Skip to content

fix(code-review): keep every BHA partition within the agent budget - #381

Open
mikeangstadt wants to merge 7 commits into
mainfrom
fix/code-review-bha-partition-cap
Open

mikeangstadt wants to merge 7 commits into
mainfrom
fix/code-review-bha-partition-cap

Conversation

@mikeangstadt

Copy link
Copy Markdown
Collaborator

Purpose

/code-review could leave Bug Hunter A partitions with no agent on diffs above the unified threshold (5,000 changed LOC). Two caps were computed independently:

  • route passed the partitioner --max-bha-agents = 9 - 2 - domain_critics (up to 7)
  • arbitrate-budget reserved bha_partitions = min(ceil(LOC/1200), 5)

When the partitioner emitted more partitions than arbitrate reserved, _derive_spawn_agents_from_plan prefix-took the cap and dropped the rest as budget_capped. Those files were never bug-hunted. On standard/deep this surfaced as a required coverage gap (false CHANGES_REQUESTED); on shallow it was silent.

What changed (plugins/code-review, v3.13.0 → 3.13.1)

  • One cap source. The partition stage resolves its own cap from --cr-dir: arbitrate-budget's budget.bha_partitions when present, else _max_bha_partitions_by_loc(diff_data) (shallow). run-prefix and the golden harness no longer compute or pass it; an explicit --max-bha-agents still overrides. _budget_bha_partitions is the one reader (rejects bools), used by partition and derive-spawn-spec.
  • Partition shape. Trivial partitions fold before the budget merges, so they never hold a slot. The force-merge keeps its smallest-combined-LOC rule with a same-type tie-break.
  • reviewBudget clamp. A negative reviewBudget no longer turns the critic slice into a tail-drop.
  • Coverage gaps. One writer (_write_coverage_gaps) keyed by gap_producer; every producer (arbitrate-budget, spawn-spec, verify-spawn, verify-consolidate) replaces only its own findings on every return path, so a resumed walk neither duplicates nor keeps a stale gap. derive-static-spec now records budget_capped as a required gap like derive-spawn-spec. setup creates coverage_gaps.json only when absent.
  • Docs: SCHEMA.md, commands/start.md, skills/spawn-reviewers/SKILL.md, CHANGELOG.

Commands run

uv sync --frozen --group dev
pytest plugins/code-review/tools/python/        # 1508 passed, 3 skipped
ruff check plugins/code-review/tools/python/
pyright --pythonpath .venv/bin/python <changed .py files>   # 0 errors

Counterfactual: test_large_diff_without_domain_critic_spawns_every_partition fails on the pre-fix tree ("partitioner emitted 7 partitions for 5 BHA agents"); test_arbitrated_cap_below_loc_count_bounds_the_partitioner and the trivial/balance tests fail on the intermediate commits.

Reviewed locally with /code-review across four rounds; a final local review of 88bf000 is still running and any findings will land as follow-up commits on this branch.

🤖 Generated with Claude Code

mikeangstadt and others added 7 commits November 3, 2026 10:25
- route handed partition --max-bha-agents = 9 - 2 - <domain critics>
  (7 with none), while arbitrate-budget reserved
  min(ceil(LOC / 1200), 5) Bug Hunter A agents. Above the unified
  threshold (or with it disabled) the partitioner kept up to 7
  partitions, and derive-spawn-spec / derive-static-spec prefix-took
  the cap and dropped the rest as budget_capped, so those files were
  never bug-hunted
- _max_bha_partitions_by_loc is now the single source: route calls it
  with its fleet-slot cap, arbitrate-budget calls it on the same
  diff_data.json, so partitions <= bha_partitions on every path; the
  partitioner's force-merge folds the extra files into fewer
  partitions
- derive-static-spec now reports a budget_capped partition as a
  required coverage gap, as derive-spawn-spec already did, so a drop
  can no longer pass silently on shallow
- Prefix goldens: route.max_bha_agents 7 -> 1 for the small fixtures

Testing: pytest plugins/code-review/tools/python/ (1483 passed,
3 skipped); new route->arbitrate->partition->spawn regression cases
fail on the unfixed source (7 or 6 partitions for 5 agents, 7 for 2)

Risks: Diffs between 5001 and ~6000 LOC with impl/test splits now get
5 larger partitions instead of 6-7, one of which was previously dropped

