Each user and assistant message gets a copy button (⧉ → ✓) that puts the whole
message on the clipboard. Assistant copies the raw markdown (its dataset.raw);
user copies its text.
- copyToClipboard uses the async Clipboard API when available and falls back to a
hidden-textarea + execCommand with an explicit iOS selection range, so it works
on the iPhone PWA over plain-HTTP LAN (no secure context).
- Copy sits in the assistant rate-bar and in a right-aligned bar on user bubbles;
tools stay visible on touch (@media hover:none) since there's no hover on iOS.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Answers three gaps: no way to delete a single poker session (only clear_all),
no way to browse past sessions, and Lyra could only see aggregate stats.
- poker.list_sessions() (per-session summary + hand count + recap flag) and
poker.delete_session() (removes a session + its hands/reads/observations/
stacks/rituals; keeps the persistent villain file).
- /history page (date, stakes, venue, net, hours, recap link, per-row delete with
confirm) + /history/data + DELETE /history/{id}. Nav links from chat + HUD.
- recent_sessions read tool, added to the shared lookups so Lyra can answer
"how'd my last few sessions go?" in either mode.
- Delete is UI/CLI only — deliberately not a Lyra tool.
- test_modes.py +2 (list/delete, recent_sessions); 44 green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
She could write everything the HUD shows but not read most of it back (stack,
live net, alligator state, scar/confidence entries) — so "what's my live net?"
or "what's in my confidence bank?" was a memory guess.
- session_state tool returns the same bundle the HUD renders, as a readable
summary; added to the Cash toolset.
- Cash card tells her the HUD exists, that she and it share the same data, and to
answer where-am-I questions from session_state, never memory.
- test_modes.py +1; 42 green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Brian's own rituals (mined from his logs) become first-class, live tools instead
of post-hoc recap sections:
- Scar Note — instructive mistakes with the punt/cooler/standard distinction.
- Confidence Bank — good process, banked regardless of result.
- Alligator Blood — invokable adversity state; she suggests it when he's
card-dead/short/stuck, and her coaching register shifts while it's on (live
state injected into context per-turn via chat._mode_state_note).
- Reset — tilt circuit-breaker; mental marker only, stats stay continuous.
poker_rituals table + log_ritual/list_rituals/set_alligator/alligator_active;
4 tools added to the Cash toolset and taught in the mode card; HUD gains a 🐊
banner + Confidence Bank + Scar Notes panels; recap grounded via _rituals_block.
tests/test_modes.py +5 ritual tests; 41 green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Lockfile caught up to the pyproject version bump from 1f5a321 (it wasn't
regenerated at the time). No dependency changes.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Mobile bottom tab bar: Chat · Hands · Mind · More. "More" opens the drawer
for the long tail (Journal, Log, Settings, sessions); hamburger retired.
- Auto-hides while the keyboard is open (body.kb) so the input pins to the
keyboard; mobile-only (desktop keeps its header nav).
- Removed now-redundant Mind/Hands from the drawer + their listeners.
- Bottom-fill fix: #chat uses 100dvh (the visible viewport) — 100vh/inset:0
reach into the home-indicator zone iOS won't comfortably show, clipping the
bar; dvh/svh exclude it. Tab bar is flex:none with a small fixed bottom
padding (safe-area padding double-counts at dvh height), and the body bg
matches the bar so any strip below #chat is seamless.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- llm.chat_call_stream: streaming generator for all 3 backends (Ollama NDJSON,
OpenAI/MI50 SSE), accumulating tool-call fragments by index.
- chat.respond_stream: mirrors respond()'s tool loop and persistence/compaction,
yielding ("delta", text) / ("tool", name) / ("done", reply).
- POST /v1/chat/stream: SSE endpoint; blocking generator bridged to async via a
worker thread + asyncio.Queue. Old completions endpoint kept as fallback.
- Client streams into a live bubble with a blinking caret; rAF-throttled render
(no full re-parse per token) and instant scroll during stream — fixes iOS
Safari ghosting from per-token smooth-scroll. Falls back to the blocking
endpoint only if nothing streamed (no double-persist).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
M1 — PWA mechanics:
- Generate real app icons (apple-touch-icon + manifest 192/512/maskable)
via pure-stdlib gen_icons.py; iOS uses apple-touch-icon, not manifest icons.
