Repository navigation
v3.2.0
The native tool loop has always rationed its own evidence gathering against a fixed assumption: about 50 KB of corpus on the first turn and 12 KB captured per tool result, no matter which model was listening on the other end. Declaring a model window already sized the review corpus, diff and file budgets to it — but the tool loop kept living hand-to-mouth under defaults tuned for the smallest plausible setup.
3.2.0 makes the window you declare the source of truth for the whole evidence pipeline, front to back. It is a steady minor release: no breaking changes, no renames, nothing you must reconfigure — though if you run large-window models, your reviewer is about to read a lot more of the PR per turn.
primary-model-context-tokens: ${{ vars.PRIMARY_CONTEXT_TOKENS }}
smart-model-context-tokens: ${{ vars.SMART_CONTEXT_TOKENS }}
fallback-model-context-tokens: ${{ vars.FALLBACK_CONTEXT_TOKENS }}Context that actually fits the model
- Declaring a tier's window already sized the review corpus, diff and file budgets; the native-loop budgets were the holdout.
tool-corpus-max-bytesandtool-max-response-bytesnow derive from the declared window too (#923): the first-turn corpus normally lands near 15% of the window (capped at 600 KB) and each tool result near 2% (capped at 64 KB), with envelope guards so the loop's conversation can never overflow the window you declared. - New
fallback-model-context-tokens(#922, #942): the fallback reuses the primary's review corpus, so declaring the fallback window sizes the shared corpus to the smaller of the two capacities. A compact fallback model stops receiving a corpus sized for a large primary. - No window declared? Nothing changes: the derived values resolve to exactly the previous defaults (50000 and 12000).
- The honest flip side: if you already declare large windows, expect materially larger review and tool context, and the token spend that comes with it. Explicit
tool-corpus-max-bytes/tool-max-response-bytesvalues still win — keep them set if you want the old ceiling.
Better evidence, less harness friction
A cluster of fixes with one theme: the reviewer spends its context on evidence instead of fighting its own harness.
gh_apidecodes Contents API file reads instead of handing the model truncated base64 (#926). The model was being shown encoded bytes and asked to review code it couldn't read.- Changed files whose complete diff is already in the review corpus are credited as read (#930) — no redundant tool calls re-reading what the corpus already carried, and no honest review made to look partial.
- Repository-content files are masked before truncation rather than after (#943), so a size cut can no longer land inside what the redaction needed to see.
- Direct smart routes actually run on the selected model (#933) — the smarter model you configured is the one that shows up.
- The availability-fallback corpus is bounded by the fallback model's declared capacity (#942).
Operator-mode foundations
#946 lands the typed groundwork for the self-hosted Operator — the Renovate-style, multi-repository reviewer: a strictly validated, fail-closed operator instance config; named model/executor/evidence profiles that carry credential references, never inline secrets; a deterministic adoption policy; and an ecfg-v1 effective-config fingerprint that future review jobs will hash for identity. Repository-referenced instruction files are read from the trusted base side — byte-capped, symlink-refusing, resolved from the base ref, never the PR head.
None of this is wired into the normal Action runtime yet. It is scaffolding, exercised by tests and consumed by the controller work ahead; Action users should notice exactly nothing today.
Small things that make operating it nicer
- Skipped reviews now say why, in the log and a step-summary line (#944, #903). Silent skips were a support-ticket generator.
- Published reviews carry the action version in both the engine line and the managed metadata marker (#941), so "which build wrote this review" stops being an archaeology question.
action_versionis an additive key in the marker (the schema stays version 1) — custom parsers should tolerate it. For the same reason, note that theverdict-sourceoutput may now readenforcementwhen a fail-closed layer forced the verdict. - Claim extraction got sharper: whole-sentence claim units, body claims outranking titles, enumerated items handled (#939).
- Requirement tracing is scoped to the requirements actually relevant to the change (#937).
- Image-digest-only manifest refreshes — the pin bumps that change nothing but digests — now classify correctly (#925, #909).
- The README split into focused pages with a generated inputs reference (#917), and the Marketplace listing metadata is refreshed (#918).
Toward Tangled
Another read-side step for the Tangled platform (#854, #584): a Bobbin/XRPC read client and a canonical pull resolver, with pulls addressed by AT-URI/CID and resolution that fails closed on ambiguity. This is groundwork — Tangled review support is not complete yet.
Upgrading
- No declared model-context windows? You should see no context-budget behavior change from #923 — the derived values resolve to the previous defaults.
- Declaring large windows? Expect larger review and tool context, and the token spend that comes with it.
- Running primary + fallback models? Consider setting
fallback-model-context-tokensso the shared corpus fits the smaller window. - Existing explicit
tool-corpus-max-bytes/tool-max-response-bytesremain authoritative. Remove them only if you want window-derived budgets. @v3follows the newly released v3.2.0 build. Immutable-version and SHA pins must bump deliberately.- No input renames, removals, or other major-version migration. Every input's current contract lives in docs/inputs.md.
Full changelog: v3.1.0...v3.2.0
Generated changelog (release-please)
Features
- build: stamp action version into the engine line and marker (#941) (c0721cf)
- config: define the operator instance/repo config and adoption contract (#946) (1529026)
- context: derive context budgets from each tier's declared window (#923) (3ef1fd2)
- v3: add Tangled Bobbin read client and canonical pull resolver (#584) (#854) (4899ece)
Bug Fixes
- action: log the precheck skip reason and write a step summary line (#944) (69a91d7), closes #903
- claims: whole-sentence claim units, body-claim priority, item enumeration (#939) (a625193), closes #898
- classification: classify image-digest-only manifest refreshes (#909) (#925) (7f6ab64)
- deps: update dependency @types/node (24.19.0 → 24.19.1) (#936) (0562534)
- eval: carry env-only stage knobs so blind replays really skip human reviews (#928) (#929) (14900d9)
- requirement-trace: scope the trace to relevant requirements (#937) (b0a3d6e)
- review: bound the availability-fallback corpus by the declared fallback capacity (#942) (0162197)
- review: use selected model for direct smart routes (#933) (d5923ee)
- test: stop the semantic fixture racing git's background gc (#920) (#945) (19f0f6b)
- tools: credit changed files whose full diff is already in the review corpus (#930) (0a5c700)
- tools: gh_api decodes Contents API files instead of returning truncated base64 (#926) (d1fdbd4)
- tools: mask repo_contents files before truncating (#943) (861a326)