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Pass 2b force-merge now merges the two smallest partitions of one
  type, falling back to a cross-type pair only when no type has two
  left; with the route cap it is the normal path for partitioned
  diffs, and the old any-type merge folded test files into Opus
  impl partitions
- derive-static-spec resets coverage_gaps.json before writing its
  spawn gaps (shallow has no arbitrate-budget to reset it), so a
  resumed or clean re-run neither duplicates nor keeps a gap
- One module-level _SPAWN_BENIGN_REQUIRED_SKIPS shared by
  derive-spawn-spec and derive-static-spec
- Docstring, SCHEMA.md and CHANGELOG state the cache-path bound:
  the partitioner may read uncached_diff_data.json (LOC <= full
  diff) and still force-merges to the route's value
- Tests: dropped the non-independent domain-critic parameter and
  fixed its comment; renamed the cross-type force-merge test to
  assert same-type merges, added the cross-type fallback case, an
  even 10 impl / 10 test 5,500 LOC case, and static-spec re-run cases

Testing: pytest plugins/code-review/tools/python/ (1485 passed,
3 skipped); new force-merge and re-run tests fail on 44f837f; the
main regression still fails on 44f837f^

Risks: slot_cap kept: a negative reviewBudget makes it bind below 5

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- route clamps defaults.reviewBudget at 0. A negative value made
  [:max_domain_critics] a tail-drop slice (-1 kept 6 of 7 matching
  critics) and, through 9 - 2 - critics, shrank the BHA count
- With at most one domain critic, 9 - 2 - critics >= 6 exceeds
  DEFAULT_MAX_BHA_AGENTS, so route now calls
  _max_bha_partitions_by_loc(diff_data) exactly as arbitrate-budget
  does; the slot_cap parameter, branch and test are gone
- Docstring notes BHB + Auditor + critic + 5 BHA = 8 stays within the
  9-agent route fleet; SCHEMA.md and CHANGELOG updated to match

Testing: pytest plugins/code-review/tools/python/ (1486 passed,
3 skipped); the negative-budget route test fails on 930a6a5 and the
main regression still fails on 44f837f^

Risks: None identified; no other path reads reviewBudget

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- derive-spawn-spec and derive-static-spec fold a partitions.json with
  more partitions than a positive BHA cap: the partitioner's
  force-merge (_force_merge_to_cap) runs again down to the cap,
  partitions.json and patches_p<N>.txt are rewritten, and the spec
  records folded_partitions plus a fleet note. budget_capped remains
  for a zero cap and for a fold without a diff scope to rebuild
  patches, both still required coverage gaps
- run-prefix passes _max_bha_partitions_by_loc(diff_data.json) when no
  route value exists instead of omitting --max-bha-agents
- The partitioner merges trivial partitions before the force-merge,
  and the force-merge takes the smallest combined-LOC pair, preferring
  a same-type pair only on a tie
- setup writes an empty coverage_gaps.json per run; the spawn-spec
  stages replace their own spawn_* findings (_replace_spawn_gaps)
- Comments and SCHEMA.md name docs-only as the intentional exception
  to partitions <= bha_partitions; benign-skip comment lists the set
- Tests: fold, cache path, missing route value, both balance
  reproductions, critic-independent route count; the even 10/10 split
  test now asserts balance instead of impl/test purity

Testing: pytest plugins/code-review/tools/python/ (1491 passed,
3 skipped); the fold, balance and fallback tests fail on 40b62d8;
the main regression still fails on 44f837f^

Risks: The fold shells out to git diff for patches with the diff
scope from scope.json / auto_incremental.json and cwd as the partition
stage does

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
arbitrate-budget runs before stage_17_partition, so run-prefix now passes
the partitioner coverage.json final.budget.bha_partitions (floored at 1),
falling back to the LOC count on shallow. The post-hoc fold in the
spawn-spec stages and folded_partitions are removed; budget_capped with
its required coverage gap stays as the guard for an externally produced
partitions.json. The golden harness uses the same helper.

Spawn-spec and verify-spawn gap findings carry gap_producer and
producer-prefixed ids; each stage replaces only its own entries on a
re-run, so a resume neither duplicates a gap nor collides ids with
arbitrate-budget's.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
partition derives its Bug Hunter A cap from --cr-dir when --max-bha-agents
is omitted (arbitrated budget.bha_partitions, else the LOC count, floored
at 1), through one bool-rejecting budget reader shared with
derive-spawn-spec. run-prefix and the golden harness no longer pass it.
Trivial partitions fold before the budget-respecting merge, so they no
longer drive a merge or take a slot.

coverage_gaps.json has one writer, _write_coverage_gaps, keyed on
gap_producer for every producer (arbitrate-budget, spawn-spec,
verify-spawn, verify-consolidate). Each clears its own gaps on every
return path. setup creates the file only when absent and warns on
failure.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…r run

setup no longer pre-creates coverage_gaps.json, so the file stays
arbitrate-budget's own completion signal for resume. Untagged gaps from a
3.13.0 run directory are attributed to their producer by shape and
replaced. The spawn-spec stages write their gaps once per run, after the
review-root check. An explicit --max-bha-agents below 1 is floored with
a warning. The golden harness drops its unread max_bha_agents field.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant