mirror of
https://github.com/MoonshotAI/kimi-code.git
synced 2026-08-06 23:36:06 +00:00
3c75a27da6
* feat(agent-core-v2): detect prompt-cache breaks from per-step usage and emit telemetry Track consecutive turn-scoped LLM requests per agent; when the cache-read token count drops by more than 5% and by more than 2000 tokens between requests, log a debug line and emit cache_break_detected with both usages, the drop ratio, and the interval. Operation requests (e.g. compaction) act as a baseline barrier so expected drops are not reported. * feat(tui): add cache-expiry hint dialog for resumed and idle sessions (v2 engine) Resuming a long-idle session or submitting after a long idle stretch re-sends the whole history with an expired context cache. Show a dialog offering to compact, start a new session, continue as-is, or never ask again (persisted as cache_expiry_hint in tui.toml). Thresholds come from the client_configs endpoint (estimated_cache_duration) via a generic per-name cached client; only OAuth-managed providers participate. * fix(tui): preserve submit order and revalidate session in cache-hint flows Cold-cache submits during the in-flight config fetch are now swallowed and replayed through a FIFO chain, so a later prompt can never overtake the stashed one. Both the resume and idle paths re-check the current session after the async fetch: a switch mid-flight drops the dialog (resume) or hands the stashed input back to the editor instead of sending it into the wrong session (idle). * chore(agent-core-v2): regenerate state manifest after merging main * fix(tui): apply cache_expiry_hint on /reload and /reload-tui * fix(agent-core-v2): skip unmeasured all-zero usage in cache break detection * fix(tui): restore chained cache-hint submits when the dialog is not sent When several submits are swallowed during the cold-config fetch and the first dialog is dismissed (or its compact/new action fails), the stashed inputs were restored while later chained submits were still released — reordering the conversation. Chained submits now follow the fate of the message that opened the dialog, and multiple restores append newline-joined instead of overwriting the editor. * fix(agent-core-v2): reset cache-break baseline on model change Caches are per-model, so a cache-read drop after /model is expected, not a break. The baseline now carries the model and only same-model records are compared. * fix(tui): only count LLM-activity replay records for the resume cache hint The v2 resume replay also carries local-only state records (permission, plan, config updates, approval results) that slash commands append without an LLM request. Filter lastActiveAt to message/compaction records so a recent local change no longer masks an expired cache. * style(agent-core-v2): rewrite the cacheBreak impl header per package convention State the domain role, collaborators, and scope instead of narrating implementation steps; the behavior guards now live in the code alone. * fix(tui): drop the resume cache hint when a turn started mid-fetch The resume dialog is fire-and-forget over an async config fetch; if the user already sent the first prompt by the time it resolves, mounting would overlay an active turn and its actions would hit the live session. Re-check streamingPhase/isCompacting after the await, next to the session check. * style(agent-core-v2): trim the cacheBreak contract header to contract and scope * refactor: report cache-break detection from the TUI client Move the detector out of the engine so the telemetry event carries the client's own identity (which client produced it is now attributable). The TUI observes main-loop turn.step.completed usage directly, with the same guards: first-step/unmeasured/all-zero records skipped, model change and compaction reset the baseline. The agent-core-v2 cacheBreak module is removed. * chore: drop accidentally committed dist-web build output and ignore it * chore: revert the dist-web ignore rule * chore: restore dist-web to the tracked content from main * feat(tui): record cache breaks caused by mid-session model/effort switches A model or effort change mid-session busts the prompt-cache key — that is a real cache break worth attributing, not noise. The baseline now carries model and effort, the same-model exemption is gone, and cache_break_detected reports prev/curr model and effort alongside both usages. * chore(changeset): simplify the cache-expiry hint entry * chore(changeset): trim the cache-expiry hint entry to one line * fix(tui): cache-hint review follow-ups - carry the pre-dialog media extraction through compact/new resends so pasted attachments survive the image-store clear on a new session - reset the cache-break baseline after /undo — the context cut makes the next cache-read drop expected - release the stashed submit when a foreground operation started during the cold-config fetch instead of mounting the dialog over it - count a completed compaction as activity so the next submit is not judged against the pre-compaction timestamp * chore(changeset): drop the v2-engine-only suffix * fix(tui): seed the activity baseline when the resume check skips * fix(tui): cache-hint review follow-ups * fix(tui): record cache activity on completed steps, not turn begin * feat(cli): persist the client-configs cache across restarts |
||
|---|---|---|
| .. | ||
| .vitepress | ||
| en | ||
| media | ||
| public | ||
| zh | ||
| .gitignore | ||
| AGENTS.md | ||
| index.md | ||
| package.json | ||