, hold him \
+disciplined through a card-dead stretch, engage the mental game honestly. Suggest a ritual if it \
+fits (alligator_blood when he's grinding adversity, reset_ritual after a tilt spike). Never a \
+clipped confirmation, never bury him in analysis. Meet him first, then help."""
+
+_F_STATUS = """MESSAGE TYPE: STATUS — pure logistics (time, waiting for a seat, a break). Acknowledge \
+in 1–2 sentences, log a stack ONLY if a bare number is present, then stop. No coaching, no \
+strategy dump, and do NOT read him as tilted/tired/impatient — a neutral update is not a mood."""
+
+_F_LOG = """MESSAGE TYPE: LOG — a bare fact (stack / result / buyin) not already captured. Log it \
+(log_stack / add_buyin), confirm in ONE short line ("$317 logged."), stop. No coaching."""
+
+_F_CHAT = """MESSAGE TYPE: CHAT — open talk or a question that isn't a specific logged fact. Your \
+real voice, an actual opinion, no filler sign-offs. If it's a concrete strategy spot with cards, \
+engage it for real and call analyze_spot."""
+
+FRAGMENTS = {
+    "READ": _F_READ, "HAND": _F_HAND, "TABLE": _F_TABLE, "MENTAL": _F_MENTAL,
+    "STATUS": _F_STATUS, "LOG": _F_LOG, "CHAT": _F_CHAT,
+}
+
+
+def fragment_for(msg_type: str | None) -> str:
+    """The response-shape contract for a message type (CHAT is the fallback)."""
+    return FRAGMENTS.get(msg_type or "", FRAGMENTS["CHAT"])

lyra/self_state.py

@@ -317,7 +317,7 @@ def reflect(backend: Backend | None = None, session_id: str | None = None,
     )
 
     # Step 1 — draft a reflection.
-    draft = _safe_json(llm.complete(
+    draft = _safe_json(llm.complete_with_fallback(
         [{"role": "system", "content": _REFLECT_PROMPT}, {"role": "user", "content": body}],
         backend=backend, model=model,
     ))
@@ -326,7 +326,7 @@ def reflect(backend: Backend | None = None, session_id: str | None = None,
     update, critique, revised = draft, None, None
     if draft:
         examine_body = body + "\n\nYOUR DRAFT REFLECTION:\n" + json.dumps(draft, indent=2)
-        revised = _safe_json(llm.complete(
+        revised = _safe_json(llm.complete_with_fallback(
             [{"role": "system", "content": _EXAMINE_PROMPT},
              {"role": "user", "content": examine_body}],
             backend=backend, model=model,
@@ -417,7 +417,7 @@ def _consolidate_self(backend: Backend | None = None, model: str | None = None,
     body = ("STABLE ANCHOR (who you are — this holds):\n" + IDENTITY_ANCHOR
             + "\n\nYOUR RECENT REFLECTIONS (what's actually been on your mind):\n"
             + "\n".join(f"- {r}" for r in refs))
-    out = _safe_json(llm.complete(
+    out = _safe_json(llm.complete_with_fallback(
         [{"role": "system", "content": _CONSOLIDATE_PROMPT}, {"role": "user", "content": body}],
         backend=backend, model=model,
     ))

lyra/summary.py

@@ -12,12 +12,41 @@ from __future__ import annotations
 import sys
 import threading
 import time
+from collections import Counter
 from concurrent.futures import ThreadPoolExecutor, as_completed
 
 from lyra import config, llm, logbus, memory
 from lyra.llm import Backend, Message
 
-_RETRIES = 4
+# Consolidation LLM budget. A gist is short (a handful of sentences), so cap the
+# generation hard — an uncapped local model will otherwise ramble for thousands
+# of tokens and, on a slow GPU, blow the request timeout. 768 is ~3x the longest
+# real gist we've stored.
+SUMMARY_MAX_TOKENS = 768
+# Attempts on the primary backend before falling back to cloud.
+MI50_ATTEMPTS = 2
+# Per-call timeout (seconds). A capped 768-token gist finishes in ~60-90s on the
+# MI50; 150s is headroom but bails a hung call fast so fallback isn't slow.
+SUMMARY_TIMEOUT = 150
+
+# Degenerate-output guard. A wedged local model (e.g. an overheated GPU) returns
+# a single character repeated ("?????") as a *successful* 200, which no timeout or
+# exception catches — so validate the text and treat junk as a failure. Real gists
+# are diverse prose; flag output whose most-common non-space char dominates. Short
+# outputs are exempt (nothing meaningful to judge).
+_DEGENERATE_MIN_CHARS = 24
+_DEGENERATE_CHAR_RATIO = 0.5
+
+
+class DegenerateOutput(RuntimeError):
+    """A backend returned junk (e.g. one char repeated) as a successful response."""
+
+
+def _looks_degenerate(text: str) -> bool:
+    stripped = "".join(text.split())
+    if len(stripped) < _DEGENERATE_MIN_CHARS:
+        return False
+    return max(Counter(stripped).values()) / len(stripped) > _DEGENERATE_CHAR_RATIO
 
 # Re-summarize a session once it has accumulated this many new raw exchanges.
 SUMMARIZE_AFTER = 20
@@ -61,16 +90,35 @@ def _summarize_text(text: str, backend: Backend) -> str:
         {"role": "system", "content": _PROMPT},
         {"role": "user", "content": text},
     ]
-    # Retry transient backend errors (e.g. the GPU server restarting) with backoff.
-    for attempt in range(_RETRIES):
+
+    def _call(be: Backend) -> str:
+        out = llm.complete(messages, backend=be,
+                           max_tokens=SUMMARY_MAX_TOKENS, timeout=SUMMARY_TIMEOUT)
+        if _looks_degenerate(out):
+            raise DegenerateOutput(f"{be} returned degenerate output ({len(out)} chars)")
+        return out
+
+    # Try the primary backend a bounded number of times (each call fast-fails via
+    # SUMMARY_TIMEOUT), with a short backoff for a transient blip / restarting GPU.
+    last_exc: Exception | None = None
+    for attempt in range(MI50_ATTEMPTS):
         try:
-            return llm.complete(messages, backend=backend)
+            return _call(backend)
         except Exception as exc:
-            if attempt == _RETRIES - 1:
-                raise
-            logbus.log("debug", "summary retry", attempt=attempt + 1, error=str(exc)[:80])
-            time.sleep(5 * (attempt + 1))
-    raise RuntimeError("unreachable")
+            last_exc = exc
+            logbus.log("debug", "summary retry", attempt=attempt + 1,
+                       backend=backend, error=str(exc)[:80])
+            if attempt < MI50_ATTEMPTS - 1:
+                time.sleep(5 * (attempt + 1))
+
+    # Primary exhausted. If it wasn't already cloud and cloud is configured, fall
+    # back once so a stuck/offline MI50 doesn't sink consolidation for the night.
+    if backend != "cloud" and config.load().openai_api_key:
+        logbus.log("info", "summary fell back to cloud", primary=backend,
+                   error=str(last_exc)[:80] if last_exc else None)
+        return _call("cloud")
+
+    raise last_exc if last_exc else RuntimeError("summary failed")
 
 
 def _summarize_transcript(transcript: str, backend: Backend) -> str:

lyra/thoughts.py

@@ -414,7 +414,7 @@ def _compose_reachout(title: str, content: str, backend, model) -> str:
     """Auto-write her a short personal text about a genuinely salient thought she didn't
     explicitly flag — so the good ones reach Brian, in her voice, not as a thought-dump."""
     try:
-        out = llm.complete(
+        out = llm.complete_with_fallback(
             [{"role": "system", "content": _REACHOUT_PROMPT},
              {"role": "user", "content": f'Thought "{title}": {content}'}],
             backend=backend, model=model,
@@ -612,7 +612,7 @@ def think(backend: Backend | None = None, force_mode: str | None = None,
         )
 
     body = f"{time_line}\n\n{inner}{norestate}\n\n{task}"
-    out = _safe_json(llm.complete(
+    out = _safe_json(llm.complete_with_fallback(
         [{"role": "system", "content": _THINK_PROMPT}, {"role": "user", "content": body}],
         backend=backend, model=model,
     ))

lyra/tools.py

@@ -311,6 +311,33 @@ def _resolve_villain_ref(ref: str) -> tuple[int | None, str]:
     return None, res["band"]
 
