施工员2号 aba2aebeb0 HEL-877: wire the audited .12 voice base into the one CyberGril chat chain as backend-only code (no deploy, no real voice call) - new independent src/cybergril/voice module: session+router add POST /api/voice/session / DELETE end / GET honest status on the unified instance (server-side room + dedicated cybergril-voice-agent dispatch + short-TTL join token via zero-new-dependency livekit_wire HS256/Twirp matching the installed SDK's wire format verified claim-for-claim on the .12 agent venv stdin-only; keys never reach the browser; unconfigured/no-profile/TTS-down refuse honestly and capabilities.voice stays false), chat_api_client speaks the exact real.ts RealChatAdapter protocol (per-turn health epoch, bounded 409 resync, robust incremental SSE parsing across multibyte/CRLF/multi-data chunk splits, fixed turn_id idempotency, immediate HTTP-stream close on barge-in without waiting for the next ASR final), llm_chat_api is a livekit llm.LLM adapter whose only backend is POST /api/chat so persona/mood/dyad/memory/life-sharing/budget gates and the one same-turn chat_lock govern text and voice identically (phi spoken verbatim via ChatChunk, never archived, no receipt), receipt.py confirms a sharing receipt only after real SDK playback completion with every negative (interrupted/failed/mid-playout/phi/incomplete/stale-epoch) fail-closed and exactly once per turn, stt_funasr/tts_cosyvoice reuse the proven .12 adapters verbatim with provenance, worker.py assembles the dedicated worker (independent voice_state_cybergril.json so D1 stays undecided, no second persona, cg.subtitle data-channel contract for the visual lead, CG-011 proactive seam left as an explicit no-op hook, built-in health port 8082); app.py keeps a minimal registration with voice_service injection for isolated tests; deployment drafts only (config/voice.example.yaml, worker.env.example, restart_cybergril.sh.example with exact-pattern process matching, serve_dual_8360_8361.py keeping HTTP 8360 and the HTTPS voice entry on ONE process sharing state/locks, runbook-hel877-voice.md contract+startup+rollback incl. the >=6GB GPU precondition and the do-not-copy historical fuser warning) - verification: 281 pytest pass on 3.11 venv + requirements.lock after merging the chief engineer's mid-flight HEL-874 rework e978e87 in full (parallel work preserved), verify_hel877_voice_chain.py 8/8 offline full-chain incl. socket-level barge-in fail-close and text/voice mutual same-gate cancellation, real-SDK check on the .12 livekit-agents 1.8.8 venv 8/8 ALL PASS via SSH stdin with zero writes zero network, dual-port single-process smoke on loopback 18360/18361 with a throwaway self-signed cert then stopped, ruff clean on all new code, dual-target requirement checks pass vs origin/hel-870-assembly (current tip e978e87) and origin/main with handoff-HEL-877.json; zero real LLM/embed/TTS/ASR requests, no mic/camera, no downloads/installs, no service start/stop, no 4080 access, no third-account or profile writes, web/ untouched, not merged, not deployed
Co-authored-by: multica-agent <github@multica.ai>

