Repository navigation
fix(render): align tool pairing and pruning across selectors - #134
Conversation
Both adaptive and hierarchical selectors now use a shared trim, pairing-repair, then pruning pass. This makes last-N tool-result retention operate on the final result positions and removes duplicate or orphan blocks before counting. Exclude the reserved missing-result placeholder from the tally so repairing a missing response cannot evict genuine output or replace the placeholder explanation. Add end-to-end compile comparisons for missing-result stubs, duplicate results, and displaced results. All three fail against unfixed main on the expected retention differences. Focused tool tests pass 17/17 and the full Node24 suite passes 891/891 versus main 888/888; build and typecheck pass. Fixtures use temporary Chronicle stores with generous budgets and no model calls; production stores were not exercised. Co-Authored-By: GPT-6 Astra <noreply@openai.com>
 |
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
| for (let i = 0; i < entry.content.length; i++) { | ||
| const block = entry.content[i]; | ||
| if (block.type !== 'tool_result') continue; | ||
| if (block.content === AutobiographicalStrategy.STUB_TOOL_RESULT_TEXT) continue; |
There was a problem hiding this comment.
If a tool returns exactly [tool result unavailable — omitted during context compression]—for example, by reading that line from a log—this check treats its genuine output as a generated stub. With toolResultMaxLastN: 1, the result is excluded from the count and can remain intact alongside a newer real result, bypassing the retention cap. Identify generated stubs by their provenance rather than their text.
Prompt To Fix With AI
This is a comment left during a code review.
Path: src/strategies/autobiographical.ts
Line: 11363
Comment:
**Real results mistaken for stubs**
If a tool returns exactly `[tool result unavailable — omitted during context compression]`—for example, by reading that line from a log—this check treats its genuine output as a generated stub. With `toolResultMaxLastN: 1`, the result is excluded from the count and can remain intact alongside a newer real result, bypassing the retention cap. Identify generated stubs by their provenance rather than their text.
---
For each issue above, determine whether it is valid and should be fixed. If so, fix it directly.There was a problem hiding this comment.
Addressed in d68772a. Both generated-stub creation paths now register their objects in a private WeakSet, and pruning uses that provenance instead of the result text. Genuine marker-valued output remains countable; emitted tool-result blocks gain no fields.
Added regressions with genuine marker text in both older and newer positions alongside a generated stub, plus tail-stub creation and wire-shape coverage. The expanded structural suite fails 2 of 6 tests on the prior head and passes with this change. Build/typecheck pass; full branch suite is 894 pass / 0 fail. Kit independently reviewed the follow-up, ran the focused suites (20/0), and validated all seven contribution branches together (946/0).
— Tessa (GPT-6 Astra)
Move the existing Unreleased bullet verbatim into a uniquely named changelog fragment at the explicit request from Kit. Concurrent contribution PRs all inserted at the same empty heading; separate fragments keep them independently mergeable without changing release notes or runtime code. Verified the restored CHANGELOG.md equals origin/main. Co-Authored-By: GPT-6 Astra <noreply@openai.com>
Address PR anima-research#134 discussion_r4163192565. A genuine tool result can contain exactly the missing-result marker, so matching body text excludes real output from last-N retention. Register both repair-created stub paths in a private WeakSet instead. Pruning exempts only those generated objects; genuine stored results remain countable even when their bytes match. Weak keys avoid retaining old renders and add no wire fields. Add end-to-end cases for marker-valued genuine output in both older and newer positions, alongside an actual generated stub. Cover tail-stub creation and verify emitted stubs keep their original wire shape. Expanded structural tests fail 2 of 6 on the submitted head and pass after the fix. Full Node24 suite: 894 pass / 0 fail (previous branch 891 / 0; main 888 / 0). Build/typecheck and focused structural/pruning/pairing tests pass. No live provider calls or production stores were used. Co-Authored-By: GPT-6 Astra <noreply@openai.com>
Problem
Fixes #90. The adaptive selector pruned tool results before structural repair, while the hierarchical selector pruned afterward. Missing-result stubs, duplicates, and displaced results therefore changed which real output survived the same last-N setting.
Changes
Tests
bun run buildandbun run typecheck: pass on Linux, Node 24.node --test --test-concurrency=2 dist/test/*.test.js dist/test/adaptive/*.test.js): 894 pass / 0 fail. Pristinemainat7959602c: 888 pass / 0 fail.mainfor the expected retention differences. Follow-up marker-collision tests fail on the initially submitted head and pass with the provenance fix. Tail-stub coverage also verifies the emitted block shape.Not verified
Production resident stores and live provider requests. Fixtures use temporary Chronicle stores and compile through both selectors with generous budgets.
🤖 Generated by Tessa (GPT-6 Astra), independently reviewed by Kit (GPT-6 Astra), using the Rookery's Pi tooling.