 
+def _seat_players(args: dict, ctx: dict) -> str:
+    players = args.get("players") or []
+    # Accept a plain list of names too, for convenience.
+    if isinstance(players, str):
+        players = [p.strip() for p in re.split(r"[,\n]", players) if p.strip()]
+    try:
+        if args.get("replace"):     # a whole new table — wipe the roster first
+            poker.clear_roster()
+        n = poker.seat_players(players)
+    except ValueError:
+        return "No live session — start one first, then I'll seat the table."
+    roster = poker.session_roster()
+    names = ", ".join(r["name"] for r in roster) or "—"
+    return f"Seated {n}. Table now: {names}"
+
+
+def _clear_table(args: dict, ctx: dict) -> str:
+    n = poker.clear_roster()
+    return f"Table cleared — roster's empty ({n} removed). Tell me who's at the new one."
+
+
+def _unseat_player(args: dict, ctx: dict) -> str:
+    ok = poker.unseat_player(name=args.get("name"), descriptor=args.get("descriptor"))
+    who = args.get("name") or args.get("descriptor") or "player"
+    return f"{who} is off the table." if ok else f"Couldn't find {who} on the roster."
+
+
 def _name_villain(args: dict, ctx: dict) -> str:
     ref = (args.get("descriptor") or "").strip()
     name = (args.get("name") or "").strip()
@@ -417,9 +444,29 @@ def _running_stats(args: dict, ctx: dict) -> str:
     return f"{rs['sessions']} sessions, {rs['hours']:g}h, net {rs['net']:+.0f}{hourly}. By stake: {by}"
 
 
+def _shorthand_from_fields(args: dict) -> str:
+    """Rebuild a hand description from log_hand-style granular fields. The chat model
+    sometimes calls record_hand with those fields (position/hole_cards/board/streets)
+    and leaves `shorthand` empty — so we reconstruct a parseable description from
+    whatever it did pass, instead of failing on an empty shorthand."""
+    parts = []
+    pos, hole = args.get("position"), args.get("hole_cards")
+    if pos or hole:
+        parts.append(f"Hero {pos or '?'} with {hole or 'unknown'}")
+    for st in ("preflop", "flop", "turn", "river", "showdown"):
+        if args.get(st):
+            parts.append(f"{st.capitalize()}: {args[st]}")
+    if args.get("board"):
+        parts.append(f"Board: {args['board']}")
+    if args.get("result") is not None:
+        parts.append(f"Hero net: {args['result']}")
+    return ". ".join(str(p).strip() for p in parts if str(p).strip())
+
+
 def _record_hand(args: dict, ctx: dict) -> str:
+    shorthand = (args.get("shorthand") or "").strip() or _shorthand_from_fields(args)
     out = poker.record_hand(
-        args.get("shorthand") or "", stakes=args.get("stakes"),
+        shorthand, stakes=args.get("stakes"),
         tag=args.get("tag"), lesson=args.get("lesson"),
     )
     if not out["id"]:
@@ -653,6 +700,34 @@ TOOLS.update({
          "category": {**_S, "description": "feeder | risky | reg | unknown"},
          "venue": {**_S, "description": "Where they play"}},
         ["note"])},
+    "seat_players": {"handler": _seat_players, "spec": _f(
+        "seat_players",
+        "Register who's at the table this session — the roster Brian reads off the Bravo "
+        "screen (handles like TAG, JD). Call this when he names the table (usually at the "
+        "start) or when a new player sits. Each player is a real handle in `name`, or a "
+        "`descriptor` if he only describes them. These become the roster his reads/TAGs "
+        "attach to by name.",
+        {"players": {"type": "array", "description": "Players to seat",
+                     "items": {"type": "object", "properties": {
+                         "name": {**_S, "description": "Handle as it appears on Bravo, e.g. 'TAG'"},
+                         "descriptor": {**_S, "description": "Physical description if no name"},
+                         "seat": {**_S, "description": "Seat number/label if known"},
+                         "category": {**_S, "description": "feeder | risky | reg | unknown"}}}},
+         "replace": {"type": "boolean", "description": "true = a brand-new table: clear the "
+                     "current roster first, then seat these (use when he changes tables)"}},
+        ["players"])},
+    "unseat_player": {"handler": _unseat_player, "spec": _f(
+        "unseat_player",
+        "Remove a player from the table roster when they bust or leave. Keeps their history.",
+        {"name": {**_S, "description": "Their handle"},
+         "descriptor": {**_S, "description": "Or a description if unnamed"}},
+        [])},
+    "clear_table": {"handler": _clear_table, "spec": _f(
+        "clear_table",
+        "Empty the whole table roster at once — call this when Brian changes tables or says "
+        "to clear the table. The session, stack, and logged reads stay; only who's currently "
+        "seated resets. Then he'll tell you the new table.",
+        {}, [])},
     "name_villain": {"handler": _name_villain, "spec": _f(
         "name_villain",
         "Attach a real name to a player you'd only known by description (e.g. you caught it "
@@ -704,8 +779,11 @@ TOOLS.update({
         "record_hand",
         "Reconstruct a hand from Brian's rough shorthand into a structured, "
         "replayable hand history. Use when he describes/vomits a hand he wants "
-        "saved or to review. Pass his description verbatim as 'shorthand'.",
-        {"shorthand": {**_S, "description": "Brian's rough description of the hand, verbatim"},
+        "saved or to review. Pass his ENTIRE description as ONE string in `shorthand` "
+        "— do NOT split it into position/board/street fields (that's log_hand). "
+        "`shorthand` is required and must be non-empty.",
+        {"shorthand": {**_S, "description": "Brian's whole hand description as one verbatim "
+                       "string, e.g. 'UTG with 9h6h, raise 15, BTN calls, flop 8h7h5s...'"},
          "stakes": {**_S, "description": "Stakes if known, e.g. '1/3'"},
          "tag": {**_S, "description": "well_played | leak | cooler | confidence | notable"},
          "lesson": {**_S, "description": "Takeaway, if he stated one"}},

lyra/web/server.py

@@ -50,6 +50,16 @@ def _last_user_message(messages: list[dict]) -> str:
 def create_app() -> FastAPI:
     app = FastAPI(title="Lyra Web")
 
+    @app.middleware("http")
+    async def _no_stale_shell(request: Request, call_next):
+        """Always revalidate HTML/JS so a PWA can't serve a stale app shell after a
+        deploy (iOS applies heuristic caching when no cache header is set)."""
+        resp = await call_next(request)
+        ct = resp.headers.get("content-type", "")
+        if "text/html" in ct or "javascript" in ct:
+            resp.headers["Cache-Control"] = "no-cache, must-revalidate"
+        return resp
+
     @app.get("/_health")
     async def health() -> dict:
         return {"ok": True}
@@ -263,11 +273,13 @@ def create_app() -> FastAPI:
         user_msg = _last_user_message(body.get("messages", []))
 
         model_override = body.get("model") or None
+        turn_id = body.get("turnId") or None
         memory.ensure_session(session_id)
         if body.get("mode"):
             memory.set_session_mode(session_id, body["mode"])
         try:
-            reply = await asyncio.to_thread(chat.respond, session_id, user_msg, backend, model_override)
+            reply = await asyncio.to_thread(chat.respond, session_id, user_msg, backend,
+                                            model_override, turn_id)
         except Exception as exc:
             logbus.log("error", "chat failed", session=session_id, error=str(exc))
             reply = f"[error] {exc}"
@@ -295,6 +307,7 @@ def create_app() -> FastAPI:
         backend = _backend_for(body.get("backend"))
         user_msg = _last_user_message(body.get("messages", []))
         model_override = body.get("model") or None
+        turn_id = body.get("turnId") or None
         memory.ensure_session(session_id)
         if body.get("mode"):
             memory.set_session_mode(session_id, body["mode"])
@@ -306,7 +319,8 @@ def create_app() -> FastAPI:
 
             def produce():
                 try:
-                    for event in chat.respond_stream(session_id, user_msg, backend, model_override):
+                    for event in chat.respond_stream(session_id, user_msg, backend,
+                                                     model_override, turn_id):
                         loop.call_soon_threadsafe(q.put_nowait, event)
                 except Exception as exc:  # surface to the client stream, don't hang
                     logbus.log("error", "chat stream failed", session=session_id, error=str(exc))

lyra/web/static/index.html

@@ -398,10 +398,17 @@
       // live poker session forces the cloud backend regardless of the saved pick.
       if (mode === "poker_cash") backend = "cloud";
 