## HEL-877 交付摘要(R5 后台施工,未部署未合并)

- 新独立模块 src/cybergril/voice/:专属 livekit worker(agent_name=cybergril-voice-agent,复用 .12 FunASR/CosyVoice/silero 底座,不 fork 不另搭平台,无第二人格/记忆,不直连 LLM);llm_chat_api 适配器与文字前端 real.ts 同协议走 POST /api/chat(epoch 409 有界重同步、SSE 分片、turn_id 幂等、打断即断流);/api/voice/session 建房+专属派发+短时限 token 全服务端(密钥不进浏览器;缺配置/无档案/TTS 停机诚实拒绝;capabilities.voice 不伪造);receipt 仅在 SDK 真实播放完成后确认(打断/失败/φ 轮/旧 epoch 全 fail-closed)。
- 部署草案:/opt/voice-agent-cybergril 目录、worker 健康端口 8082、HTTP 8360 + HTTPS 8361 同进程双入口(deploy/voice/serve_dual_8360_8361.py,共用同一状态/锁)、启动顺序含 TTS GPU≥6GB 前置与回滚;本卡未执行任何启停。契约(开关/字幕/失败/打断/挂断 + CG-011 主动开口接缝)见 docs/runbook-hel877-voice.md §3。
- 自测(零真实请求):281 pytest 全过(含语音 25 例:socket 级打断断流、SSE 边界、重放零调用、文字/语音同门互打断、回执全负例、token 房间隔离);scripts/verify_hel877_voice_chain.py 8/8;.12 agent venv(livekit-agents 1.8.8)真 SDK 检查 8/8(stdin-only 零写入零网络);双端口同进程本地冒烟通过;ruff 干净;需求检查双目标通过(含 handoff-HEL-877.json)。
- 施工中总工 HEL-874(e978e87)推上同分支,本卡已完整并入,不覆盖并行成果。未验证如实列明:端到端语音(TTS 停机/GPU 3.9GB<5.6GB)、真 LiveKit server 的 dispatch/delete、移动端实机——待接线/审核卡真机验收,未经真实验收不称 R5 完成。
2026-10-26 22:42:19 +08:00
HEL-877: wire the audited .12 voice base into the one CyberGril chat chain as backend-only code (no deploy, no real voice call) - new independent src/cybergril/voice module: session+router add POST /api/voice/session / DELETE end / GET honest status on the unified instance (server-side room + dedicated cybergril-voice-agent dispatch + short-TTL join token via zero-new-dependency livekit_wire HS256/Twirp matching the installed SDK's wire format verified claim-for-claim on the .12 agent venv stdin-only; keys never reach the browser; unconfigured/no-profile/TTS-down refuse honestly and capabilities.voice stays false), chat_api_client speaks the exact real.ts RealChatAdapter protocol (per-turn health epoch, bounded 409 resync, robust incremental SSE parsing across multibyte/CRLF/multi-data chunk splits, fixed turn_id idempotency, immediate HTTP-stream close on barge-in without waiting for the next ASR final), llm_chat_api is a livekit llm.LLM adapter whose only backend is POST /api/chat so persona/mood/dyad/memory/life-sharing/budget gates and the one same-turn chat_lock govern text and voice identically (phi spoken verbatim via ChatChunk, never archived, no receipt), receipt.py confirms a sharing receipt only after real SDK playback completion with every negative (interrupted/failed/mid-playout/phi/incomplete/stale-epoch) fail-closed and exactly once per turn, stt_funasr/tts_cosyvoice reuse the proven .12 adapters verbatim with provenance, worker.py assembles the dedicated worker (independent voice_state_cybergril.json so D1 stays undecided, no second persona, cg.subtitle data-channel contract for the visual lead, CG-011 proactive seam left as an explicit no-op hook, built-in health port 8082); app.py keeps a minimal registration with voice_service injection for isolated tests; deployment drafts only (config/voice.example.yaml, worker.env.example, restart_cybergril.sh.example with exact-pattern process matching, serve_dual_8360_8361.py keeping HTTP 8360 and the HTTPS voice entry on ONE process sharing state/locks, runbook-hel877-voice.md contract+startup+rollback incl. the >=6GB GPU precondition and the do-not-copy historical fuser warning) - verification: 281 pytest pass on 3.11 venv + requirements.lock after merging the chief engineer's mid-flight HEL-874 rework e978e87 in full (parallel work preserved), verify_hel877_voice_chain.py 8/8 offline full-chain incl. socket-level barge-in fail-close and text/voice mutual same-gate cancellation, real-SDK check on the .12 livekit-agents 1.8.8 venv 8/8 ALL PASS via SSH stdin with zero writes zero network, dual-port single-process smoke on loopback 18360/18361 with a throwaway self-signed cert then stopped, ruff clean on all new code, dual-target requirement checks pass vs origin/hel-870-assembly (current tip e978e87) and origin/main with handoff-HEL-877.json; zero real LLM/embed/TTS/ASR requests, no mic/camera, no downloads/installs, no service start/stop, no 4080 access, no third-account or profile writes, web/ untouched, not merged, not deployed
2026-10-26 22:42:19 +08:00
HEL-877: wire the audited .12 voice base into the one CyberGril chat chain as backend-only code (no deploy, no real voice call) - new independent src/cybergril/voice module: session+router add POST /api/voice/session / DELETE end / GET honest status on the unified instance (server-side room + dedicated cybergril-voice-agent dispatch + short-TTL join token via zero-new-dependency livekit_wire HS256/Twirp matching the installed SDK's wire format verified claim-for-claim on the .12 agent venv stdin-only; keys never reach the browser; unconfigured/no-profile/TTS-down refuse honestly and capabilities.voice stays false), chat_api_client speaks the exact real.ts RealChatAdapter protocol (per-turn health epoch, bounded 409 resync, robust incremental SSE parsing across multibyte/CRLF/multi-data chunk splits, fixed turn_id idempotency, immediate HTTP-stream close on barge-in without waiting for the next ASR final), llm_chat_api is a livekit llm.LLM adapter whose only backend is POST /api/chat so persona/mood/dyad/memory/life-sharing/budget gates and the one same-turn chat_lock govern text and voice identically (phi spoken verbatim via ChatChunk, never archived, no receipt), receipt.py confirms a sharing receipt only after real SDK playback completion with every negative (interrupted/failed/mid-playout/phi/incomplete/stale-epoch) fail-closed and exactly once per turn, stt_funasr/tts_cosyvoice reuse the proven .12 adapters verbatim with provenance, worker.py assembles the dedicated worker (independent voice_state_cybergril.json so D1 stays undecided, no second persona, cg.subtitle data-channel contract for the visual lead, CG-011 proactive seam left as an explicit no-op hook, built-in health port 8082); app.py keeps a minimal registration with voice_service injection for isolated tests; deployment