8000 [pull] main from can1357:main by pull[bot] · Pull Request #121 · ericrihm/oh-my-pi · GitHub
[go: up one dir, main page]

Skip to content

[pull] main from can1357:main - #121

Merged
pull[bot] merged 19 commits into
ericrihm:mainfrom
can1357:main
Sep 25, 2026
Merged

pull[bot] merged 19 commits into
ericrihm:mainfrom
can1357:main

Conversation

@pull
@pull pull Bot commented Sep 25, 2026 •
Copy link
Copy Markdown

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 : )

recon and others added 19 commits September 25, 2026 15:47
…iction counter

Per review: the advisor's anchor search now starts after the newest usage report made at or before the latest prunedAt, so the gate falls back to the local estimate. Drops evictedSinceAnchor, its resets, coveredBefore and coveredTokensSaved.
…rune

Also drops the isWorthPruning dependency (#13128 was closed) in favor of a local MIN_EVICT_TOKENS, and dedupes the changelog entry.
Adaptive and interleaved thinking share max_tokens with the answer, but streamSimple passed a caller's maxTokens through as the whole request cap. At high effort the thinking pass could spend all of it: an on-demand compaction capped at 13,107 tokens on Opus 5.5 at max effort ended at max_tokens with no compaction block, and every native compaction of that session fell back. Treat the cap as the requested output on every thinking path and add the effort's thinking budget on top, bounded by the model ceiling, as the budget-only path already did. Uncapped requests keep the provider default.
…s no block

A compaction response without a compaction block was logged as 'carried no compaction block' with nothing else, so a request that ran out of max_tokens looked the same as one that called a tool or was refused. Append the stop reason.
…r good

A speculative native compaction that failed for a reason a retry would hit again (an exhausted output budget, a tool call, a refusal) was re-sent at every maintenance boundary, and the threshold pass kept deferring to the retries until the grace cap. One session re-sent a ~600k-token Anthropic compaction 10 times over 31 minutes before falling back. Remember the failure for the session and model until a compaction commits: speculation stops re-sending it, the threshold pass no longer defers to it, and it starts from the next usable method. Retriable failures keep speculating.
Extension factories are rebound to every subagent session, but handlers
could not tell whether they ran in the top-level session or a subagent.
Pass the identity the SDK already resolves (kind, registry id, agent
definition name, task depth, parent id) to the ExtensionRunner and
expose it on ExtensionContext.
…eview, add advisor.evictStaleResults

- Move the prunedAt anchor rule into transcript-tokens.ts (skipPrunedAnchors) instead of copying it
- Reuse compaction's MIN_PRUNE_TOKENS
- Leave recall and other non-re-derivable tool results in place
- Never evict the review that just finished
- Add advisor.evictStaleResults (default on)
…d reviews (@alnaggar-dev)

feat(advisor): evict stale tool results from finished reviews
…equests (@alphastorm)

fix(ai): add thinking headroom to capped anthropic requests
…on that failed for good (@alphastorm)

fix(coding-agent): stop re-sending a native compaction that failed for good
…y as ctx.agent (@andrebrait)

feat(extensions): expose the running agent's identity as ctx.agent
@pull pull Bot locked and limited conversation to collaborators Sep 25, 2026
@pull pull Bot added the ⤵️ pull label Sep 25, 2026
@pull
pull Bot merged commit 533826a into ericrihm:main Sep 25, 2026
Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants

0