fix: shell test env isolation, Gemini worker mode, improve loop sequencing - #113
Closed
laird wants to merge 25 commits into
Closed
fix: shell test env isolation, Gemini worker mode, improve loop sequencing#113laird wants to merge 25 commits into
laird wants to merge 25 commits into
Conversation
…g to N-clean; scope + exercise preflight; swarm cadence)
Design spec for adding a human-gated planning capability to the manager session (brainstorm → spec → critical review → decompose to stories → existing fleet implements), plus renaming /fix→/dev and /fix-loop→/dev-loop with aliases. Reuses existing skills and the decomposed/subtask convention; adds two backlog-aware behaviors to the manager loop. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Critical-design-review round 1 approved the planning-pipeline spec as-is: 0 literal-wrongness findings, 0 forced decisions. Empirically confirmed the two load-bearing hand-off facts (subtask is claimable; decomposed parents are excluded from claiming). Recorded the CDR approval in the spec status and carried the three CIR-deferred items forward as notes B8/B9 (+ existing B5). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
…eview-1 to autocoder-planning-pipeline-implementation-plan
Tracks the 2026-07-28 handoff documenting the /fix→/dev rename and peters-toolkit:bugfix integration, plus the still-unaddressed critical review of the planning-pipeline implementation plan. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…aim lock until a terminal outcome
…ver reached the ship branch
…ini/droid markers, stale comments)
…rs.sh (GraphQL replay lands new files 644)
… exclusion, drift repair), marketplace to 3.32.0
…ound/improved/learned per round, verdict chapter)
…e v2 removal Atlassian removed POST /rest/api/2/search from Jira Cloud (HTTP 410, CHANGE-2046). list and any-claimable now use POST /rest/api/3/search/jql with an explicit fields list, nextPageToken pagination (no startAt), and existence-of-first-page instead of the removed total for any-claimable. Issue lifecycle endpoints stay on v2 (not removed; plain-string bodies). The fake serves the new contract, enforces a settable page cap to exercise the token loop, and returns 410 on v2 search so regressions cannot pass. Live verification against real Jira Cloud pending by coordinator.
…text Live smoke showed /rest/api/3/search/jql returns description as an ADF document object, not v2's plain string. The list reshape now walks ADF content nodes collecting text (paragraphs joined with newlines; tolerates null, plain strings, malformed nodes), so body stays a plain string for all consumers. v2 CRUD read paths are untouched (still plain strings). The fake now serves ADF descriptions from v3 search to pin the contract.
…mp autocoder 4.10.1 / marketplace 3.32.1
…-smoke-test.sh; live validation pending an ADO sandbox)
…gents The SKILL.md had no model guidance. Top-level coordinator (scope/grade/fix decisions) uses the deep model; subagents and swarm workers use the balanced model. Mirrors the autocoder manager/worker tier pattern. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… model selection (#110) * feat(autocoder): add model-config startup confirmation and env-var-driven model selection Autocoder agents now resolve model tiers from env vars → .autocoder.json → built-in defaults (claude-opus-5 / claude-sonnet-5 / claude-haiku-4-5 on Claude Code; gemini-2.5-pro / gemini-2.0-flash on Antigravity). On first run with no model config found, the agent presents the defaults and asks the user to confirm or override before proceeding. The confirmed models are exported as MANAGER_MODEL / WORKER_MODEL / FAST_MODEL and referenced in every Task tool call in the workflow so agents actually use them. Also adds skills/autocoder/references/model-config.md documenting the tier system and how to persist overrides via .autocoder.json. Parallel maintenance: plugins/autocoder/commands/fix.md ↔ .agent/workflows/fix.md Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> * chore: bump autocoder to 4.11.0, marketplace to 3.33.0 Reflects model-config startup confirmation and env-var-driven model selection added in the previous commit. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com> --------- Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