drafts only (config/voice.example.yaml, worker.env.example, restart_cybergril.sh.example with exact-pattern process matching, serve_dual_8360_8361.py keeping HTTP 8360 and the HTTPS voice entry on ONE process sharing state/locks, runbook-hel877-voice.md contract+startup+rollback incl. the >=6GB GPU precondition and the do-not-copy historical fuser warning) - verification: 281 pytest pass on 3.11 venv + requirements.lock after merging the chief engineer's mid-flight HEL-874 rework e978e87 in full (parallel work preserved), verify_hel877_voice_chain.py 8/8 offline full-chain incl. socket-level barge-in fail-close and text/voice mutual same-gate cancellation, real-SDK check on the .12 livekit-agents 1.8.8 venv 8/8 ALL PASS via SSH stdin with zero writes zero network, dual-port single-process smoke on loopback 18360/18361 with a throwaway self-signed cert then stopped, ruff clean on all new code, dual-target requirement checks pass vs origin/hel-870-assembly (current tip e978e87) and origin/main with handoff-HEL-877.json; zero real LLM/embed/TTS/ASR requests, no mic/camera, no downloads/installs, no service start/stop, no 4080 access, no third-account or profile writes, web/ untouched, not merged, not deployed
2026-10-26 22:42:19 +08:00
HEL-877: wire the audited .12 voice base into the one CyberGril chat chain as backend-only code (no deploy, no real voice call) - new independent src/cybergril/voice module: session+router add POST /api/voice/session / DELETE end / GET honest status on the unified instance (server-side room + dedicated cybergril-voice-agent dispatch + short-TTL join token via zero-new-dependency livekit_wire HS256/Twirp matching the installed SDK's wire format verified claim-for-claim on the .12 agent venv stdin-only; keys never reach the browser; unconfigured/no-profile/TTS-down refuse honestly and capabilities.voice stays false), chat_api_client speaks the exact real.ts RealChatAdapter protocol (per-turn health epoch, bounded 409 resync, robust incremental SSE parsing across multibyte/CRLF/multi-data chunk splits, fixed turn_id idempotency, immediate HTTP-stream close on barge-in without waiting for the next ASR final), llm_chat_api is a livekit llm.LLM adapter whose only backend is POST /api/chat so persona/mood/dyad/memory/life-sharing/budget gates and the one same-turn chat_lock govern text and voice identically (phi spoken verbatim via ChatChunk, never archived, no receipt), receipt.py confirms a sharing receipt only after real SDK playback completion with every negative (interrupted/failed/mid-playout/phi/incomplete/stale-epoch) fail-closed and exactly once per turn, stt_funasr/tts_cosyvoice reuse the proven .12 adapters verbatim with provenance, worker.py assembles the dedicated worker (independent voice_state_cybergril.json so D1 stays undecided, no second persona, cg.subtitle data-channel contract for the visual lead, CG-011 proactive seam left as an explicit no-op hook, built-in health port 8082); app.py keeps a minimal registration with voice_service injection for isolated tests; deployment drafts only (config/voice.example.yaml, worker.env.example, restart_cybergril.sh.example with exact-pattern process matching, serve_dual_8360_8361.py keeping HTTP 8360 and the HTTPS voice entry on ONE process sharing state/locks, runbook-hel877-voice.md contract+startup+rollback incl. the >=6GB GPU precondition and the do-not-copy historical fuser warning) - verification: 281 pytest pass on 3.11 venv + requirements.lock after merging the chief engineer's mid-flight HEL-874 rework e978e87 in full (parallel work preserved), verify_hel877_voice_chain.py 8/8 offline full-chain incl. socket-level barge-in fail-close and text/voice mutual same-gate cancellation, real-SDK check on the .12 livekit-agents 1.8.8 venv 8/8 ALL PASS via SSH stdin with zero writes zero network, dual-port single-process smoke on loopback 18360/18361 with a throwaway self-signed cert then stopped, ruff clean on all new code, dual-target requirement checks pass vs origin/hel-870-assembly (current tip e978e87) and origin/main with handoff-HEL-877.json; zero real LLM/embed/TTS/ASR requests, no mic/camera, no downloads/installs, no service start/stop, no 4080 access, no third-account or profile writes, web/ untouched, not merged, not deployed
2026-10-26 22:42:19 +08:00
HEL-877: wire the audited .12 voice base into the one CyberGril chat chain as backend-only code (no deploy, no real voice call) - new independent src/cybergril/voice module: session+router add POST /api/voice/session / DELETE end / GET honest status on the unified instance (server-side room + dedicated cybergril-voice-agent dispatch + short-TTL join token via zero-new-dependency livekit_wire HS256/Twirp matching the installed SDK's wire format verified claim-for-claim on the .12 agent venv stdin-only; keys never reach the browser; unconfigured/no-profile/TTS-down refuse honestly and capabilities.voice stays false), chat_api_client speaks the exact real.ts RealChatAdapter protocol (per-turn health epoch, bounded 409 resync, robust incremental SSE parsing across multibyte/CRLF/multi-data chunk splits, fixed turn_id idempotency, immediate HTTP-stream close on barge-in without waiting for the next ASR final), llm_chat_api is a livekit llm.LLM adapter whose only backend is POST /api/chat so persona/mood/dyad/memory/life-sharing/budget gates and the one same-turn chat_lock govern text and voice identically (phi spoken verbatim via ChatChunk, never archived, no receipt), receipt.py confirms a sharing receipt only after real SDK playback completion with every negative (interrupted/failed/mid-playout/phi/incomplete/stale-epoch) fail-closed and exactly once per turn, stt_funasr/tts_cosyvoice reuse the proven .12 adapters verbatim with provenance, worker.py assembles the dedicated worker (independent voice_state_cybergril.json so D1 stays undecided, no second persona, cg.subtitle data-channel contract for the visual lead, CG-011 proactive seam