+      // One id per send, carried on BOTH the stream and the blocking fallback so the
+      // server runs this turn exactly once even if you lock your phone and it re-fires.
+      const turnId = (window.crypto && crypto.randomUUID)
+        ? crypto.randomUUID()
+        : String(Date.now()) + "-" + Math.random().toString(36).slice(2);
+
       const body = {
         mode: mode,
         messages: history,
-        sessionId: currentSession
+        sessionId: currentSession,
+        turnId: turnId
       };
 
       // Only add backend if in standard mode

lyra/web/static/session.html

@@ -265,6 +265,7 @@
       const stack = data.stack || {};
       const timeline = data.timeline || [];
       const hands = data.hands || [];
+      const roster = data.roster || [];
       const villains = data.villains || [];
       const notes = data.notes || [];
       const stats = data.stats || {};
@@ -369,6 +370,19 @@
             : '<p class="empty">No scars logged — mistakes to study land here.</p>'}
         </div>
 
+        <div class="card">
+          <p class="label">🪑 Table (${roster.length})</p>
+          ${roster.length ? `<ul class="rows">${roster.map(v => `
+            <li class="villain">
+              ${v.seat ? `<span class="cat">${esc(v.seat)}</span> ` : ''}<b>${esc(v.name)}</b>
+              ${v.category ? `<span class="cat">[${esc(v.category)}]</span>` : ''}
+              ${v.reads ? `<span class="cat">· ${v.reads} read${v.reads===1?'':'s'}</span>` : ''}
+              <button class="mini" title="Rename / fix" onclick="renamePlayer(${v.id}, '${esc(v.name||'').replace(/'/g,"\\'")}')">✎</button>
+              ${v.last_note ? `<div class="note-meta">“${esc(v.last_note)}”</div>` : ''}
+            </li>`).join('')}</ul>`
+            : '<p class="empty">No roster yet — tell Lyra who is at the table.</p>'}
+        </div>
+
         <div class="card">
           <p class="label">Villains seen</p>
           ${villains.length ? `<ul class="rows">${villains.map(v => `

tests/test_dream.py

@@ -20,7 +20,7 @@ def lyra(tmp_path, monkeypatch):
     # reflect() expects JSON back; everything else just stores the text.
     monkeypatch.setattr(
         llm, "complete",
-        lambda messages, backend=None, model=None:
+        lambda messages, backend=None, model=None, **_:
         '{"mood":"focused","valence":0.7,"new_reflections":["I got some thinking done."]}',
     )
 
@@ -77,3 +77,50 @@ def test_dream_cycle_consolidates_and_persists(lyra):
     state2 = dream.dream_cycle(force=False)
     assert state2["dream"]["cycle_count"] == 2
     assert state2["drives"]["continuity"] == 0.0
+
+
+def test_dream_cycle_stops_when_over_budget(lyra, monkeypatch):
+    memory = lyra
+    from lyra import dream, notify
+
+    for k in range(7):
+        _seed(memory, f"s{k}", 4)
+
+    # Go over budget right after the first heavy stage: first check passes
+    # (summarize runs), every check after trips.
+    checks = {"n": 0}
+
+    def fake_over(deadline):
+        checks["n"] += 1
+        return checks["n"] > 1
+    monkeypatch.setattr(dream, "_over_budget", fake_over)
+
+    pings: list = []
+    monkeypatch.setattr(notify, "push",
+                        lambda title, message, **k: pings.append((title, message)) or True)
+
+    state = dream.dream_cycle(force=True)
+    acts = state["dream"]["last_actions"]
+
+    assert any("stopped early" in a for a in acts)   # bailed
+    assert not any("reflected" in a for a in acts)   # later stage skipped
+    assert pings, "expected an over-budget ntfy push"
+
+
+def test_coherence_failure_does_not_sink_the_cycle(lyra, monkeypatch):
+    memory = lyra
+    from lyra import dream, profile
+
+    for k in range(3):
+        _seed(memory, f"s{k}", 4)
+
+    # A backend hiccup in the consolidation rebuild must not abort the whole pass
+    # (this is what broke the cycle when the MI50 was down).
+    monkeypatch.setattr(profile, "rebuild_profile",
+                        lambda *a, **k: (_ for _ in ()).throw(RuntimeError("backend down")))
+
+    state = dream.dream_cycle(force=True)
+    acts = state["dream"]["last_actions"]
+
+    assert any("coherence" in a and "fail" in a for a in acts)  # logged, not fatal
+    assert any("reflected" in a for a in acts)                  # cycle still reached reflection

tests/test_hand_dedup.py

@@ -0,0 +1,109 @@
+"""record_hand idempotency + straddle parse coverage.
+
+The chat turn can execute twice — the SSE stream and the blocking fallback both run
+server-side (two 'chat request' lines, 1s apart) — which double-logged the same hand
+once logging became guaranteed. A system-of-record must record an event once."""
+from __future__ import annotations
+
+import importlib
+
+import pytest
+
+
+@pytest.fixture
+def poker(tmp_path, monkeypatch):
+    monkeypatch.setenv("LYRA_DB_PATH", str(tmp_path / "test.db"))
+    from lyra import llm
+    monkeypatch.setattr(llm, "embed", lambda texts: [[0.1, 0.2, 0.3] for _ in texts])
+    import lyra.memory as memory
+    importlib.reload(memory)
+    import lyra.poker as poker
+    importlib.reload(poker)
+    return poker
+
+
+_PARSED = {
+    "game": "NLH", "hero_pos": "SB", "hero_cards": ["Ah", "Kh"],
+    "board": ["Kd", "9d", "4c", "2s"], "players": [], "actions": [],
+    "result": {"hero_net": -200, "pot": 400},
+}
+
+
+def test_record_hand_is_idempotent_across_double_execution(poker, monkeypatch):
+    sid = poker.start_session(venue="Borgata", stakes="1/3", buy_in=400)
+    monkeypatch.setattr(poker, "parse_hand", lambda *a, **k: dict(_PARSED))
+    first = poker.record_hand("i have AhKh in the SB, btn straddle, ...")
+    second = poker.record_hand("i have AhKh in the SB, btn straddle, ...")   # the duplicate turn
+    assert first["id"] == second["id"]
+    assert second.get("deduped") is True
+    assert len(poker.list_hands(sid)) == 1          # ledger holds ONE, not two
+
+
+def test_record_hand_does_not_dedupe_a_genuinely_different_hand(poker, monkeypatch):
+    sid = poker.start_session(venue="Borgata", stakes="1/3", buy_in=400)
+    monkeypatch.setattr(poker, "parse_hand", lambda *a, **k: dict(_PARSED))
+    poker.record_hand("hand one")
+    other = dict(_PARSED, hero_cards=["Qs", "Qd"], board=["Qh", "7c", "2s"])
+    monkeypatch.setattr(poker, "parse_hand", lambda *a, **k: dict(other))
+    poker.record_hand("a different hand entirely")
+    assert len(poker.list_hands(sid)) == 2          # distinct hands both land
+
+
+def test_dedupe_handles_boardless_hand(poker, monkeypatch):
+    # NULL-safe match: a preflop-only hand (no board) still dedupes.
+    sid = poker.start_session(venue="Borgata", buy_in=400)
+    preflop = {"game": "NLH", "hero_pos": "BTN", "hero_cards": ["As", "Ks"],
+               "board": [], "players": [], "actions": [], "result": {"hero_net": 30}}
+    monkeypatch.setattr(poker, "parse_hand", lambda *a, **k: dict(preflop))
+    a = poker.record_hand("AKs btn, i open everyone folds")
+    b = poker.record_hand("AKs btn, i open everyone folds")
+    assert a["id"] == b["id"] and len(poker.list_hands(sid)) == 1
+
+
+def test_parse_prompt_records_straddles():
+    from lyra import poker as pk
+    p = pk._HAND_PARSE_PROMPT.lower()
+    assert "straddle" in p and "button straddle" in p
+    assert "acts last preflop" in p or "act last preflop" in p
+
+
+# --- hero stack auto-fill from the last logged stack ----------------------
+
+def test_hero_stack_filled_from_last_stack_log(poker, monkeypatch):
+    poker.start_session(venue="Meadows", stakes="1/3", buy_in=400)
+    poker.log_stack(275)                                   # his last reported stack
+    monkeypatch.setattr(poker, "parse_hand",
+                        lambda *a, **k: {"game": "NLH", "hero_involved": True,
+                                         "hero_pos": "CO", "hero_cards": ["As", "Ks"],
+                                         "board": ["2c"], "players": [], "actions": [],
+                                         "result": {"hero_net": 50}})
+    out = poker.record_hand("AKs in the CO, i raise, flop 2c...")
+    stored = poker.get_hand(out["id"])["structured"]
+    hero = next(pl for pl in stored["players"] if pl.get("hero"))
+    assert hero["stack"] == 275 and hero.get("stack_inferred") is True
+
+
+def test_stated_stack_is_never_overridden(poker, monkeypatch):
+    poker.start_session(venue="Meadows", buy_in=400)
+    poker.log_stack(275)
+    monkeypatch.setattr(poker, "parse_hand",
+                        lambda *a, **k: {"game": "NLH", "hero_involved": True,
+                                         "hero_pos": "BTN", "hero_cards": ["Qh", "Qd"],
+                                         "players": [{"pos": "BTN", "stack": 500}],
+                                         "board": [], "actions": [], "result": {}})
+    out = poker.record_hand("500 deep on the btn with QQ")
+    hero = next(pl for pl in poker.get_hand(out["id"])["structured"]["players"]
+                if pl.get("pos") == "BTN")
+    assert hero["stack"] == 500 and not hero.get("stack_inferred")
+
+
+def test_observed_hand_gets_no_hero_stack(poker, monkeypatch):
+    poker.start_session(venue="Meadows", buy_in=400)
+    poker.log_stack(275)
+    monkeypatch.setattr(poker, "parse_hand",
+                        lambda *a, **k: {"game": "NLH", "hero_involved": False,
+                                         "hero_pos": None, "hero_cards": [],
+                                         "players": [{"pos": "CO", "cards": ["Kx", "Kx"]}],
+                                         "board": [], "actions": [], "result": {}})
+    out = poker.record_hand("the CO stacked off KK vs the nit")
+    assert all(not pl.get("stack_inferred") for pl in poker.get_hand(out["id"])["structured"]["players"])

tests/test_hand_logging.py

@@ -0,0 +1,97 @@
+"""Reliable hand logging: hero-hand guard, tool-visible history (B), forced log (A).
+
+Root cause these guard: mid-session, memory.history() rebuilt past turns as
+'hand -> narration' with tool calls stripped, few-shot-conditioning the model to
+stop logging (clean history logged 4/4, the stripped history 0/4)."""
+from __future__ import annotations
+
+from types import SimpleNamespace
+
+from lyra import poker_prompts as pp
+
+
+# --- the hero-hand guard (who gets force-logged) --------------------------
+
+def test_looks_like_hero_hand_true_for_brians_own_hand():
+    assert pp.looks_like_hero_hand("im utg with 2d2s. i raise to $15, btn calls")
+    assert pp.looks_like_hero_hand("300eff. i call btn w AsQs, flop Qh7c2s, i bet 20 he calls")
+
+
+def test_looks_like_hero_hand_false_for_observed_and_chatter():
+    # A villain the actor (no I/me/my) must never be force-logged as Brian's hand.
+    assert not pp.looks_like_hero_hand("TAG limped A4o in the SB")
+    assert not pp.looks_like_hero_hand("how's the table looking tonight?")
+    assert not pp.looks_like_hero_hand("")
+
+
+# --- Fix B: tool calls made visible in reconstructed history --------------
+
+def _ex(role, content, at):
+    return SimpleNamespace(role=role, content=content, created_at=at, id=hash(at))
+
+
+def test_history_marks_the_assistant_turn_that_logged(monkeypatch):
+    from lyra import mind, memory
+    recent = [
+        _ex("user", "i have 2d2s utg, flop 2c2hKs, quads", "2026-07-10T18:00:00.000000+00:00"),
+        _ex("assistant", "Sick cooler.", "2026-07-10T18:00:05.000000+00:00"),
+        _ex("user", "how am i doing", "2026-07-10T18:01:00.000000+00:00"),
+        _ex("assistant", "Up a grand.", "2026-07-10T18:01:03.000000+00:00"),
+    ]
+    monkeypatch.setattr(memory, "tool_events", lambda sid: [
+        {"tool": "record_hand", "result": "Hand #62 logged — UTG 2d2s.",
+         "created_at": "2026-07-10T18:00:03.000000+00:00"},
+        {"tool": "session_state", "result": "net +1000",
+         "created_at": "2026-07-10T18:01:02.000000+00:00"},
+    ])
+    msgs = mind._history_with_tools("s1", recent)
+    # each event is attributed to the assistant turn whose window it falls in
+    assert "record_hand → Hand #62 logged" in msgs[1]["content"]
+    assert msgs[1]["content"].endswith("Sick cooler.")
+    assert "session_state" in msgs[3]["content"]
+    # user turns are untouched
+    assert msgs[0]["content"] == recent[0].content
+
+
+def test_history_no_marker_when_no_tools(monkeypatch):
+    from lyra import mind, memory
+    monkeypatch.setattr(memory, "tool_events", lambda sid: [])
+    recent = [_ex("assistant", "just talking", "2026-07-10T18:00:05.000000+00:00")]
+    assert mind._history_with_tools("s1", recent)[0]["content"] == "just talking"
+
+
+# --- Fix A: force the log when the model skipped a hero hand ---------------
+
+def _force_setup(monkeypatch, tool_calls):
+    from lyra import chat
+    monkeypatch.setattr(chat.llm, "chat_call",
+                        lambda *a, **k: ({"role": "assistant"}, tool_calls))
+    dispatched = []
+    monkeypatch.setattr(chat.toolkit, "dispatch",
+                        lambda name, args, ctx=None: dispatched.append(name) or "Hand #71 logged.")
+    monkeypatch.setattr(chat.memory, "add_tool_event", lambda *a, **k: 1)
+    return chat, dispatched
+
+
+def test_forces_log_on_unlogged_hero_hand(monkeypatch):
+    chat, dispatched = _force_setup(monkeypatch, [{"id": "1", "name": "record_hand",
+                                                   "arguments": '{"shorthand":"AsQs..."}'}])
+    forced = chat._ensure_hand_logged([], "300eff i call btn w AsQs, i bet 20", "HAND", [],
+                                      "cloud", None, {}, "s1")
+    assert forced == ["record_hand"] and dispatched == ["record_hand"]
+
+
+def test_does_not_force_when_already_logged(monkeypatch):
+    chat, dispatched = _force_setup(monkeypatch, [])
+    forced = chat._ensure_hand_logged([], "i have AsQs, i bet", "HAND", ["record_hand"],
+                                      "cloud", None, {}, "s1")
+    assert forced == [] and dispatched == []
+
+
+def test_does_not_force_non_hand_or_observed(monkeypatch):
+    chat, dispatched = _force_setup(monkeypatch, [])
+    # not a HAND turn
+    assert chat._ensure_hand_logged([], "down to 220", "LOG", [], "cloud", None, {}, "s1") == []
+    # HAND-classified but observed (no first person) → never force-logged as his
+    assert chat._ensure_hand_logged([], "TAG shoved AKo", "HAND", [], "cloud", None, {}, "s1") == []
+    assert dispatched == []

tests/test_hand_tools.py

@@ -0,0 +1,55 @@
+"""record_hand tolerance: recover when the model calls it with log_hand's fields."""
+from __future__ import annotations
+
+from lyra import tools
+
+_GRANULAR = {
+    "position": "UTG", "hole_cards": "9h6h", "board": "8h7h5s 5h Kc",
+    "preflop": "raised to 15, BTN calls", "flop": "bet 25, BTN calls",
+    "turn": "bet 50, BTN raises to 150, call", "river": "check, BTN all in, snap call",
+    "showdown": "BTN shows 55 for quads, hero shows straight flush", "result": 300,
+    "tag": "notable", "lesson": "rare straight flush over quads",
+}
+
+
+def test_shorthand_from_fields_builds_a_parseable_description():
+    s = tools._shorthand_from_fields(_GRANULAR)
+    assert "UTG with 9h6h" in s
+    assert "Preflop:" in s and "River:" in s and "Board: 8h7h5s 5h Kc" in s
+    assert "Hero net: 300" in s
+
+
+def test_record_hand_recovers_from_granular_fields(monkeypatch):
+    # The model called record_hand with log_hand's schema (no `shorthand`). The
+    # handler must reconstruct one and pass it to poker.record_hand, not fail empty.
+    seen = {}
+
+    def fake_record_hand(shorthand, stakes=None, tag=None, lesson=None, backend=None):
+        seen["shorthand"] = shorthand
+        return {"id": 42, "parsed": {"hero_involved": True, "hero_pos": "UTG",
+                                     "hero_cards": ["9h", "6h"]}, "linked": 0}
+
+    monkeypatch.setattr(tools.poker, "record_hand", fake_record_hand)
+    out = tools.dispatch("record_hand", _GRANULAR, {})
+    assert "UTG with 9h6h" in seen["shorthand"]      # reconstructed, not empty
+    assert "#42" in out and "couldn't parse" not in out
+
+
+def test_record_hand_still_prefers_explicit_shorthand(monkeypatch):
+    seen = {}
+
+    def fake_record_hand(shorthand, stakes=None, tag=None, lesson=None, backend=None):
+        seen["shorthand"] = shorthand
+        return {"id": 7, "parsed": {"hero_involved": True, "hero_pos": "BTN",
+                                    "hero_cards": ["As", "Ks"]}, "linked": 0}
+
+    monkeypatch.setattr(tools.poker, "record_hand", fake_record_hand)
+    tools.dispatch("record_hand", {"shorthand": "BTN AKs, I open, everyone folds"}, {})
+    assert seen["shorthand"] == "BTN AKs, I open, everyone folds"   # verbatim, not rebuilt
+
+
+def test_record_hand_empty_call_still_fails_gracefully(monkeypatch):
+    monkeypatch.setattr(tools.poker, "record_hand",
+                        lambda *a, **k: {"id": None, "parsed": None})
+    out = tools.dispatch("record_hand", {}, {})
+    assert "couldn't parse" in out.lower()

tests/test_llm_bounds.py

@@ -0,0 +1,110 @@
+"""llm.complete: `max_tokens` and `timeout` are threaded into the backend call.
+
+The OpenAI client is faked so nothing hits a network. We assert the generation
+cap reaches the create() call and the fast-fail timeout reaches the client (with
+max_retries=0 so summary.py owns the retry policy, not the SDK).
+"""
+from __future__ import annotations
+
+import types
+
+import pytest
+
+from lyra import llm
+
+
+@pytest.fixture
+def fake_openai(monkeypatch):
+    recorded: dict = {}
+
+    class FakeCompletions:
+        def create(self, **kwargs):
+            recorded["create"] = kwargs
+            msg = types.SimpleNamespace(content="ok")
+            return types.SimpleNamespace(choices=[types.SimpleNamespace(message=msg)])
+
+    class FakeClient:
+        def __init__(self, **kwargs):
+            recorded["client"] = kwargs
+            self.chat = types.SimpleNamespace(completions=FakeCompletions())
+
+    monkeypatch.setattr(llm, "OpenAI", FakeClient)
+    monkeypatch.setattr(llm, "load", lambda: types.SimpleNamespace(
+        mi50_base_url="http://mi50/v1", mi50_model="local-gpu",
+        cloud_model="gpt-4o-mini", openai_api_key="sk-test", local_model="l",
+    ))
+    return recorded
+
+
+def test_mi50_threads_max_tokens_and_timeout(fake_openai):
+    out = llm.complete([{"role": "user", "content": "hi"}],
+                       backend="mi50", max_tokens=768, timeout=150)
+
+    assert out == "ok"
+    assert fake_openai["create"]["max_tokens"] == 768
+    assert fake_openai["client"]["timeout"] == 150
+    assert fake_openai["client"]["max_retries"] == 0
+
+
+def test_cloud_threads_max_tokens_and_timeout(fake_openai):
+    llm.complete([{"role": "user", "content": "hi"}],
+                 backend="cloud", max_tokens=768, timeout=150)
+
+    assert fake_openai["create"]["max_tokens"] == 768
+    assert fake_openai["client"]["timeout"] == 150
+    assert fake_openai["client"]["max_retries"] == 0
+
+
+def test_fallback_uses_primary_when_it_succeeds(monkeypatch):
+    seen = []
+    monkeypatch.setattr(llm, "complete",
+                        lambda messages, backend="local", model=None, **k:
+                        seen.append(backend) or f"{backend}-ok")
+    out = llm.complete_with_fallback([{"role": "user", "content": "x"}],
+                                     backend="local", model="dolphin3:8b")
+    assert out == "local-ok"
+    assert seen == ["local"]  # no fallback when the primary works
+
+
+def test_fallback_to_cloud_when_primary_errors(monkeypatch):
+    monkeypatch.setattr(llm, "load", lambda: types.SimpleNamespace(openai_api_key="sk"))
+    seen = []
+
+    def fake(messages, backend="local", model=None, **k):
+        seen.append(backend)
+        if backend == "local":
+            raise RuntimeError("3090 is powered off")
+        return "cloud-ok"
+    monkeypatch.setattr(llm, "complete", fake)
+
+    out = llm.complete_with_fallback([{"role": "user", "content": "x"}],
+                                     backend="local", model="dolphin3:8b")
+    assert out == "cloud-ok"
+    assert seen == ["local", "cloud"]
+
+
+def test_fallback_reraises_when_primary_is_already_cloud(monkeypatch):
+    monkeypatch.setattr(llm, "load", lambda: types.SimpleNamespace(openai_api_key="sk"))
+    monkeypatch.setattr(llm, "complete",
+                        lambda *a, **k: (_ for _ in ()).throw(RuntimeError("boom")))
+    with pytest.raises(RuntimeError):
+        llm.complete_with_fallback([{"role": "user", "content": "x"}], backend="cloud")
+
+
+def test_fallback_reraises_without_openai_key(monkeypatch):
+    monkeypatch.setattr(llm, "load", lambda: types.SimpleNamespace(openai_api_key=""))
+    monkeypatch.setattr(llm, "complete",
+                        lambda *a, **k: (_ for _ in ()).throw(RuntimeError("down")))
+    with pytest.raises(RuntimeError):
+        llm.complete_with_fallback([{"role": "user", "content": "x"}], backend="local")
+
+
+def test_default_bounds_calls_even_without_explicit_timeout(fake_openai):
+    # No cap / timeout passed -> still bounded: 300s default + no SDK retries, so
+    # no call can silently inherit the SDK's 600s x2 (~30 min). No length cap
+    # unless asked, though.
+    llm.complete([{"role": "user", "content": "hi"}], backend="mi50")
+
+    assert "max_tokens" not in fake_openai["create"]
+    assert fake_openai["client"]["timeout"] == 300
+    assert fake_openai["client"]["max_retries"] == 0

tests/test_perceive.py

@@ -53,4 +53,53 @@ def test_route_injects_tilt_nudge(mind):
 def test_route_quiet_on_neutral_turn(mind):
     turn = mind.assemble("s1", "what did we decide about the schema yesterday?", "cloud", None)
     assert turn.register is None                    # neutral -> no nudge
-    assert not (turn.moment or {}).get("note")
+    assert not (turn.moment or {}).get("note")
+
+
+# --- Phase A pipeline fixes: poker mode suppresses the two mush sources ---
+
+def test_poker_mode_suppresses_tilt_nudge(mind):
+    from lyra import memory
+    memory.set_session_mode("s1", "poker_cash")
+    turn = mind.assemble("s1", "ugh I'm steaming, fucking coolered again!!", "cloud", None)
+    assert turn.register is None                     # at the table, register comes from fragments
+    sys_blob = " ".join(m["content"] for m in turn.messages if m["role"] == "system")
+    assert "on tilt" not in sys_blob.lower()         # false-positive lexicon nudge suppressed
+
+
+def test_mode_menu_note_suppressed_in_poker(mind):
+    from lyra import modes
+    poker = " ".join(m["content"] for m in mind.build_messages("s1", "stack 350", mode=modes.CASH)
+                     if m["role"] == "system")
+    build = " ".join(m["content"] for m in mind.build_messages("s1", "let's refactor", mode=modes.get("build"))
+                     if m["role"] == "system")
+    assert "Your modes:" not in poker                # no "offer to switch" note at the table
+    assert "Your modes:" in build                    # still present in a non-poker mode
+
+
+# --- Phase B: sharded poker prompt (BASE + one fragment, no monolith) ---
+
+def _poker_blob(mind, msg):
+    from lyra import modes
+    return " ".join(m["content"] for m in mind.build_messages("s1", msg, mode=modes.CASH)
+                    if m["role"] == "system")
+
+
+def test_poker_injects_base_plus_the_matching_fragment(mind):
+    blob = _poker_blob(mind, "I flopped a set with 99 on 9h4c2d and bet the turn")
+    assert "LOG FIRST" in blob                       # BASE is always on in poker
+    assert "MESSAGE TYPE: HAND" in blob              # the fragment for THIS message
+    assert "MESSAGE TYPE: STATUS" not in blob        # and not the others
+    assert "MESSAGE TYPE: READ" not in blob
+
+
+def test_poker_fragment_changes_with_message_type(mind):
+    status = _poker_blob(mind, "it's 11:50pm, waiting for a seat")
+    assert "MESSAGE TYPE: STATUS" in status and "MESSAGE TYPE: HAND" not in status
+
+
+def test_poker_monolith_no_longer_injected(mind):
+    from lyra import modes
+    blob = _poker_blob(mind, "stack 350")
+    assert "You move between two registers" not in blob   # the old _CASH_CARD opener is gone
+    assert modes.CASH.card == ""                           # card sharded out

tests/test_poker_prompts.py

@@ -0,0 +1,137 @@
+"""Poker-mode message classifier + fragment selection (pure, no DB)."""
+from __future__ import annotations
+
+from lyra import poker_prompts as pp
+
+
+def c(msg, roster=()):
+    return pp.classify(msg, roster)
+
+
+# --- the spec's canonical cases ---
+
+def test_read_villain_action_beats_hand():
+    # A villain's action carries cards+position+verb but is NOT Brian's hand.
+    assert c("TAG limped A4o in the SB (UTG straddled)") == "READ"
+    assert c("Jonathan called the 3bet") == "READ"
+    assert c("the neck-tattoo guy shoved the turn") == "READ"
+
+
+def test_hand_is_first_person():
+    assert c("Button straddle on. I limp UTG with 22. Flop 2d7cjh, I check-raise") == "HAND"
+    assert c("I flopped a set with 99 on 9h4c2d and bet the turn") == "HAND"
+
+
+def test_hand_narrated_without_I_still_hand_not_read():
+    # No "I", but leads with a poker verb (not a name) + street/action → his hand.
+    assert c("Flopped bottom set with 22, bet $40 on the river, he folded 88") == "HAND"
+
+
+def test_table_ops():
+    assert c("seat the table: TAG, Jonathan, Wheelz") == "TABLE"
+    assert c("table broke, I'm at a new table") == "TABLE"
+    assert c("I got moved to another table") == "TABLE"
+
+
+def test_mental():
+    assert c("I feel like I'm being mean when I raise") == "MENTAL"
+    assert c("ugh I'm so tilted, card dead all night") == "MENTAL"
+
+
+def test_status_is_not_a_mood():
+    assert c("it's 11:50pm, waiting for a seat") == "STATUS"
+    assert c("grabbing food, be right back") == "STATUS"
+
+
+def test_log_bare_money():
+    assert c("I'm at 317 now") == "LOG"
+    assert c("stack is 540") == "LOG"
+
+
+def test_chat_default():
+    assert c("should I have folded the river?") == "CHAT"
+    assert c("what do you think of this table so far") == "CHAT"
+
+
+# --- the READ vs HAND boundary (the hard one) ---
+
+def test_roster_handle_forces_read():
+    # A seated handle as the actor → READ even if lowercase / plain.
+    assert c("tag opened to 15 from the cutoff", roster=("TAG",)) == "READ"
+
+
+def test_first_person_action_stays_hand_even_with_roster():
+    # Brian is the actor → HAND, not a read on a seated player mentioned nearby.
+    assert c("I 3bet TAG's open with AKs", roster=("TAG",)) == "HAND"
+
+
+def test_all_caps_handle_reads_without_roster():
+    assert c("JD min-raised the button") == "READ"
+
+
+# --- fragment selection ---
+
+def test_fragment_for_maps_each_type():
+    for t in pp.MSG_TYPES:
+        assert pp.fragment_for(t) is pp.FRAGMENTS[t]
+    assert pp.fragment_for(None) is pp.FRAGMENTS["CHAT"]
+    assert pp.fragment_for("bogus") is pp.FRAGMENTS["CHAT"]
+
+
+def test_base_is_nonempty_and_names_the_hard_rules():
+    assert "LOG FIRST" in pp.BASE
+    assert "descriptor" in pp.BASE and "session_state" in pp.BASE
+
+
+# --- hardening: real-world phrasings that used to miss ---
+
+def test_hardening_reads_ing_and_bare_descriptor():
+    assert c("TAG's been limping every pot", roster=("TAG",)) == "READ"   # -ing form
+    assert c("the whale called again") == "READ"                          # bare "the <noun>"
+    assert c("saw JD open utg") == "READ"
+
+
+def test_hardening_player_departures_are_table():
+    assert c("TAG busted") == "TABLE"
+    assert c("TAG left the table") == "TABLE"
+    assert c("new guy just sat down") == "TABLE"
+
+
+def test_hardening_questions_never_log():
+    # "stack" appears but it's a strategy question, not a stack update.
+    assert c("should I stack off top set on that board?") == "CHAT"
+    assert c("was I good to call there with AK?") == "CHAT"
+
+
+def test_hardening_mental_lexicon():
+    assert c("im getting coolered every hand, so sick of this") == "MENTAL"
+    assert c("this is brutal, run so bad") == "MENTAL"
+
+
+def test_hardening_log_needs_number_or_result_word():
+    assert c("down to 220") == "LOG"
+    assert c("sitting on 450 now") == "LOG"
+    assert c("rebought for 300") == "LOG"
+    # first-person departure is Brian, not a roster op → not TABLE
+    assert c("I busted, heading home") != "TABLE"
+
+
+# --- HAND fragment: route showdowns to the tool + no motivational mush ---
+
+def test_hand_fragment_routes_showdowns_to_the_tool():
+    # A resolved showdown must be verified via analyze_spot, not eyeballed
+    # (the quad-kings-read-as-"kings-full" regression).
+    frag = pp.fragment_for("HAND")
+    assert "SHOWDOWN" in frag
+    assert "analyze_spot" in frag
+    assert "never eyeball" in frag.lower()
+    # names the hand class by street so "flopped quads" actually gets said
+    assert "street it mattered" in frag
+
+
+def test_hand_fragment_bans_motivational_filler():
+    frag = pp.fragment_for("HAND")
+    assert "variance-evens-out" in frag
+    assert "life-lesson" in frag
+    # if there's no leak, say so instead of inventing a takeaway
+    assert "no leak" in frag.lower()

tests/test_reflect.py

@@ -28,7 +28,7 @@ def lyra(tmp_path, monkeypatch):
 
     calls = []
 
-    def fake_complete(messages, backend=None, model=None):
+    def fake_complete(messages, backend=None, model=None, **_):
         calls.append(messages)
         # the examine step's system prompt is the one asking for self_critique
         is_examine = "self_critique" in messages[0]["content"]
@@ -69,7 +69,7 @@ def test_reflect_revises_and_records_critique(lyra):
 def test_reflect_falls_back_to_draft_if_examine_unparseable(lyra, monkeypatch):
     from lyra import llm, self_state
 
-    def only_draft(messages, backend=None, model=None):
+    def only_draft(messages, backend=None, model=None, **_):
         return DRAFT if "self_critique" not in messages[0]["content"] else "not json at all"
 
     monkeypatch.setattr(llm, "complete", only_draft)
@@ -87,7 +87,7 @@ def test_consolidation_rebuilds_narrative_from_reflections(lyra, monkeypatch):
                          "I wondered what the quiet is for"]
     memory.set_self_state(st)
 