#112) Switch fix.md and model-config.md from full API model IDs (claude-opus-5) to platform tier shorthand (opus/sonnet/haiku for Claude Code, pro/flash for Gemini). Agents inherit credentials from the running session — no separate API keys needed. Rename skills/harden → skills/improve to reflect the general improvement and validation loop, not just security hardening. Bump autocoder 4.11.0 → 4.12.0, marketplace 3.33.0 → 3.34.0. Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
…ncing - tests/test_issue_source_lib.sh: replace VAR=value function-prefix syntax with explicit `export` before the call. Bash restores the caller's env after a shell-function-prefix invocation, undoing the function's own export; using persistent exports avoids this. - plugins/autocoder/scripts/worker-launch-lib.sh: switch Gemini workers from a shell-script wrapper (gemini-fix-loop.sh) to interactive mode with /fix-loop, matching the Antigravity parallel implementation and what test_worker_launch_lib.sh has always asserted. - skills/improve/SKILL.md: enforce one-run-at-a-time in swarm-mode cadence. Remove the 'early rerun to verify' coordinator shortcut that could start a new run while the swarm was still implementing fixes. Coordinator now lands all entangled fixes first, then dispatches the swarm batch; the single rerun at step 5 is the only run for that batch. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
BASH_ENV=/home/laird/.bashenv prepends $HOME/bin to PATH for every bash subprocess. This causes the real gh wrapper to shadow the test stub in capture_search, making all search-string assertions fail (empty capture). Setting BASH_ENV="" for the backend subprocess ensures the stub directory stays first in PATH. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This was referenced Aug 17, 2026
laird
added a commit
that referenced
this pull request
Aug 17, 2026
Removed the "early rerun to verify" coordinator shortcut from the swarm-mode cadence, which could start a new exercise run while the swarm was still implementing the current batch's fixes -- wasting the run and making grading ambiguous about which fixes it was actually exercising. The coordinator now lands all entangled/seam-uncertain fixes first, dispatches the independent findings to the swarm, waits for the batch to fully complete, then does a single batch acceptance rerun. Includes the packaged mirror at plugins/autocoder/skills/improve/SKILL.md (tests/test_skill_packaging.sh requires the two stay identical). The other three fixes originally bundled in PR #113 (test_issue_source_lib.sh explicit-export syntax, worker-launch-lib.sh Gemini interactive mode, test_issues_gh_search.sh BASH_ENV clearing) are already present on current master -- each conflicted on cherry-pick with only a comment-wording difference, no functional diff. Reapplying them would have been a no-op.
Owner
Author
|
Superseded by #129, which lands the one genuine fix from this PR's four bundled changes. Checked each of the four independently on cherry-pick onto current master:
Closing rather than merging to avoid a no-op re-diff on (1), (2), (4). |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
test_issue_source_lib.sh: ReplaceVAR=value funcprefix syntax with explicitexportbefore the call. Bash saves/restores per-call env for shell functions, undoing the function's ownexporton return; persistent exports avoid this.worker-launch-lib.sh: Switch Gemini workers from a shell-script wrapper (gemini-fix-loop.sh) to interactive mode with/fix-loop, matching the Antigravity parallel implementation and whattest_worker_launch_lib.shhas always asserted.skills/improve/SKILL.md: Enforce one-run-at-a-time in the swarm-mode cadence. Remove the "early rerun to verify" coordinator shortcut that could start a new run while the swarm was still implementing fixes. Coordinator now lands all entangled fixes first, dispatches the swarm batch, waits for completion, then does the single batch acceptance rerun at step 5.test_issues_gh_search.sh: ClearBASH_ENVbefore running the backend subprocess incapture_search.BASH_ENV=/home/laird/.bashenvprepends$HOME/binto PATH inside every bash subprocess, causing the realghauth-switching wrapper to shadow the test stub.Test plan
bash tests/test_issue_source_lib.sh→ok issue-source-libbash tests/test_worker_launch_lib.sh→ok worker-launch-libbash tests/test_issues_gh_search.sh→Results: 15 passed, 0 failed🤖 Generated with Claude Code