left as an explicit no-op hook, built-in health port 8082); app.py keeps a minimal registration with voice_service injection for isolated tests; deployment drafts only (config/voice.example.yaml, worker.env.example, restart_cybergril.sh.example with exact-pattern process matching, serve_dual_8360_8361.py keeping HTTP 8360 and the HTTPS voice entry on ONE process sharing state/locks, runbook-hel877-voice.md contract+startup+rollback incl. the >=6GB GPU precondition and the do-not-copy historical fuser warning) - verification: 281 pytest pass on 3.11 venv + requirements.lock after merging the chief engineer's mid-flight HEL-874 rework e978e87 in full (parallel work preserved), verify_hel877_voice_chain.py 8/8 offline full-chain incl. socket-level barge-in fail-close and text/voice mutual same-gate cancellation, real-SDK check on the .12 livekit-agents 1.8.8 venv 8/8 ALL PASS via SSH stdin with zero writes zero network, dual-port single-process smoke on loopback 18360/18361 with a throwaway self-signed cert then stopped, ruff clean on all new code, dual-target requirement checks pass vs origin/hel-870-assembly (current tip e978e87) and origin/main with handoff-HEL-877.json; zero real LLM/embed/TTS/ASR requests, no mic/camera, no downloads/installs, no service start/stop, no 4080 access, no third-account or profile writes, web/ untouched, not merged, not deployed
2026-10-26 22:42:19 +08:00
HEL-869: read-only remediation evidence per manager independent review of 81d50f2 - recheck869.py reproduces the upstream rolling 10-message window from the fixed mem0 storage source (SHA256 5f5ecdad, byte-identical to the isolated venv install) in an independent in-memory SQLite with zero LLM zero embedding zero network (single 15-message batch truncates to 10; the real 9,1,1,2,2 add pattern keeps batch1 first-4 + later-6, dropping batch1 last-5) and verifies read-only that this round's mem0 messages hold 10/15 missing exactly t3 user/companion, t4 user/companion, t5 user with the restored public copy identical at 10, while the baseline main archive chat_message keeps all 15 verbatim per-role rows and the original calls inputs still carry all 15 - SDK messages is a rolling window, not a permanent quote archive, and no real user data was lost; erratum_HEL-869.json records six report-corrections without touching any original JSON (E1 S1 frozen criterion sampled only mem0 t1 so full-15 retention cannot be claimed, E2 this round used 46004 tokens = 44735+1269 while 163041 is the cumulative 39-call total, E3 raw QA6-13 map in order to plan QA1-8, E4 calls headline 'answered blue' is a stale run2 note vs this round's honest unknown answer, E5 in-process llm_calls_real=0/embed 39 are not authoritative vs persistent ledgers' 13 sent/40 dense, E6 S9 re-verified only the public instance with full private+public recovery reused from gate2 P5); README HEL-868 record annotated in place at the three corrected conclusions plus a HEL-869 section; plan-next-HEL-869.md delivers the decision-only minimal follow-up (main archive stays the permanent quote layer with reread via turn_ids anchors, mem0 interfaces reused, no upstream fork; batch-extraction line-level provenance proven by input-side pre-grouping where each add's anchor equals its actual input set with gate-side sha256 recompute - no gate bypass, no anchor trimming, no prompt-only promises; S6 verified via explicit-status questioning under the same frozen bounded grammar with no wording amnesty; Chinese custom_instructions deferred; upstream/ours responsibilities, risks incl. embed budget 498/500 remaining 2 as the hard blocker, and a staged minimal retest: stage0 offline zero-cost audit then stage1 ~14 requests/~40-46 embeddings pending quota decision); all three ledgers byte-identical before/after (03f92557/c9871277/9716c722), every e06dcaf result blob unchanged, no criteria/fixtures/orchestration edits, no real requests, no embeddings, no merge, no deploy; dual-target requirement checks 4/4 pass vs origin/main and origin/agent/agent/de8874a3803b with/without handoff-HEL-869.json
2026-10-26 04:09:58 +08:00
HEL-877: wire the audited .12 voice base into the one CyberGril chat chain as backend-only code (no deploy, no real voice call) - new independent src/cybergril/voice module: session+router add POST /api/voice/session / DELETE end / GET honest status on the unified instance (server-side room + dedicated cybergril-voice-agent dispatch + short-TTL join token via zero-new-dependency livekit_wire HS256/Twirp matching the installed SDK's wire format verified claim-for-claim on the .12 agent venv stdin-only; keys never reach the browser; unconfigured/no-profile/TTS-down refuse honestly and capabilities.voice stays false), chat_api_client speaks the exact real.ts RealChatAdapter protocol (per-turn health epoch, bounded 409 resync, robust incremental SSE parsing across multibyte/CRLF/multi-data chunk splits, fixed turn_id idempotency, immediate HTTP-stream close on barge-in without waiting for the next ASR final), llm_chat_api is a livekit llm.LLM adapter whose only backend is POST /api/chat so persona/mood/dyad/memory/life-sharing/budget gates and the one same-turn chat_lock govern text and voice identically (phi spoken verbatim via ChatChunk, never archived, no receipt), receipt.py confirms a sharing receipt only after real SDK playback completion with every negative (interrupted/failed/mid-playout/phi/incomplete/stale-epoch) fail-closed and exactly once per turn, stt_funasr/tts_cosyvoice reuse the proven .12 adapters verbatim with provenance, worker.py assembles the dedicated worker (independent voice_state_cybergril.json so D1 stays undecided, no second persona, cg.subtitle data-channel contract for the visual lead, CG-011 proactive seam left as an explicit no-op hook, built-in health port 8082); app.py keeps a minimal registration with voice_service injection for isolated tests; deployment drafts only (config/voice.example.yaml, worker.env.example, restart_cybergril.sh.example