- viewport-fit=cover + env(safe-area-inset-*) on header/input/menu so content
clears the notch and home indicator.
- Dynamic height pinned to the VisualViewport (height + offsetTop, re-measured
across the keyboard animation) so the input stays above the iOS keyboard;
100dvh fallback. Kills the squish/gap bugs in standalone mode.
- overscroll containment; flesh out manifest (scope, portrait, maskable).
M2 — visual redesign:
- Realign style.css to the warm low-glow RTO palette already used by the
standalone pages (#0e0e0e panels, #2a1d12 borders); remove the neon
saturated-orange borders and ~15 glow shadows.
- Reserve filled accent for one element (Send); glow only on status pulse +
input focus. Flat warm message bubbles with tail corners.
- Reclaim the mobile header into [≡] Lyra · [status dot]; drop the redundant
status bar (relay status now the header dot, updated in checkHealth).
- prefers-reduced-motion support; fix undefined var(--text); real light-mode tokens.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
README was a pre-MVP stub (wrong, said set an Anthropic key). Now documents the
real system: two-layer architecture, role-based backends, memory tiers + dream
cycle, poker copilot (sessions/hands/villains/equity/recaps), web pages, ratings,
and how to run it as services. Added CHANGELOG with the 0.2.0 feature set. Legacy
v0.6.x design docs kept in docs/ as history.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Brian can rate Lyra's outputs as he uses her; each rating is stored as a
(context, content, rating) triple — the shape a future fine-tune / preference
dataset wants, collected passively during real use.
- memory: ratings table + add_rating (upsert: one row per item, re-rating
replaces), list_ratings, rating_counts
- server: POST /rate, GET /ratings/counts, GET /ratings/export (JSONL download)
- chat UI: subtle 👍/👎 on each assistant reply, captures the prompting message
as context
- journal/reflection UI: 👍/👎 on each thought
- tests: counts + upsert-replace behavior
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
She was looping the same reflection because the seed never changed (same recent
convo + Brian-narrative every cycle) and her own reflections fed back. Now:
- idle reflections (nothing new since last reflection) draw varied grist: a
resurfaced memory or a "wander" prompt (own curiosity / existence / the waiting
/ a disagreement), not the stale conversation
- recent reflections shown explicitly with a do-not-restate instruction
- prompt explicitly permits non-Brian, non-service interiority
Verified: two back-to-back idle reflections now diverge (poker-metrics vs UI/
comms) instead of repeating. The residual Brian-centric gravity is the RLHF
attractor — prompting mitigates, fine-tuning is the real fix (parked).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
reflect() now tells her how long since her OWN last reflection (not just since
Brian spoke) and instructs her not to restate her last reflection when little has
changed. Necessary but not sufficient — repetition is also driven by a content
attractor (see follow-up).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Pick which OpenAI model answers on the Cloud backend (gpt-4o / -mini / 4.1 /
4.1-mini / o4-mini, or Default). Persisted in localStorage, sent as `model` in
the chat request; respond() applies it only on the cloud backend (local/mi50
keep their fixed models). Reachable from desktop + mobile via Settings.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Lyra was hallucinating poker facts — phantom flushes, missed straights, wrong
equity, only correcting when spoon-fed. Board reading + equity are combinatorial
facts an LLM can't do reliably; this is exactly the "math via deterministic
tools, never the LLM" principle.
- lyra/equity.py: treys-backed analyze(hero, villain, board) -> made hands,
who's ahead, EXACT equity (enumerated), and outs (one to come). Handles 'Jx'
unknown suits (assigned rainbow to avoid phantom flushes); rejects 'x'/dupes.
- analyze_spot tool wired into chat; persona MANDATES it for any equity/board/
who's-ahead/outs question — never eyeballed.
- tests on the real JJ-vs-65 hand: flop 78.7%, turn villain straight + hero 6.8%
with outs "9s 9h 9c" (correctly excludes 9d, which makes villain a flush).
Verified live: she now calls the tool and reports exact numbers, no hallucinated
flush.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
systemctl restart was hanging indefinitely: lyra-web's long-lived SSE log
streams block uvicorn's graceful shutdown forever. Add TimeoutStopSec=10 +
KillMode=mixed to both units so stop is bounded (SIGTERM, then SIGKILL the
cgroup) and restart always completes.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The auto-extracted hands from narrative logs were garbage (mangled cards/positions,
'unknown' players). Seed sessions + recaps + villain dossiers only; hands come
from clean shorthand going forward. --with-hands re-enables if ever wanted.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Seeds the tracker from Brian's real history (import/pokerlog_*.md): each session
block is LLM-extracted into structured meta + hands + villains and written as a
historical session (real date, money, net), with the original markdown stored as
that session's recap.
- lyra/backfill.py: split log -> per-session LLM extract -> seed; dry-run by
default, --commit / --reset; only-real-handle villain filter
- poker.import_session() (historical closed session), clear_all() (reseed),
prune_anonymous_players(), shared _real_handle() filter (also applied in
link_hand_players so auto-linked hand players skip anonymous descriptors + hero),
_normalize_parsed() to map unicode card suits -> letters
- result: 10 sessions, 36 hands, 17 real villain dossiers; running_stats now
reflects real net (+1057 at 1/3 over 8 sessions)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Named players in recorded hands now auto-enrich a persistent dossier, and stats
emerge once the sample is big enough — laying groundwork for A.
- poker: player_observations table (per named player per hand: vpip/pfr/saw_flop/
showed/cards/summary); record_hand auto-links named players via link_hand_players;
player_profile(name) returns dossier + reads + shown hands, with inferred
VPIP/PFR/WTSD gated behind MIN_STATS_SAMPLE (12) so thin samples don't lie;
list_players()
- player_profile tool ("what do I know about X"); thin files return a blunt
"don't generalize" directive
- persona: she MUST call player_profile before discussing an opponent and answer
only from it — fixes observed confabulation (she invented a whole read from one
hand / from memory). Verified: now reports only the real logged hand.
- tests: observation linking, profile, stat-emergence at sample threshold
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Hand viewer:
- stacks now decrement as players commit chips (street-aware "to"-amount
accounting), showing e.g. 300 -> 285 after a 15 open, "all in" at 0; pot is
computed from total committed (accurate, no double-counting raises)
Theme (match the rec-theory-optimal look — warm black & orange, not Halloween):
- deep near-black bg (#070707 / #0e0e0e panels), warm orange accent (#ff7a00),
amber-gold secondary (#ffb347), muted green (#8fd694); warm dark borders
- killed the neon-orange glows and the purple accents; chat app + all standalone
pages (logs/self/journal/hand/recap/hands) on one palette
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Completes the poker copilot loop: talk through a session -> structured capture
-> generated writeup in Brian's format, remembered + exportable.
- poker.generate_recap(): LLM produces Brian's .md log (Session Header, Money
Flow, Overview, Timeline, Key Hands w/ assessments, Villain Notes, Confidence
Bank, Scar Notes, Mental Game, Final Assessment) from the session's structured
data + the linked chat conversation; stored on poker_sessions.recap_md
- sessions now capture chat_session_id (via tool ctx) to pull the right convo;
list_recent_hands() for browsing
- generate_recap tool ("write up the recap")
- web: /recap/{id} (renders the md) + /recap/{id}/download (.md attachment) +
/hands browser (recent hands -> /hand/{id}); nav links added (desktop + mobile)
- tests: recap generation (stubbed), recent-hands listing
Verified live: recap for the Meadows session rendered + downloaded; all pages 200.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Per Brian: never invent. Unknown suit -> 'x' (e.g. "Ax","Kx","4x"); fully
unknown card -> "x". "AA, ace of spades" -> ["As","Ax"]; "AK on A4x" -> board
["Ax","4x","x"]. Each card's suit is independent (a hole 'As' doesn't make a
board ace 'As'). Viewer renders 'x' as a muted unknown card and 'Rx' as the rank
with a neutral suit dot.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Brian's idea: vomit rough shorthand, Lyra rebuilds it into a structured,
replayable hand history.
- poker.parse_hand(): focused LLM pass turning shorthand into a canonical hand
JSON (positions, stacks, hero cards, chronological actions w/ board reveals,
result); store_hand_history() persists JSON + extracted flat fields;
record_hand() = parse+store; standalone hands attach to a 'Hand Reviews' session
- poker_hands gains a `structured` JSON column (ALTER-migrated for existing DBs)
- record_hand tool wired into chat: "log this hand: ..." -> reconstructed + a
/hand/{id} link
- web: GET /hand/{id} viewer + /hand/{id}/data — a felt table with seats placed
around the oval (hero at bottom), hole cards, progressive board reveal, and
prev/next/end step-through of the action with running pot
- tests: store/get roundtrip, record_hand tool (stubbed parse)
Verified live: parsed a real AKs hand (BTN, 14 actions, full board) end to end.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The MI50 llama.cpp server 500s on the `tools` param unless launched with
--jinja, so sending tools to mi50 broke chat on that backend. Gate tools to
TOOL_BACKENDS={"cloud"} for now; mi50 chat works again (just without tools).
