Commit Graph
4 Commits
Author SHA1 Message Date
zonggongandmultica-agent e291058754 HEL-882 rework: execution-table caps become real hard gates at the wire, SDK hidden retries and mid-add failures recover without loss, the 8360 release package runs offline, and the HEL-874 embedding evidence is corrected - every real provider exit (create/dialogue/life/gist/judgment/growth/proactive phrasing + mem0's internal extraction LLM) now persistently pre-allocates request count (per-category daily + one-shot totals in request_usage/budget_total, BEGIN IMMEDIATE) and a conservative token estimate before sending, refuses at the tightest cap (BudgetStop, no more downgrade-and-keep-calling) and survives restarts; known usage settles to actual, unknown/disconnect/failure keeps the full reservation; ledger guards gain per-day tops (embed_daily/llm_daily) and verify_ledger_authority rejects a fresh file bypassing the frozen 498/500 prefix; the mem0 SDK client's hidden OpenAI retries are disabled (max_retries=0) so every actual send is gated and answered failures (even 5xx) settle as used with only httpx.ConnectError-grade unsent requests released; embed refusals the SDK swallows into empty adds are re-raised as budget errors so the batch re-queues without burning an attempt and resumes when the cap returns, add_group resumes at vector-store level (src_group anchors: only missing rows resent, crash/restart/concurrency never duplicate), sem coverage is per-backend (stub batches no longer block a real rebuild, real done batches never resend) with backend-scoped batch keys; the release package is offline-executable: fresh-dir git clone, cybergril service account, PYTHONPATH in the unit so the lock-only venv imports the product, shipped caps explicitly 0 (=refuse by design until approved), deploy/mem0-stack.lock + accurate 3.11 wheel gap with a no-download path (hel857-venv is 3.10; never present stub/FTS as the real memory experience), and tools/check_release_offline_hel882.py verifies checkout/import/health/fail-closed/approved-run/data-preservation on loopback with echo stand-ins; evidence corrected per manager review: docs/rework/hel874/erratum-hel882.json + README appendix record that e978e87's verification did real local embeddings from cache (provable lower bound 12, exact total unrecoverable, temp sandbox gone) while zero-real-LLM/zero-download/zero-outbound stand, old JSONs and authoritative ledgers untouched, and tools/verify_mem0_real_hel874.py now installs the deterministic 512-dim stand-in before TextEmbedding construction plus non-loopback socket rejection (zero model loads) with final_embed_rows persisted before temp cleanup; verification: 272 python tests pass on 3.11.16 venv + requirements.lock offline (16 new: manager api repro blocked with 0 provider calls, sdk repro 1-HTTP/max_retries=0/llm_ops=1, mid-add swallow requeued-then-resumed, resume-no-duplicate, stub-to-mem0 rebuild, daily embed cap, unified mem0-LLM gate, authority refusal, restart-persistent counters, tightest-constraint settle/release), ruff clean (src/tests/tools), web tsc + 42 vitest + build via npm ci --offline, pinned-install deterministic re-run 12/12, manager review scripts re-run evidence attached in delivery, dual-target requirement checks 4/4 with handoff-HEL-882.json; zero real calls, zero real embeddings, zero non-loopback outbound, authoritative ledgers and freezes untouched, not deployed, PR not merged, no visual changes
Co-authored-by: multica-agent <github@multica.ai>
2026-10-26 23:20:18 +08:00
白栖知andmultica-agent f6614d4bf9 HEL-780: connect persistent companion and real streaming text chat
Co-authored-by: multica-agent <github@multica.ai>
2026-10-22 16:25:35 +08:00
mobai 8dea68a25b feat(think): four-tier model routing with budget guard
Corrects a common misconception: the 60s heartbeat must not call any LLM.
It runs the mood decay equation -- microseconds of float math. Giving it a
model is an architecture error, not a resource problem.

Real token spend lands in two places:
  - per-turn input context (ontology + state + dyad + memory)
  - background task frequency

So route by scene, not by time:

  none      rule/math     tick, reach-out decision, fillers, avatar drive
  low       local small   life events, memory gist, fact extract
  standard  mid + short   speculative-response second segment
  high      flagship      main dialogue, proactive message, crisis

Hard constraints:
  - dialogue floor is LOW, never NONE (can get dumber, cannot lose lines)
  - crisis never degrades (product constraint, not perf tuning)
  - degradation is predictable: one rung per request, floor only at
    the hard budget threshold

Cost levers, by actual payoff:
  1. prompt caching (fixed prefix) -- by far the largest
  2. upstream compression (small model condenses 20 memories to 3)
  3. batching (one batch per 30min, not 20 single calls)
  4. swapping to a smaller model -- smallest payoff, hurts persona

Also: cost degradation is narratable. 'Net error -> I zoned out' already
exists in the projection table, so being throttled is in-character and true.

- config.py: tiers / routing / budget, replacing llm+llm_fast+llm_background
- think/: scenes, budget, router, wiring
- models.yaml restructured around tiers
- SPEC gains 9.5
- 20 new tests; 78 total pass; ruff clean
2026-10-19 20:14:41 +08:00
mobai 1012d8bb79 chore: project foundation
Contracts (the shared language):
- percept / state / context / utterance / render / audio / capabilities
- Capability model instead of vendor schemas (see below)

Providers (capability-based abstraction):
- LLM + Embedding: OpenAI-compatible shape IS the standard, adopted directly
- ASR: internal contract is an event stream (partial -> final), because
  /v1/audio/transcriptions has no partial and dialogue needs it for
  endpointing. OpenAI adapter degrades to supports_partial=False.
- TTS: internal contract carries tags/emotion/style, because
  /v1/audio/speech has no paralinguistics and [laughter]/[breath] are
  the core of feeling alive. Unsupported providers are stripped AND
  reported via Degradation -- silent degradation is forbidden.
- Avatar: no de-facto standard, custom RenderCmd protocol driven by State.

Core:
- mood: dynamical system, slow variables, negativity bias (grudge),
  numbing after prolonged negativity
- bazi: four pillars -> numeric vector; 'ming' is data, not prose
- projection: phi, system event -> in-world statement; uncovered events
  are counted as projection gaps

Storage:
- SQLite schema, three memory layers, dyad ledger
- accessibility floor 0.03 (not 0): forgettable but still there

Config split:
- models.yaml is the ONLY place model versions live
- character.yaml holds world rules / emotion params, no version numbers

58 tests pass; ruff clean.
2026-10-19 20:08:52 +08:00