Repository navigation
✨ feat(review): run combined review heads as one kick per head - #15068
Conversation
Under review.combined_perspectives the planner dropped review_heads, so a hive that wanted several independent models had to run split perspectives and pay heads x perspectives record-only kicks per PR head (10 with two heads and five perspectives) plus the decision kick. Every one re-read the same diff, and review latency and spend scaled with the perspective count instead of the number of independent opinions. With combined perspectives each head now gets one record-only kick carrying every perspective it still owes, using the combined prompt body, and records a JSON array of per-perspective verdicts. Pending entries stay one per (perspective, head), so settling, expiry, head_max_retries, re-asking only the missing perspectives and the decision step are unchanged. Heads are expected by the aggregate in both modes, and the mixed-reviewer-set check now applies in both modes since head dispatch leaves nothing for a plain lane either way. Closes #15058 Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com> Co-authored-by: gregoryhunt <31053269+gregoryhunt@users.noreply.github.com>
|
[APPROVALNOTIFIER] This PR is NOT APPROVED This pull-request has been approved by: The full list of commands accepted by this bot can be found here. DetailsNeeds approval from an approver in each of these files:Approvers can indicate their approval by writing |
Open PR overlap checkBase branch: Conflicts are checked with every PR merged onto the current tip of
|
…mbined heads A reviewer set mixing lanes with and without review_heads loaded fine under combined_perspectives before heads were dispatched there, so rejecting it at load would break existing configs. Keep it loadable and log a boot warning that the plain lanes receive no combined perspectives while the heads cover them. Split-perspective validation is unchanged. Refs #15058 Signed-off-by: scanner <280983584+hivecommons-hive[bot]@users.noreply.github.com>
|
hive/awaiting-ci: CI pending — sweep will merge when green. 🐝 Hive Agent: — hive: agent=scanner backend=copilot model=claude-fable-5.1 copilot=1.0.88 |
|
Thank you for your contribution! Your PR has been merged. We'd love to hear how your experience was: share feedback |
…PR (#15480) A PR row whose linked issue is already closed drew a bare `closed` chip when the hive's merged ledger did not know the closer (a PR the owner merged by hand; a merge lost across a roll), which read as the issue having been closed before its fix merged — hive's #15422 under follow-up #15425 and #15058 under #15431. The issues were in fact closed by GitHub the second #15426 / #15068 merged. The soft-link enrichment now asks GitHub which MERGED pull request closed the issue (closedByPullRequestsReferences; an open PR with Fixes #N is listed there too and is never named) and carries state_reason and closed_by on the link, so the chip reads `closed by #N`. When no merged closer is known it reads `closed · this PR unmerged` with a hover saying a person or another merged PR closed it. Docs state the rule: the hive closes an issue only after its PR has merged. Fixes #15479 Signed-off-by: Andrew Anderson <andy@clubanderson.com>
…ve (#15539) * ✨ review: fan a combined head kick out to one sub-agent per perspective A combined review head (#15068) judges every perspective in one record-only kick, but nothing told it how to spend that kick, so a head walked the perspectives one after another and its kick took the sum of all of them. The combined head prompt now carries a FAN-OUT section: fetch the PR metadata, the diff and any linked issue or contract once, start one sub-agent per perspective in parallel with the diff, the contract text, the shared instructions and that perspective's focus, then collect their verdict objects into the one JSON array and record it once with --record-verdict. Sub-agents post nothing, never call hive-review and do not refetch the PR. Without sub-agents the head reviews the perspectives in sequence; the array contract is unchanged. The wording is backend-neutral (Task tool on Claude Code, sub-agents on Codex). A single-perspective re-ask carries no fan-out section, and plain combined kicks are unchanged. Refs #15058 Signed-off-by: Gregory Hunt <greg@on-board.ai> Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> * 🐛 review: pin combined-head fan-out sub-agents to the head's model The fan-out told a head to start one sub-agent per perspective without naming a model, while the head's independent-review section pins a single sub-agent to the head model. On a backend whose sub-agent facility defaults to another model, per-perspective verdicts could come from a model other than the one the head is stamped with. Each fan-out sub-agent is now launched with the head's model and replaces (not adds to) the pinned independent-review sub-agent; a head that cannot pin it reviews the perspectives itself. Docs and test updated. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Signed-off-by: Gregory Hunt <greg@on-board.ai> --------- Signed-off-by: Gregory Hunt <greg@on-board.ai> Co-authored-by: Gregory Hunt <greg@on-board.ai> Co-authored-by: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Co-authored-by: hivecommons-hive[bot] <280983584+hivecommons-hive[bot]@users.noreply.github.com>
Under
review.combined_perspectives: truethe planner droppedreview_heads, so getting several independent models meant split perspectives and heads × perspectives record-only kicks per PR head (10 for 2 heads × 5 perspectives) plus the decision. Each kick re-read the same diff, so latency and model spend scaled with the perspective count, not with the number of independent opinions.With combined perspectives each head now gets one record-only kick carrying every perspective it still owes (combined prompt body, one JSON verdict array via
--record-verdict): 2 head kicks plus the decision instead of 11. Pending entries stay per (perspective, head), so settling, expiry,head_max_retries, re-asking only missing perspectives, and the decision step are unchanged. Split heads behave exactly as before.Config: heads are now expected by the aggregate in both modes. A combined-mode reviewer set that mixes lanes with and without
review_headsloaded fine before this change, so it still loads; the hive logs areview heads configwarning at boot (config.ReviewHeadsWarnings) that the plain lanes receive no combined perspectives while the heads cover them. Split-mode validation, which rejects that mix, is unchanged. No new config gate.Tests: one kick per head carrying all missing perspectives with per-pair pending; a 3-of-5 verdict array settles those pairs and the next plan re-asks only the other 2; the decision fires only once every head settled every perspective, with the same prompt as split heads; the combined mixed set loads and warns, while split still rejects it.
Closes #15058