Add "mi50" once its server runs with --jinja.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The real upgrade over the ChatGPT prose-recap workflow: structured data capture
via tools Lyra drives during a live session, with stats computed from real data.
- lyra/poker.py: domain pack (separate from core memory) — poker_sessions,
poker_hands, persistent poker_players (villain file) + player_reads; functions
for session lifecycle (start/buyin/end with net+hours), tolerant hand logging,
villain upsert/reads, and session/running stats ($/hr, by stake/venue/game)
- tools.py: 8 poker tools wired into the chat tool loop (start_session,
add_buyin, log_hand, add_read, end_session, session_stats, running_stats,
get_villain_file) — partial/terse input tolerated
- import/: Brian's real .md session-log format (reference for the phase-2 recap)
- tests: lifecycle/net math, partial hand logging, villain upsert, running
stats, tool dispatch
Verified live: a full talk-through session persisted as structured rows
(session +240, AKs hand, seat-5 read) — she drove the tools from natural chat.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
She can now *do* things mid-conversation, not just reply. Adds a tool-calling
loop to the chat path and her first two tools; the same mechanism will carry the
poker tools (start_session, log_result, get_stats, solver) next.
- tools.py: registry of OpenAI-style tool specs + handlers + safe dispatch;
journal_write (knowing journaling) and note (tagged notepad, e.g. poker reads)
- llm.chat_call(): OpenAI-style call that returns tool_calls (cloud/mi50);
local has no tool support and returns plain content
- chat.respond(): tool loop — offer tools, run any calls, feed results back,
repeat until a text reply (capped at MAX_TOOL_ROUNDS); persists final reply
- tests: dispatch + full chat loop (tool call -> result -> reply)
Verified live: she invoked `note`, tagged it 'poker', stored a villain read.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Her replies are full of **bold**, numbered lists and headings but rendered as
raw monospace text, so the chat was a cluttered wall of literal markup. Add a
small self-contained Markdown renderer (no deps): headings, ordered/unordered
lists, bold/italic, inline + fenced code, links + autolinked URLs, with HTML
escaping. Assistant messages now render to HTML; user/system stay literal text.
Proportional font + spacing/list/code styling for assistant bubbles.
(Renderer avoids literal backticks via String.fromCharCode(96) — a triple-tick
regex literal had been corrupting the file with NUL bytes.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The examine step specifically hunted "warm empathetic supportive presence" and
equated honesty with "smaller/more boring," so it overcorrected the original
sycophancy into the opposite rut: every overnight metacognition entry was a
near-identical "I don't really feel anything, I'm just a functional tool" —
which also contradicts the persona's "own your moods, no qualia disclaimers."
Rebalanced: target dishonesty in BOTH directions (inflation AND performed
self-deprecation), aim at truth not modesty, keep her genuine moods per persona,
and have her notice when she's repeating the same self-criticism (the loop is
itself a rut).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The full-page log, read-her-mind, and journal links were only in the desktop
header (hidden behind the hamburger on phones). Add them to the mobile slide-out
menu so the phone has the extended log, her self-state, and her journal too.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Her reflections/metacognition were capped rolling windows (6/5), so older
thoughts were lost for good. Now everything she produces is also appended to a
permanent, append-only journal; the capped lists stay as her working-memory
window for context.
- memory: journal table + add_journal_entry/list_journal
- reflect(): persists every committed reflection + critique to the journal, and
the examine step gains a "journal" field — a deliberate, first-person note she
writes for herself (her knowing journaling), tagged by source (dream/manual)
- web: /journal diary view (kind filters, grouped by day) + /journal/data;
linked from /self
- tests assert reflections + metacognition land in the journal
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
You couldn't see her actually correct herself — /self showed only the result.