with exact-pattern process matching, serve_dual_8360_8361.py keeping HTTP 8360 and the HTTPS voice entry on ONE process sharing state/locks, runbook-hel877-voice.md contract+startup+rollback incl. the >=6GB GPU precondition and the do-not-copy historical fuser warning) - verification: 281 pytest pass on 3.11 venv + requirements.lock after merging the chief engineer's mid-flight HEL-874 rework e978e87 in full (parallel work preserved), verify_hel877_voice_chain.py 8/8 offline full-chain incl. socket-level barge-in fail-close and text/voice mutual same-gate cancellation, real-SDK check on the .12 livekit-agents 1.8.8 venv 8/8 ALL PASS via SSH stdin with zero writes zero network, dual-port single-process smoke on loopback 18360/18361 with a throwaway self-signed cert then stopped, ruff clean on all new code, dual-target requirement checks pass vs origin/hel-870-assembly (current tip e978e87) and origin/main with handoff-HEL-877.json; zero real LLM/embed/TTS/ASR requests, no mic/camera, no downloads/installs, no service start/stop, no 4080 access, no third-account or profile writes, web/ untouched, not merged, not deployed
2026-10-26 22:42:19 +08:00
HEL-877: wire the audited .12 voice base into the one CyberGril chat chain as backend-only code (no deploy, no real voice call) - new independent src/cybergril/voice module: session+router add POST /api/voice/session / DELETE end / GET honest status on the unified instance (server-side room + dedicated cybergril-voice-agent dispatch + short-TTL join token via zero-new-dependency livekit_wire HS256/Twirp matching the installed SDK's wire format verified claim-for-claim on the .12 agent venv stdin-only; keys never reach the browser; unconfigured/no-profile/TTS-down refuse honestly and capabilities.voice stays false), chat_api_client speaks the exact real.ts RealChatAdapter protocol (per-turn health epoch, bounded 409 resync, robust incremental SSE parsing across multibyte/CRLF/multi-data chunk splits, fixed turn_id idempotency, immediate HTTP-stream close on barge-in without waiting for the next ASR final), llm_chat_api is a livekit llm.LLM adapter whose only backend is POST /api/chat so persona/mood/dyad/memory/life-sharing/budget gates and the one same-turn chat_lock govern text and voice identically (phi spoken verbatim via ChatChunk, never archived, no receipt), receipt.py confirms a sharing receipt only after real SDK playback completion with every negative (interrupted/failed/mid-playout/phi/incomplete/stale-epoch) fail-closed and exactly once per turn, stt_funasr/tts_cosyvoice reuse the proven .12 adapters verbatim with provenance, worker.py assembles the dedicated worker (independent voice_state_cybergril.json so D1 stays undecided, no second persona, cg.subtitle data-channel contract for the visual lead, CG-011 proactive seam left as an explicit no-op hook, built-in health port 8082); app.py keeps a minimal registration with voice_service injection for isolated tests; deployment drafts only (config/voice.example.yaml, worker.env.example, restart_cybergril.sh.example with exact-pattern process matching, serve_dual_8360_8361.py keeping HTTP 8360 and the HTTPS voice entry on ONE process sharing state/locks, runbook-hel877-voice.md contract+startup+rollback incl. the >=6GB GPU precondition and the do-not-copy historical fuser warning) - verification: 281 pytest pass on 3.11 venv + requirements.lock after merging the chief engineer's mid-flight HEL-874 rework e978e87 in full (parallel work preserved), verify_hel877_voice_chain.py 8/8 offline full-chain incl. socket-level barge-in fail-close and text/voice mutual same-gate cancellation, real-SDK check on the .12 livekit-agents 1.8.8 venv 8/8 ALL PASS via SSH stdin with zero writes zero network, dual-port single-process smoke on loopback 18360/18361 with a throwaway self-signed cert then stopped, ruff clean on all new code, dual-target requirement checks pass vs origin/hel-870-assembly (current tip e978e87) and origin/main with handoff-HEL-877.json; zero real LLM/embed/TTS/ASR requests, no mic/camera, no downloads/installs, no service start/stop, no 4080 access, no third-account or profile writes, web/ untouched, not merged, not deployed
2026-10-26 22:42:19 +08:00
HEL-874 rework: bounded verbatim readback reaches the real provider input, real pinned Mem0 2.2.1 SDK contract verified live and the adapter's five wiring defects fixed, budget gates made provable (atomic reserve/settle embed ledger + independent llm_ops cap), definite-cap real-check execution table, 8360 isolated release package prepared (not deployed), 2026-10-26 user originals recorded source-only - semantic.context now quotes chain-verified archive rows (original words + role + when, <=200-char window with visible trim marker, summary demoted to a one-line pointer, same min(600,budget) cap) and episodic.context flips from gist-first-120-chars to verbatim-first (<=240) + role + date, both proven by tests/test_readback_hel874.py capturing the actual messages handed to the provider through the real API path: a detail absent from the gist yet quoted verbatim from a turn pushed outside the recent-20 window, the same for a semantic chain hit, and an unshared companion row never leaking through FTS, semantic or any anchor (AM-001); tools/verify_mem0_real_hel874.py runs the FIXED install itself (mem0ai 2.2.1 2c81af3, real embedded qdrant, real fastembed from the existing cache, deterministic 127.0.0.1 fake gateway, temp ledgers only): before=31970c4 dies at construction (second QdrantClient on the same path rejected by the SDK - recorded), and source-level findings confirmed live and fixed - unscoped search raises ValueError in 2.2.1 (add now carries user_id scope, search passes filters+top_k since limit is silently ignored), payload reads ride the SDK's own vector_store.client, the old outer search debit double-counted (wrapper is now the single accounting point), the SDK's internal