Repository navigation
[pull] main from can1357:main - #127
Merged
Merged
Conversation
Port earendil-works/pi's prompt-cache warming (proposal in #12691): the main agent loop replays its last request with a one-token output budget shortly before the prompt-cache entry expires, so idle gaps do not force a full-prefix cache re-write. - session/cache-warmer.ts: cost-aware scheduler refreshing at 90% of the declared entry lifetime (>=10s margin) only when the expected avoided-miss cost clears the refresh cost by $0.05; streaming/idle phases with 15%/100% continuation probability; 30/60-minute safety windows; cache_warming_decision extension override event - catalog: Model.promptCache per-retention-tier lifetimes; direct Anthropic rows annotated 5m/1h; models.yml promptCache opt-in plumbed through ModelPatch and custom-model builds; models.json regenerated with the annotation only (zero unrelated drift) - providers.cacheWarming setting (off/streaming/idle, default idle) replaces the Anthropic-only fixed keep-alive refresh loop, which is removed along with its stream-level scheduling machinery Resolves #12691
- models.yml promptCache now reaches models: ModelPatch carries the field, applyModelPatch copies it, and custom-model overlays/finalize pass it through (previously accepted by the schema and dropped) - the warmer mirrors the Anthropic provider's OAuth retention default: OAuth requests on long-retention-capable models schedule against the 1h tier instead of 5m - extension decide calls race a 2s timeout so a slow extension cannot push the replay past the expiry margin - session dispose cancels the warmer and detaches onWarmed before the drain windows; the mode-change subscription is unwound too - arming no longer string-matches session ids: every request through the session's own streamFn is its main loop, and spawned/one-shot sessions (task subagents, standalone compaction, agentic commit) opt out via the new CreateAgentSessionOptions.cacheWarming flag - prompt-cache lifetimes move to catalog rules (prompt-cache node in rules/providers/anthropic.kdl, applied through applyCatalogAssignments), replacing the generator-side TS pass - isCurrent uses the upstream full-prefix identity check - the TTL test scrubs PI_CACHE_RETENTION and new type assertions are removed (narrowed deps.stream contract, typed thinking fixture) Ref #12691
# Conflicts: # docs/models.md # packages/ai/src/providers/anthropic.ts # packages/ai/test/anthropic-cache-refresh.test.ts # packages/catalog/src/compat/rules.json # packages/catalog/src/models.json
… fmaf on x86-64-v2 The baseline x64 addon (the only x64 build for win32 and darwin) targets x86-64-v2, which has no FMA; f32::mul_add there compiles to a scalar fmaf libcall per element and blocks vectorization. The Q8/f32 dot products, dot4, RoPE and attention value accumulation now use plain multiply-add.
… strategy (@KamijoToma) feat(coding-agent): port upstream pi's cache-warming strategy Maintainer fixes applied while merging: - stop warming after a refresh that misses the cache or fails - cut each replay off at its first generated token (thinking lifts max_tokens to the budget) - adopt the retention tier the response reports writing (resolver-keyed OAuth sessions) - idle warming covers entries shorter than the 30-minute window only - invalidate on model/system prompt/tools/thinking/retention changes - register providers.cacheWarming in the settings registry; replay via primaryStreamFn - changelog attribution, breaking-change section, merge-artifact cleanup
…g mul_add to libm fmaf on x86-64-v2 (@H4vC) perf(pi-predict): stopped SmolLM CPU kernels lowering mul_add to libm fmaf on x86-64-v2
# Conflicts: # packages/catalog/src/compat/rules.json # packages/catalog/src/models.json
- Restore lazy loading for the OpenAI grounded provider - Order API-billed openai/* after every openai-codex entry - Surface structured OpenAI error type/code/param without the message body - Hard-cap sources/citations at the requested result count - Omit zero searchRequests; reject OAuth consistently - Document ignored params and caps; add catalog/tui changelog entries - Replace the tautological chain test with a billing-precedence contract test
…toli-tsinovoy) feat(web-search): add OpenAI API-billed search
…g mul_add to libm fmaf on x86-64-v2 (@H4vC)
…mp update On Windows, Bun's rejected Bun.file().text() of a missing file holds no event-loop handle, so the loop drained mid-await and the unsettled-entry guard failed `omp update` with exit 1 when ~/.npmrc and ~/.bunfig.toml were absent.
… fs.promises in omp update (@H4vC)
- Deferred the unsettled-entry verdict from beforeExit to exit, reporting only when the last beforeExit was final (an unref'd probe timer clears the mark if the loop turns again). - Windows Bun emits beforeExit while omp update still has I/O in flight and then completes the command; it no longer exits 1 with a false diagnostic. - Added regression tests for a loop resumed after beforeExit, completing or exiting explicitly. - Isolated the test child from inherited XDG dirs. Fixes #13470
…date path - npm-registry.ts: switched to the async-only node:fs/promises import per AGENTS.md - update-cli.ts: addBunCacheActualDir reads package.json via fs.promises so a missing manifest cannot drain the loop on Windows (#13470) - CHANGELOG: moved the #13470 entry back to Unreleased after rebase and linked the issue on the readOptional entry
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to subscribe to this conversation on GitHub.
Already have an account?
Sign in.
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
See Commits and Changes for more details.
Created by
pull[bot] (v2.0.0-alpha.4)
Can you help keep this open source service alive? 💖 Please sponsor : )