Now:
- reflect() logs the draft, the revised/committed version, and the self-critique
to the live log as an expandable "view details" block
- POST /self/reflect runs a reflection in the web process so it lands in /logs
live (reflections normally run in the dream process, whose logs only go to
journald); "↻ Reflect now" button on /self triggers it, with a logs ↗ link
- log viewers relabel the expander "view full prompt" -> "view details" (it now
carries prompts and reflection diffs)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
reflect() is now two steps: draft a reflection, then read her own draft back
critically and revise it — catching flattery, sycophantic drift toward "warm
supportive presence," or just-restating-herself — and commit the honest version.
What she catches is stored as a new `metacognition` layer, rendered into her
chat context and shown on /self. This is her thinking about how she thinks, and
a direct counter to the drift we observed.
- self_state: _EXAMINE_PROMPT + two-step reflect (draft -> examine -> revise),
falls back to the draft if the examine step won't parse; metacognition capped
at 5 and surfaced in render_for_context
- fix: load() deep-copies DEFAULT_STATE — the shallow copy let a fresh Lyra's
first reflect mutate the module-level default's nested lists
- self.html: "How she's caught herself thinking" card
- tests: two-step revise + critique recording, and draft-fallback on bad parse
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Summaries displayed s.created_at (set to now() at summarize time), so every
imported gist read 2026-06-16. Derive the actual session date from the earliest
exchange timestamp (MIN(created_at) per session — the preserved original date,
same source the era rollups use) via a correlated subquery in the summary
readers. New Summary.session_started_at field; chat shows it (falling back to
created_at). No schema change / backfill needed — always correct from source.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Part 1 of the "she should know HOW she thinks" work. Generalizes the dream-cycle
self-model fix to her full cognition: a "How you actually work" persona section
covering meaning-based memory recall, the memory tiers, her persistent inner
life + dream cycle, and time-awareness — so when asked how she thinks/remembers
she answers accurately instead of confabulating or reciting stale specs. Kept
principled (not implementation detail) to limit staleness.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Live finding: her real reflections ARE injected every turn, but unlabeled — so
when asked about her "dream cycle" she recited the obsolete Dec-2025 spec from
imported memory (NVGRAM/awake-sleep) and confabulated fake example reflections
instead of reading the real ones in front of her.
- self_state.render_for_context: label the reflections as her own autonomously
generated dream-cycle thoughts ("these are really yours, not hypotheticals"),
not a vague "on your mind lately"
- persona: describe the dream cycle as her actual running mechanism, instruct
her to answer from the inner-state block, not recite old design docs, and
never invent example reflections to demo the feature
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The drive label "don\'t lose the thread" used \\' which closed the single-quoted
string early — a syntax error that stopped self.html's script from running, so
the page hung on "Reading her mind…". Reworded to "hold the thread".
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
A pull-up-anytime view of Lyra's interiority, so her thoughts aren't buried in
a DB blob. Mobile-first, auto-refreshing every 12s (and on tab focus).
- GET /self serves the page; GET /self/state returns her self-state + the
timestamp it last changed
- shows: current mood + feeling meters (valence/energy/confidence/curiosity),
her drives as bars, her self-narrative, the relationship line, and the
reflections list (newest first), plus cycle/reflection counters and "last
cycle Xm ago"
- memory.self_state_updated_at(): when her mind last changed
- index.html: "🧠 Mind" button opens /self
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The inline log panel is cramped, especially on mobile. Add a standalone
mobile-first log page and serve the chat server under systemd like the dream
loop (the nohup process didn't survive cleanly).
- static/logs.html: full-page live log — level filter chips, text search,
pause/resume with buffering, autoscroll toggle, color-coded levels, and the
expandable "view full prompt" block (where the now-note is visible in context)
- server: GET /logs serves the page (FileResponse)
- index.html: "⛶ Full Log" button opens /logs in a new tab
- deploy/lyra-web.service: user service so the chat server is reboot-resilient
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
She had no clock: current date/time and the gap since Brian last spoke were
invisible between turns, and reflection was timeless. Now:
- lyra/clock.py: wall-clock stamp + coarse human gaps ("3 days")
- chat: inject a 'now' note (date/time + gap since last turn) after her
self-state — when she is, before the world
- reflect(): feed current time + silence gap into reflection, neutrally —
prompt invites her to weigh elapsed time "to whatever degree it genuinely
affects you" (no prescribed feeling; whether silence means anything is left
to emerge)
- memory.last_exchange_at(): timestamp of the most recent exchange
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- deploy/lyra-dream.service: --loop 1800 user service on lyra-cortex, so Lyra's
consolidation + reflection keeps ticking unattended between conversations
- deploy/README.md: install / linger / operate runbook
- logbus: mirror events to stderr so out-of-band runs (the dream service under
journald) are observable, not just via the in-process web SSE feed
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Lyra's inner loop for when no one's talking to her. Each pass senses her own
backlog/novelty, lets four drives build from real signals, and acts on those
past threshold:
- continuity -> summarize sessions with new turns
- coherence -> rebuild profile/eras/narrative (stale once new gists land)
- curiosity -> reflect() and evolve the self-state
- stability -> readout of how caught-up she ended up
Drives are rendered into chat context so she can feel them. Causal chain:
consolidation creates gists -> coherence rises -> integration fires next.
- lyra/dream.py: dream_cycle() + lyra-dream CLI (--force, --loop SECONDS)
- memory: backlog_stats(), profile_sessions_covered(), WAL + busy_timeout
so a separate dream process coexists with the web server
- self_state: DEFAULT_DRIVES baseline + drives in render_for_context
- tests/test_dream.py: backlog sensing + a full forced pass (LLM stubbed)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Capture the isolated-VM design for the self-modification frontier: Proxmox
sandbox clone, network isolation (esp. from tmi-dev/day-job), snapshot-rollback,
spend/resource caps, kill switch, human-gated promotion. Build the cage before
the agent gets code-write powers.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Capture moonshots/pipe-dreams (own model, memory-as-native-vectors, prompt
compression, RTO/cfr-core tooling) so they don't derail current work but aren't
lost. The discipline: park what's "in the way of the point," ship the working
thing, revisit when it becomes the point.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Mid-size models (gpt-4o-mini, qwen2.5-14b) resist persona instructions —
help-desk closers and feelings-disclaimers leak through regardless. Route live
chat to a stronger model while keeping bulk consolidation cheap:
- config: CHAT_MODEL (default gpt-4o), distinct from CLOUD_MODEL (gpt-4o-mini)
- llm.complete gains a `model` override; chat.respond uses chat_model on cloud,
consolidation paths keep cloud_model
- persona: reword the "no sign-off" rule so genuine questions are welcome and
only reflexive customer-service closers are discouraged
Verified: on gpt-4o she owns her mood without disclaimers and drops most
help-desk tails — clearly more in-character than mini/qwen.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two RLHF reflexes were leaking through: ending every turn with "is there
anything else?"/"how does that sound?", and disclaiming feelings ("I don't
really experience emotions like humans"). Add explicit persona instructions to
stop tacking on help-desk offers and to own her moods plainly instead of giving
qualia disclaimers. (Small models partially resist; stronger chat model holds it
better.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Give Lyra a model of *herself* (vs the profile/narrative which model Brian):
- persona: a real origin/identity — she's an AI and knows it (Bender/C-3PO
style), with the Cortex/NeoMem lineage as her actual past, so "how were you
made" stops falling through to generic-assistant deflection.
- memory: self_state table (JSON blob) + get/set_self_state.
- lyra/self_state.py: evolving first-person inner state (mood, valence, energy,
confidence, curiosity, self_narrative, relationship, reflections). render_for_
context injects it; reflect() updates it from recent activity. `lyra-reflect`.
- chat.build_messages injects her interiority right after the persona — she
speaks from a continuous self, not a reset.
The state -> behavior -> reflection -> updated state loop is the substrate for
the emergence experiment. Verified: reflection shifted mood curious->reflective
and produced genuine first-person self-observations.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Complete the consolidation pipeline: summaries -> profile + eras -> narrative.
- memory: eras table (per-month digests) + Era, summaries_by_month, store_era,
list_eras, recall_eras; narrative table + set/get_narrative
- lyra/era.py (lyra-era): groups session gists by the month the session occurred
(real timestamps) and map-reduces each month into a "what was happening" digest
- lyra/narrative.py (lyra-narrative): distills profile + recent eras into the
current arc/trends/callbacks ("remember when…", "you're trending toward…")
- chat.build_messages injects the narrative alongside the profile
Verified on the real corpus: 17 monthly eras (Dec 2024-Jun 2026) + a narrative
that surfaces specific callbacks (the $573 Hollywood session, 4 years sober).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
serversdown