Commit Graph
21 Commits
Author SHA1 Message Date
Dudu-0223 ec7030afd3 fix(compaction): default summary output cap to headroom 2026-09-19 19:56:54 +08:00
Dudu-0223 9b016b6425 refactor(compaction): require an explicit output reservation 2026-09-19 19:56:53 +08:00
Dudu-0223 555b664b08 fix(compaction): reserve configurable 64K pressure headroom 2026-09-19 19:56:53 +08:00
Dudu-0223 b02b20d713 fix(compaction): pin output caps and cover proactive reserve replay 2026-09-19 19:56:51 +08:00
Dudu-0223 0fadb08fbd fix(compaction): gate pressure on the message budget, not the whole window
Proactive pressure scaled `thresholdRatio` by the routed model's entire
context window, but a request also reserves output tokens the provider
charges to that same window — the LLM service deliberately never clamps
them (`packages/llm/llm/src/index.ts`).

With a 1,000,000-token window and the default 256,000-token request cap the
gate therefore sat at 800,000 while the provider refused any prompt above
792,576. Observed in a real session: three requests died at 792,850 /
797,639 / 793,352 message tokens ("maximum context length is 1048576 ...
you requested 1048850 tokens"), no pressure compaction ever fired, and the
overflow recovery that did run could not condense an ~800k-token history
into the 8192-token summary cap.

Scale both ratios by the message budget left after the reserve, sourced from
the effective request envelope's `maxTokens` and falling back to the
adapter's per-request output default. A reserve that leaves no message
budget reports a target-specific pressure configuration error, matching the
existing per-route warning policy rather than silently disabling proactive
compaction. The headless compaction e2e advertises a coherent capacity/cap
pair so its 8000-token message budget and 4000-token threshold are unchanged.

Related #1841
2026-09-19 19:56:51 +08:00
_KermanandTianyi Cui fb79a944f5 refactor(llm): separate durable producer sources from request inputs
Give durable user context its producer-owned source kind instead of a generic plugin wrapper. The V3-to-V4 conversion preserves the frozen rename and collision rules, all own JSON metadata, event coordinates, and historical request-only kinds. Native catalog and JSONL admission retain source validation so unknown attribution can survive without its producer while malformed owned sources are refused.

Use an explicit user-only RequestUserInput for auxiliary LLM prompts that do not own durable message identity. Auto Review and compaction remove their request-only source registrations while retaining prompt content and freezing; provider serializers and projections keep the same wire data, assistant replay isolation, image ordering, and tool correlation. Session writes, Agent delivery, and recorded title requests continue to require identified durable messages.

Bind the saved compatibility policy only to user sources and qualify tmux location attribution without weakening fixed system, model, or tool semantics. Rebuild the current inventory and unpublished V4 acknowledgement on the actual tool-role parent, preserving its rationale and all accepted historical records. The catalog format version records tooling metadata; Session remains V4. Developer-role policy and activation belong to their own feature.

The latest-master Messages adapter, serialization, request-input typing, EpochHeader reservation, and attribution checks pass 103 tests across five files. The retained V3 Messages input-history snapshot passes through the source profile; Host/Client typechecks and all 41 documentation gates pass. The generated comparison records four changed roots and 447 unchanged type fingerprints. Latest-master ARIA changes are retained, with producer source labels applied independently; new Messages fixtures retain their producer names as declared source kinds and use the fixed system-message helper; recorded Session generations are unchanged by this regrouping.

Co-authored-by: Tianyi Cui <[email protected]>
2026-09-18 19:07:12 +08:00
creatixchu debb4b9a9c fix(compaction): recover durable image offload in summary requests 2026-09-11 10:56:48 +08:00
j-xiang 8e8fb2fc62 docs(i18n): proofread bilingual README corpus 2026-09-10 00:21:32 +08:00
Tianyi Cui e11b2a5558 merge: refresh session-log-v3 from master 345248f470 2026-09-08 21:15:06 +08:00
Turtle f1c1ff5a08 docs: constrain package README summaries 2026-09-08 13:52:40 +08:00
Tianyi Cui c988a6796f feat(agent-loop): append system prompt changes on capable routes
Consolidate the existing in-history capability, loop, UI and artifacts into one baseline. Preserve the reviewed tree so each independently verified correction has a subsequent rationale-rich commit.
2026-09-07 11:09:39 +08:00
Tianyi Cui 8bde913784 refactor(compaction): replay the derived system head as a message
The session already derives the protected system head as a Message, and both adapters accept leading system history. Passing it through a separate SummarizationInput.system string unnecessarily flattens that value and rebuilds the same wire message in the adapter. Prepend the derived head to messages and remove textContent, the separate field, and GenerateOptions.system plumbing from the summarizer.