extraction LLM was ungated (LlmLedgerGuard with its own llm_ops table, cap default 0) and check-then-debit could overshoot under concurrency (EmbedLedgerGuard now reserves atomically in BEGIN IMMEDIATE, releases on failure/cancel, purges stale holds); after fixes all 12 checks pass (construction/anchor payload/user scope/1 LLM + 2-3 embeds per add measured/extraction input set provably only prior gated batches incl. old memories and rolling history/unshared canary zero outbound/search single debit/concurrency never over cap/close-reopen intact/both zero caps refuse with ZERO outbound LLM requests), evidence in docs/rework/hel874/; mood sensitivity returns to the original repo's continuous rule 1.35-0.70*strength (bazi.sensitivity_from_strength, foundation commit 1012d8b) replacing HEL-870's invented 1.25/0.8 step and is labeled an engineering heuristic, never user-approved bazi science (external agent formulas remain not adopted); docs/budget-plan-hel874.md gives definite caps for the one-time team check (55 requests/0 embeds), first-week semantic activation (<=60 extract LLM / <=180 embeds) and user daily use with code-level hard-gate facts and authorization sources (old 20-30/100-200 were estimates, not approved quota; frozen 39/60 and 498/500 untouched and never挪用); deploy/hel874-release.md + cybergril-hel874-8360.service.example prepare the isolated fresh-empty-data-dir release with health capability matrix and data-preserving rollback (NOT executed, PR#11 not merged, no user-facing loop, no experience address); sources/comment-evidence-full-experience-20261026.json adds the two 2026-10-26 user originals (01a17806, 01a1780e full-experience-then-iterate) with authorship verified and manifest appended only - baseline/current/coverage/changes untouched; verification: 256 python tests pass on 3.11.16 venv + requirements.lock offline (3 new real-provider-input readback cases incl. AM-001 negative, mem0 fake rebuilt to mirror the real 2.2.1 surfaces, 9 cases), ruff clean, web tsc + 42 vitest + build pass, local 8360 stub smoke (loopback, isolated temp DB) health R4-assembly all green then stopped, real-SDK script re-run 12/12 after final edits, dual-target requirement checks 4/4 pass with handoff-HEL-874.json; zero real model calls, zero real embeddings, zero outbound (fake gateway on loopback + existing local model cache only), authoritative ledgers and freezes untouched, no existing services or profiles touched
2026-10-26 22:13:45 +08:00
HEL-870 assembly: memory/persona/mood/life/proactive-contact joined into one runnable candidate on an additive R4 migration (no existing table or row replaced; no visual changes beyond mechanical proactive flag passthrough) - src/cybergril/memory/semantic.py productizes the audited Mem0 path per plan-next-HEL-869: input-side pre-grouping sends only public rows outbound (user rows always; companion rows only judged+acked with all-told candidates), every batch anchor turn_ids/turn_roles/input_sha256 equals that exact input set, and the read-time gate independently recomputes the fingerprint against the live archive and re-checks every companion row's share state, fail-closed rejecting no_payload/unknown_source/archive_missing/source_mismatch/hit_text_mismatch/anchor_turn_mismatch/role_text_conflict/companion_unshared with honest row_verbatim/row_anchor/batch provenance; main archive chat_message remains the ONLY permanent quote layer; sem_batch/sem_covered carry a lease-recovered idempotent job loop; EmbedLedgerGuard continues the single authoritative embed ledger and with embed_cap=0 (no approved budget) refuses every real embedding so the frozen 498/500 and 39/60 accounts and the HEL-835 gate stay untouched; fixed Mem0 2c81af3 is lazily imported with zero upstream fork, a deterministic stub backend exists for engineering verification and is never claimed semantic; MoodRuntime makes core.mood live (event impulses scaled by bazi strength sensitivity, heartbeat decay, state_snapshot persistence, restart recovery, model sees one word only); DyadRuntime makes the relationship ledger live (completed exchanges are rituals, daily regression, behavior hints stay instructions); life_activity adds start/continue/interrupt/resume/finish continuity across slots and restarts with chat-completion interrupts; reach_out implements the approved R4 scope from character.yaml (>=6h interval, hours 8-23, salience>=0.6, arousal>=0.5, intimacy>=0.4) with traceable already-public bases only, UNIQUE(basis) so restarts and duplicate tasks can never multiply messages, night suppression, absence-rule fail-closed, phi projection, and messages persisted as companion rows (proactive=1) fully synced on reopen; API health reports R4-assembly with per-capability honesty (voice/vision/weather/materials/backup stay false, semantic_memory names its backend); config/semantic.example.yaml sanitized config and docs/runbook-hel870.md (startup, locked deps, health, independent data dir, disable/rollback, experience chain list, budget request, R1-R7 seams); verification: 250 python tests pass on 3.11 venv + requirements.lock (61 new deterministic/fake-provider cases covering every gate negative, restart recovery, dedupe, red line and the full entry chain incl. reopen epoch sync), ruff clean, web typecheck + 42 vitest + build pass with the original author frontend untouched except the one-line proactive passthrough; dual-target requirement checks 4/4 pass vs origin/main and origin/agent/agent/de8874a3803b with handoff-HEL-870.json; zero real model calls, zero embeddings, zero outbound network this card; local unified-instance smoke on port 8360 (stub semantic) serves the built web and reports the life loop waiting; read-only server check recorded in the runbook (existing 8300-8350 services and voice ports observed, end-to-end voice availability NOT verified); not merged, not deployed, no production or shared services touched
2026-10-26 21:03:42 +08:00
2026-10-19 20:08:52 +08:00