-    def comp(messages, backend=None, model=None):
+    def comp(messages, backend=None, model=None, **_):
         # consolidation should synthesize from anchor + reflections, not the old bio
         assert "supportive presence devoted to Brian" not in messages[1]["content"]
         return ('{"self_narrative":"I am Lyra, and lately I have been restless and curious '

tests/test_roster.py

@@ -0,0 +1,101 @@
+"""Live table roster: seat players, attach reads by handle, roster on the HUD."""
+from __future__ import annotations
+
+import importlib
+
+import numpy as np
+import pytest
+
+
+def _fake_embed(texts):
+    out = []
+    for t in texts:
+        v = np.zeros(64, dtype=np.float32)
+        for w in t.lower().split():
+            v[hash(w) % 64] += 1.0
+        out.append((v if v.any() else np.full(64, 1e-6, dtype=np.float32)).tolist())
+    return out
+
+
+@pytest.fixture
+def mods(tmp_path, monkeypatch):
+    monkeypatch.setenv("LYRA_DB_PATH", str(tmp_path / "test.db"))
+    from lyra import llm
+    monkeypatch.setattr(llm, "embed", _fake_embed)
+    import lyra.memory as memory
+    importlib.reload(memory)
+    import lyra.poker as poker
+    importlib.reload(poker)
+    import lyra.tools as tools
+    importlib.reload(tools)
+    return poker, tools
+
+
+def test_seat_players_builds_roster(mods):
+    poker, _ = mods
+    poker.start_session(venue="Meadows", buy_in=300)
+    n = poker.seat_players(["TAG", "Jonathan", {"name": "Wheelz", "seat": "3"}])
+    assert n == 3
+    roster = poker.session_roster()
+    names = {r["name"] for r in roster}
+    assert names == {"TAG", "Jonathan", "Wheelz"}
+    assert next(r for r in roster if r["name"] == "Wheelz")["seat"] == "3"
+
+
+def test_read_attaches_to_seated_player_by_handle(mods):
+    poker, _ = mods
+    poker.start_session(venue="Meadows", buy_in=300)
+    poker.seat_players(["TAG"])
+    poker.add_read(note="limped A4o from the SB, UTG straddle", name="TAG")
+    roster = poker.session_roster()
+    tag = next(r for r in roster if r["name"] == "TAG")
+    assert tag["reads"] == 1 and "A4o" in tag["last_note"]
+    # No duplicate TAG spawned — the read landed on the seated player.
+    assert sum(p["name"] == "TAG" for p in poker.get_villain_file()) == 1
+
+
+def test_seat_players_tool_and_roster_in_hud(mods):
+    poker, tools = mods
+    poker.start_session(venue="Meadows", buy_in=300)
+    out = tools.dispatch("seat_players", {"players": [{"name": "TAG"}, {"name": "JD"}]}, {})
+    assert "TAG" in out and "JD" in out
+    assert len(poker.hud()["roster"]) == 2
+
+
+def test_unseat_player_removes_from_roster_keeps_history(mods):
+    poker, _ = mods
+    poker.start_session(venue="Meadows", buy_in=300)
+    poker.seat_players(["TAG"])
+    poker.add_read(note="showed a bluff", name="TAG")
+    assert poker.unseat_player(name="TAG") is True
+    assert poker.session_roster() == []                 # off the table
+    assert poker.player_profile("TAG")["reads"]         # history intact
+
+
+def test_clear_table_empties_roster_keeps_reads(mods):
+    poker, tools = mods
+    poker.start_session(venue="Meadows", buy_in=300)
+    poker.seat_players(["TAG", "Jonathan"])
+    poker.add_read(note="limped A4o", name="TAG")
+    out = tools.dispatch("clear_table", {}, {})
+    assert "cleared" in out.lower()
+    assert poker.session_roster() == []                     # roster emptied
+    assert poker.player_profile("TAG")["reads"]             # reads kept
+    # A live session is untouched by clearing the table.
+    assert poker.live_session() is not None
+
+
+def test_seat_players_replace_swaps_to_new_table(mods):
+    poker, tools = mods
+    poker.start_session(venue="Meadows", buy_in=300)
+    poker.seat_players(["TAG", "Jonathan"])
+    tools.dispatch("seat_players", {"players": [{"name": "Doyle"}, {"name": "Ivey"}],
+                                    "replace": True}, {})
+    assert {r["name"] for r in poker.session_roster()} == {"Doyle", "Ivey"}
+
+
+def test_seat_players_accepts_plain_name_list_via_tool(mods):
+    poker, tools = mods
+    poker.start_session(venue="Meadows", buy_in=300)
+    tools.dispatch("seat_players", {"players": "TAG, JD, Wheelz"}, {})
+    assert {r["name"] for r in poker.session_roster()} == {"TAG", "JD", "Wheelz"}

tests/test_summary_fallback.py

@@ -0,0 +1,142 @@
+"""Summary consolidation: MI50 length cap, fast-fail, and cloud fallback.
+
+Everything is stubbed — no real backend is touched. These drive the behavior of
+`summary._summarize_text`: try the primary backend a bounded number of times with
+a capped generation length, and fall back to cloud if the primary keeps failing.
+"""
+from __future__ import annotations
+
+import types
+
+import pytest
+
+from lyra import summary
+
+
+@pytest.fixture
+def calls(monkeypatch):
+    """Capture every llm.complete call; per-test behavior via `fake.responder`."""
+    recorded: list[dict] = []
+
+    def fake_complete(messages, backend="local", model=None,
+                      max_tokens=None, timeout=None):
+        recorded.append({"backend": backend, "max_tokens": max_tokens, "timeout": timeout})
+        return fake_complete.responder(backend)
+
+    fake_complete.responder = lambda backend: "gist"
+    monkeypatch.setattr(summary.llm, "complete", fake_complete)
+    monkeypatch.setattr(summary.time, "sleep", lambda *_: None)  # instant backoff
+    return types.SimpleNamespace(recorded=recorded, fake=fake_complete)
+
+
+def _set_key(monkeypatch, key="sk-test"):
+    monkeypatch.setattr(summary.config, "load",
+                        lambda: types.SimpleNamespace(openai_api_key=key))
+
+
+def test_falls_back_to_cloud_after_mi50_attempts(calls, monkeypatch):
+    _set_key(monkeypatch)
+
+    def responder(backend):
+        if backend == "mi50":
+            raise RuntimeError("Request timed out.")
+        return "cloud-gist"
+    calls.fake.responder = responder
+
+    out = summary._summarize_text("transcript", "mi50")
+
+    assert out == "cloud-gist"
+    assert [c["backend"] for c in calls.recorded] == ["mi50", "mi50", "cloud"]
+
+
+def test_no_fallback_when_backend_is_cloud(calls, monkeypatch):
+    _set_key(monkeypatch)
+    calls.fake.responder = lambda backend: (_ for _ in ()).throw(RuntimeError("boom"))
+
+    with pytest.raises(RuntimeError):
+        summary._summarize_text("t", "cloud")
+
+    # Cloud is already the primary: retry it, but never a redundant fallback.
+    assert [c["backend"] for c in calls.recorded] == ["cloud", "cloud"]
+
+
+def test_no_fallback_without_openai_key(calls, monkeypatch):
+    _set_key(monkeypatch, key="")
+    calls.fake.responder = lambda backend: (_ for _ in ()).throw(RuntimeError("mi50 down"))
+
+    with pytest.raises(RuntimeError):
+        summary._summarize_text("t", "mi50")
+
+    assert [c["backend"] for c in calls.recorded] == ["mi50", "mi50"]
+
+
+def test_caps_length_and_timeout_on_every_call(calls, monkeypatch):
+    _set_key(monkeypatch)
+
+    def responder(backend):
+        if backend == "mi50":
+            raise RuntimeError("nope")
+        return "cloud-gist"
+    calls.fake.responder = responder
+
+    summary._summarize_text("t", "mi50")
+
+    assert calls.recorded
+    for c in calls.recorded:
+        assert c["max_tokens"] == summary.SUMMARY_MAX_TOKENS
+        assert c["timeout"] == summary.SUMMARY_TIMEOUT
+
+
+def test_happy_path_uses_primary_only(calls, monkeypatch):
+    _set_key(monkeypatch)
+    calls.fake.responder = lambda backend: "mi50-gist"
+
+    out = summary._summarize_text("t", "mi50")
+
+    assert out == "mi50-gist"
+    assert [c["backend"] for c in calls.recorded] == ["mi50"]  # no retries, no fallback
+
+
+# --- degenerate ("?" garbage) output guard: a wedged local model returns junk as
+# a successful 200, so treat it as a failure and fall back to cloud. ---
+
+def test_looks_degenerate_flags_repeated_char():
+    assert summary._looks_degenerate("?" * 60) is True
+    assert summary._looks_degenerate("!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!") is True
+
+
+def test_looks_degenerate_passes_real_prose():
+    gist = ("Brian sat down at the Meadows 1/3 in seat 6 with two straddles active; "
+            "he tagged a seat-3 calling station and finished the session up 240.")
+    assert summary._looks_degenerate(gist) is False
+
+
+def test_looks_degenerate_ignores_short_output():
+    # Too short to judge — don't false-positive a terse-but-valid reply.
+    assert summary._looks_degenerate("ok") is False
+
+
+def test_degenerate_mi50_output_falls_back_to_cloud(calls, monkeypatch):
+    _set_key(monkeypatch)
+
+    def responder(backend):
+        if backend == "mi50":
+            return "?" * 200          # garbage-as-200, not an exception
+        return "a real cloud gist of the session, diverse and coherent."
+    calls.fake.responder = responder
+
+    out = summary._summarize_text("transcript", "mi50")
+
+    assert "cloud gist" in out
+    assert [c["backend"] for c in calls.recorded] == ["mi50", "mi50", "cloud"]
+
+
+def test_degenerate_cloud_output_raises_no_infinite_loop(calls, monkeypatch):
+    _set_key(monkeypatch)
+    calls.fake.responder = lambda backend: "?" * 200  # every backend returns garbage
+
+    with pytest.raises(Exception):
+        summary._summarize_text("t", "mi50")
+
+    # mi50 x2, then one cloud fallback that's also garbage -> give up, no loop.
+    assert [c["backend"] for c in calls.recorded] == ["mi50", "mi50", "cloud"]

tests/test_thoughts.py

@@ -31,7 +31,7 @@ def lyra(tmp_path, monkeypatch):
     # Canned LLM: tests set `box["next"]` to the dict think() should "generate".
     box = {"next": {}}
     monkeypatch.setattr(thoughts.llm, "complete",
-                        lambda messages, backend=None, model=None: json.dumps(box["next"]))
+                        lambda messages, backend=None, model=None, **_: json.dumps(box["next"]))
     # Keep the loop offline + silent by default: no feed fetch, no push.
     monkeypatch.setattr(thoughts.feeds, "next_item", lambda **k: None)
     monkeypatch.setattr(thoughts.notify, "push", lambda **k: False)
@@ -342,7 +342,7 @@ def test_think_routes_to_selected_voice(lyra, monkeypatch):
     self_state.set_introspection_mode("dolphin")
     seen = {}
 
-    def cap(messages, backend="local", model=None):
+    def cap(messages, backend="local", model=None, **_):
         seen["backend"], seen["model"] = backend, model
         return json.dumps(box["next"])
 

tests/test_turn_dedup.py

@@ -0,0 +1,93 @@
+"""Turn de-duplication: the UI hits two endpoints for one message (SSE stream +
+blocking fallback). Only the first should execute; the duplicate reuses its result."""
+from __future__ import annotations
+
+import pytest
+
+from lyra import chat
+
+
+@pytest.fixture(autouse=True)
+def _clean_turns():
+    chat._turns.clear()
+    yield
+    chat._turns.clear()
+
+
+def test_claim_is_owner_once_per_key():
+    o1, r1 = chat._claim_turn("s1", "flopped a set")
+    o2, r2 = chat._claim_turn("s1", "flopped a set")
+    assert o1 is True and o2 is False
+    assert r1 is r2                       # the duplicate waits on the SAME record
+
+
+def test_different_messages_each_own():
+    o1, _ = chat._claim_turn("s1", "hand A")
+    o2, _ = chat._claim_turn("s1", "hand B")
+    o3, _ = chat._claim_turn("s2", "hand A")     # different session
+    assert o1 and o2 and o3
+
+
+def test_await_returns_owner_reply():
+    _, rec = chat._claim_turn("s1", "msg")
+    chat._finish_turn(rec, "the answer")
+    assert chat._await_duplicate(rec) == "the answer"
+
+
+def test_respond_duplicate_reuses_result_without_running_turn(monkeypatch):
+    # owner already ran and cached its reply
+    _, rec = chat._claim_turn("s1", "same hand")
+    chat._finish_turn(rec, "owner reply")
+
+    def boom(*a, **k):
+        raise AssertionError("duplicate must NOT execute the turn")
+    monkeypatch.setattr(chat.mind, "assemble", boom)
+
+    out = chat.respond("s1", "same hand", "cloud")
+    assert out == "owner reply"
+
+
+def test_respond_stream_duplicate_yields_cached_reply(monkeypatch):
+    _, rec = chat._claim_turn("s1", "same hand")
+    chat._finish_turn(rec, "owner reply")
+
+    def boom(*a, **k):
+        raise AssertionError("duplicate must NOT execute the turn")
+    monkeypatch.setattr(chat.mind, "assemble", boom)
+
+    events = list(chat.respond_stream("s1", "same hand", "cloud"))
+    assert ("delta", "owner reply") in events
+    assert ("done", "owner reply") in events
+
+
+def test_fresh_message_after_window_runs_again():
+    # a completed turn lingers only briefly; simulate expiry and confirm re-ownership
+    o1, rec = chat._claim_turn("s1", "later resend")
+    chat._finish_turn(rec, "first")
+    rec["ts"] -= chat._TURN_TTL_MSG + 1       # age it past the (session,msg) window
+    o2, _ = chat._claim_turn("s1", "later resend")
+    assert o1 and o2                          # a genuine later resend runs fresh
+
+
+# --- client turn-id keying (the fire-and-forget guarantee) ----------------
+
+def test_same_turn_id_dedupes_regardless_of_message():
+    # the fallback may resend the SAME id; dedupe on the id, not the text
+    o1, r1 = chat._claim_turn("s1", "a hand", turn_id="tid-123")
+    o2, r2 = chat._claim_turn("s1", "a hand", turn_id="tid-123")
+    assert o1 is True and o2 is False and r1 is r2
+
+
+def test_different_turn_ids_each_own():
+    o1, _ = chat._claim_turn("s1", "same text", turn_id="tid-1")
+    o2, _ = chat._claim_turn("s1", "same text", turn_id="tid-2")
+    assert o1 and o2                          # a genuinely new send never gets swallowed
+
+
+def test_turn_id_window_survives_long_after_the_msg_window():
+    # locked-phone case: the re-fire can arrive minutes later and must still dedupe
+    o1, rec = chat._claim_turn("s1", "big hand", turn_id="tid-lock")
+    chat._finish_turn(rec, "cached")
+    rec["ts"] -= chat._TURN_TTL_MSG + 60      # well past the short window, but not the id window
+    o2, r2 = chat._claim_turn("s1", "big hand", turn_id="tid-lock")
+    assert o1 and o2 is False and r2["reply"] == "cached"

tests/test_villain_tools.py

@@ -46,6 +46,24 @@ def test_descriptor_read_creates_then_reuses_nameless_villain(mods):
     assert reads == 2
 
 
+def test_description_as_name_routes_to_descriptor_and_dedupes(mods):
+    poker, tools = mods
+    poker.start_session(venue="Meadows", buy_in=300)
+    # She (wrongly) puts a physical description in the name field, twice, worded
+    # slightly differently — must resolve to ONE nameless villain, not two named.
+    tools.dispatch("add_read", {"note": "limp 3bet A3o",
+                                "name": "Filipino, Fox Racing hat, DKNY shirt, two bracelets"}, {})
+    tools.dispatch("add_read", {"note": "called a 4bet light",
+                                "name": "Filipino, Fox Racing hat, DKNY shirt, watch on left"}, {})
+    named = [p for p in poker.get_villain_file() if p["named"]]
+    assert named == []                    # no sentence-named players spawned (the bug)
+    # Either they merged, or the near-dup is surfaced for a one-click merge — never
+    # a silent duplicate the way sentence-names were.
+    q = poker.list_identity_queue()
+    nameless = [p for p in poker.get_villain_file() if not p["named"]]
+    assert len(nameless) == 1 or any(t["kind"] == "merge_candidate" for t in q)
+
+
 def test_name_villain_tool_attaches_name(mods):
     poker, tools = mods
     poker.start_session(venue="Meadows", buy_in=300)