Keep range selection, shadowed seq accounting, session head protection, routed tools, image references, target policy, and the model-visible compaction instruction unchanged. Empty-content heads still derive to null and contribute no request message, but their surface node remains protected. Update subclass consumers/tests, EN/ZH package and subsystem prose, and the existing system-prompt surface owning note with refreshed pairing records.

Evidence: pnpm exec vitest run packages/compaction/compaction-basic/tests packages/llm/llm-deepseek/tests/serialize.spec.ts packages/llm/llm-pi-ai/tests/context.spec.ts --coverage --coverage.include='packages/compaction/compaction-basic/src/region.ts' --coverage.include='packages/compaction/compaction-basic/src/summarizer.ts' passed 203 tests in 6 files; both changed sources have 100% statements, branches, functions, and lines. Region-to-default-summarizer cases pin exact prefix and tools for nonempty Unicode/multiline, empty, and absent heads. DeepSeek JSON byte equality and pi-ai context equality pin leading-message vs separate-system equivalence on text and image-capable conversion paths.

pnpm run doc-sync passed all 33 gates including doc-typecheck, documentation build, translation pairing and model-experience checks. git diff --check passed. Own dependencies installed with pnpm install --frozen-lockfile. An initial test iteration used a nonexistent ctx.dispose teardown on the in-memory fixture; corrected to its existing fixture lifecycle and reran successfully. No runtime/model behavior, normalizer marker, main worktree, push, or rebase changes.
2026-09-06 20:42:43 +08:00
Tianyi Cui ee956c720d refactor(session): represent the system prompt as surface node zero
Consolidate the representation-change PR and its rebase reconciliations into one baseline. Preserve the exact tree and keep the in-history feature in the dependent PR. Follow-up fixes remain separate.
2026-09-06 20:42:43 +08:00
Tianyi Cui e974a655a0 Revert "feat(session, agent, web): support same-session message editing" 2026-09-02 21:20:07 +08:00
lsdsjy ef88756f13 feat(session, agent, web): support same-session message editing 2026-09-02 18:26:55 +08:00
Turtle 15f2997bcb cleanup: omit unneeded invariant companions 2026-08-28 13:12:52 +08:00
Magolor 0b5eba0c8d docs: rebuild the documentation skill and standards (#2983) 2026-08-25 23:47:20 +08:00
creatixchu 5183bc2b65 fix(compaction): 摘要收缩改按路由价并补齐定价访问路径
ds-review-bot 首轮意见修复:

- 摘要收缩比较改用所选节点的路由价 shadowedRouteTokenCount,修复图片消息启发式价低于带框摘要时压缩被误拒;日志影子价仍为启发式
- DeepSeek 定价经序列化器同一套 access 解析构建句柄与占位文本,消除逐图数十 token 的低估;uncatalogued 分支 JSDoc 指明复现 projectImagesForTextModel 替换
- llm-replay 在加载时拒绝纯文本模型上的 imageRequestTokens 声明
- contextBreakdown 的 README 与 JSDoc 改为等于 heuristicTokens 之和,不再声称等于路由价 surfaceTokens
2026-08-24 19:48:52 +08:00
creatixchu 42164508c8 feat(llm): 在 compaction 中按路由为图片请求压力计价
Closes #2848.

- dsh-llm 新增 LlmAdapter.imageRequestPricing 同步钩子与 LlmImageRequestPricing/LlmImageRequestPrice 词汇,ctx.llm 按路由解析
- llm-deepseek 用官方公布的 v4 视觉计算器逐句移植(14px patch、3:1 降采样、384 上限、最坏对齐 pad)实现该钩子,复现请求投影的最旧优先 offload 与像素预算缩放;纯几何 requestImageDimensions 上移到 dsh-attachment
- token-meter 表层 fold 存储与路由无关的节点事实,measure() 按生效 envelope 的路由为图片出现处定价;锚点存快照并按同一路由重定价;TokenSurfaceNode 同时携带路由价 tokens 与固定启发式 heuristicTokens
- compaction 触发、保留与选段读取同一套路由价,记录的 shadowedTokenCount 保持启发式以维持 O(1) 投影 fold 一致
- llm-replay 支持按模型的 imageRequestTokens 声明;新增 keyless 的 image-compaction ACP 快照场景端到端验证装配应用
2026-08-24 19:15:30 +08:00
Turtle 1540e76598 docs: capitalize Service Provider across repository 2026-08-13 13:31:30 +08:00
Tianyi Cui a2d0f7f411 refactor: apply repository naming contract
Apply the accepted pre-release package, service, type, directory, and role renames as one repository-wide change.
2026-08-13 00:54:38 +08:00