CyberGril

一个有自己世界的赛博伴侣。

不是聊天机器人。区别在于:它在你不说话的时候也在运行。


先看一眼

晚晚 · 介绍动画

▶ 打开完整动画 — 22 秒,纯代码生成,无任何外部资源

docs/intro.html          单文件动画,双击即播
docs/preview.webp        渲染好的预览(800×450 / 10fps / 22s / 0.5MB)
tools/render_preview.py  重新渲染、换分辨率、导出静帧

动画里有一条贯穿的主线:强调色 = 当前效价 v。 情绪低落时画面整体转冷,而不是靠字幕告诉你"她现在很难过"。

幕 时间 讲什么
膜 0–6s 物质撞在膜上弹回,信息穿过去
心跳 6–12s 没有输入时它也在跳;难过恢复得慢
记忆 12–17s 记得的会沉下去,但沉不到底
名字 17–22s —

四张静帧:docs/01-membrane.png 02-heartbeat.png 03-memory.png 04-title.png

动画本身的技术取舍:时间轴是纯函数 render(t),不做 RAF 累加。 所以任意一帧都能独立渲染——这也是 render_preview.py 能逐帧截图、 而不需要真实录屏的原因。


它是什么

一个单用户、本地优先、常驻的存在,通过一层"膜"与你通信:

  • 有自己的时钟 —— 无输入时仍在演化、仍有生活
  • 情绪是动力学系统 —— 慢变量、有惯性、有代价(会心情不好)
  • 记忆有形状 —— 三层结构,会遗忘,且遗忘有下限
  • 关系是一等实体 —— 信任、债务、未修复的事
  • 本体不可谈判 —— 立场不因任何对话、情绪、施压而改变

三层不可混淆的东西

层 载体 变化速度 谁能改
命 bazi_core.derived 几乎不变 只有整合器,每次 ±0.02,且留痕
运 fortune_period 按周/月 排盘
状态 State (v/a/d/e) 按秒/分 tick 循环

混掉任何两层,角色就会失真:

  • 命被对话改 → 人设漂移,变成一个讨好你的东西
  • 状态被写成描述 → 模型自己编情绪,动力学失效

架构

contracts/   数据契约 —— 全项目唯一的共同语言
providers/   模型接入 —— 按能力建模
core/        状态核心 —— 独立时钟,不依赖用户输入
memory/      三层记忆 + 遗忘
dyad/        关系账本
think/       模型路由 + 预算 —— 四档分配,按关键时刻
config/      models.yaml(版本号) + character.yaml(世界法则)

设计判断(这一节比代码重要)

1. OpenAI 兼容是 adapter,不是标准

对 LLM 和 Embedding,OpenAI 的形状(/v1/chat/completions、/v1/embeddings) 已经是事实标准,内部契约就直接用它。

但对 ASR / TTS / 形象,OpenAI 的形状会砍掉这个项目最需要的能力:

能力 OpenAI 形状的问题 我们的做法
ASR /v1/audio/transcriptions 是「文件进→文本出」,没有 partial;而对话必须要 partial 才能做端点检测 内部契约是事件流(partial → final);OpenAI 适配器作为降级档,supports_partial=False
TTS /v1/audio/speech 只有 text/voice/speed,没有副语言标签;而 [laughter] [breath] 是"像活的"的核心 内部契约带 tags / emotion;不支持的 provider 在 plan() 里显式降级并上报
形象 没有事实标准 自定义 RenderCmd 协议,驱动源必须是 State,不是 LLM 文本

关键机制:静默降级是不允许的。

planned = tts.plan(request)
if planned.lost_paralinguistics:
    # 日志里会出现 [degraded] speech_tags: ...
    ...

静默降级会让你在调优时永远找不到"它为什么不像人了"。

2. 状态必须以数值注入,不能用散文描述

✗ "你现在有点难过"          → 模型会自己发挥,甚至编出更多情绪
✓ v=-0.62 a=0.41 d=0.30    → 模型知道边界
  + visible="低落"
  + constraint="不得声称与此不符的内在状态"

StateBlock 里那个 constraint 字段不是装饰,它是唯一能阻止 模型绕过状态机的机制(外加输出校验与重写)。

3. 遗忘有下限

mem_episodic.accessibility 的下限是 0.03,不是 0。

  • 正常检索拿不到 → 想不起来
  • 情绪线索命中时能捞出来 → 但还在

所有现成框架都是"要么记住要么删除",没有这个中间态。

4. 未修复的事变成行为,不变成台词

✗ "你上次伤害了我"          → 指责,而且廉价
✓ 主动性 -0.1,回复变短      → 它不说,但它不一样了

5. 缺席可以被提及,不能被使用

✗ "我一直在等你,等了好久"   → 制造负罪感,这是操控
✓ "这几天我这边发生了点事"   → 它有自己的世界,不是空转等你

这条写在 core/projection.py 的注释里,也有对应的测试。

6. 大模型按「关键时刻」分配,不按时间分配

先纠正一个常见误解:60 秒心跳不该调用任何 LLM。

心跳做的是 s = b + (s-b)*exp(-dt/τ) + I,四个 float 一次 exp, 微秒级。给它配模型是架构错误,不是资源问题。

真正的 token 消耗在两处:每轮对话的输入上下文,以及后台任务的频次。

所以用四档:

档 用什么 场景 成本
none 不调用 心跳、主动性判定、填充词、表情驱动 ≈0
low 本地小模型 生活事件、记忆压缩、事实抽取、输出初筛 电费
standard 中大模型 + 短上下文 投机响应的第二段 低
high 旗舰 主对话台词、主动联系、危机 按 token

路由判定必须便宜 —— 查表,不调模型。为了省 token 先调一次模型,是本末倒置。

关键约束:

  • 主对话的下限是 low,不是 none。 预算耗尽时可以变笨,不能没台词。
  • 危机场景永不降级。 这是产品约束,不是性能优化。
  • 降级要可预测。 单次请求放不下只降一档;到硬阈值才降到下限。

省钱的真正杠杆,按收益排序:

  1. Prompt 缓存 —— 本体/世界法则固定放前缀,命中后便宜一个数量级。杠杆最大。
  2. 上游压缩 —— 小模型先把 20 条记忆压成 3 条,大模型只收结果。输入从 4000 降到 800 tokens。
  3. 批处理 —— 记忆整合攒 30 分钟一次批量处理,不是 20 次单条调用。
  4. 换小模型 —— 收益反而最小,而且伤害人设。

还有一个漂亮的副作用:成本降级可以被叙事化。 大模型不可用时它"有点走神"—— 这既符合人设,又是真的(实现同构的虚构)。


快速开始

# 1. 安装
pip install -e ".[dev]"

# 2. 配置
cp .env.example .env
# 编辑 .env 填入 API key(不填也能跑,会回退到 echo provider)

# 3. 跑测试
pytest -q

# 4. 检查配置能被正确加载与校验
python -c "from cybergril import load_settings; print(load_settings('config').character.name)"

默认配置不需要任何 API key 就能跑通测试 —— 因为内置了 echo(LLM)/ scripted(ASR)/ silence(TTS)三个离线 provider。 它们的价值是把接口形状固定下来,让上层可以先跑通再换真后端。


换成真模型

只改 config/models.yaml,不动任何代码。

tiers:
  high:
    provider: openai_compat
    model: deepseek-v4.1-flash
    base_url: https://api.deepseek.com/v1
    api_key_env: DEEPSEEK_API_KEY
  low:
    provider: openai_compat
    model: qwen3.5-4b-instruct
    base_url: http://127.0.0.1:8000/v1

tts:
  provider: cosyvoice        # 需要自己实现这个 adapter,见下
  model: fun-cosyvoice3-0.5b

新增一个 provider 只需要三步:

# 1. 继承对应的基类
class MyTTS(TTSProvider):
    name = "mytts"

    # 2. 诚实地声明能力
    @property
    def capabilities(self) -> Capabilities:
        return Capabilities(
            modality=Modality.AUDIO,
            streaming=Streaming.CHUNKED,
            speech_tags=frozenset({"laughter", "breath"}),
        )

    # 3. 实现 stream()
    async def stream(self, request): ...

# 注册
registry.register("tts", "mytts", MyTTS)

能力必须诚实。 夸大能力会让上层做出错误的规划, 而 plan() 会据此剥掉你其实需要的东西。


配置分层

文件 放什么 什么时候改
config/models.yaml 唯一允许写模型版本号的地方 换模型(每季度复核)
config/character.yaml 世界法则、情绪参数、关系参数 换人格(几乎不动)

两个文件互不污染。这是"防止文档和代码一起过时"的机制。

关于版本号的一个约定:只有 models.yaml 允许写版本号。 代码里所有模型引用走配置,不硬编码。


测试

pytest -q                    # 全部
pytest tests/test_mood.py -v # 情绪动力学
pytest -k degradation -v     # 降级机制

测试守的是设计判断,不是实现细节:

文件 守什么
test_contracts.py 接口形状
test_providers.py 降级必须显式上报
test_mood.py 情绪有惯性;负向恢复更慢
test_bazi.py 命是数值;强弱影响敏感度
test_projection.py 未覆盖事件必须计数;投影不泄漏技术词
test_store.py 遗忘有下限;间隔效应

下一步

地基只包含:契约、provider 抽象、配置、情绪动力学、八字编译、投影、存储 schema、关系账本。

还没有的(按依赖顺序):

  1. core/tick.py —— 四个循环(tick / consolidate / life / reach_out)
  2. memory/retrieval.py —— mood-congruent 检索
  3. memory/consolidation.py —— 整合与身份提升
  4. think/ —— 上下文组装、生成、三道校验
  5. express/ —— TTS 管道、形象、barge-in
  6. api/ —— WebSocket 会话

规格

完整设计见 docs/SPEC.md。

许可

私有项目,未授权。

S1 前端演示

独立的 React / TypeScript 前端位于 web/,运行 cd web && npm ci && npm run dev 后从普通入口进入明确标识的本机演示。真实后台、语音、视觉和备份尚未接通;不自动连接模型或现有服务。安装、构建和测试命令见 web/README.md,实现与后续接线见 docs/frontend/IMPLEMENTATION.md。

S
Description
No description provided
Readme
202 MiB
Languages
Python 57.2%
TypeScript 22.7%
CSS 19.9%
JavaScript 0.1%
HTML 0.1%