fix(commands,workflows): make the planning step read issue comments as authoritative - #2404
Conversation
…s authoritative Three reliability fixes to the fix-github-issue pipeline, all found by running it against real issues and reading the resulting event streams. **Planning ignored decisions recorded in comments.** `fetch-issue` already pulls `comments`, and `archon-investigate-issue` already listed them as context — but nothing said they outrank the issue body. On 2026-08-03 a run implemented an issue's body while a maintainer comment on that same issue, present in the run's fetched input and posted 16 seconds before it started, specified a different design. The PR was discarded. Both planning entry points now state that a later maintainer comment supersedes the body, that a comment stating a decision IS the specification, and that linked issues are followed one level for their decisions. **A failed gate silently reduced review coverage.** `review-gate` read `$smoke-validate.output.claims_accurate` with strict field access. When smoke-validate timed out and emitted nothing, the gate failed, every conditional reviewer skipped via trigger_rule, and the run still reported success with one reviewer instead of the intended set. The gate now reads the whole output defensively, runs under `trigger_rule: all_done` so a failed producer still reaches it, and biases toward MORE review whenever the claims are not positively confirmed. **The PR number was recorded by the model and was usually wrong.** `create-pr` asked the model to echo `gh pr view --json number` into `$ARTIFACTS_DIR/.pr-number`. Across four parallel runs it wrote the ISSUE number twice, skipped the file once, and was correct once — so `review-scope` resolved a nonexistent PR, correctly refused to fabricate a scope, and the classifier then declined every specialist for lack of evidence. Three of four PRs got one generalist reviewer while the run reported success. Replaced with a deterministic bash node that resolves the PR from the branch and fails loudly. Verified in the exact worktree that failed: the node returns 2388 where the model wrote 2323. `synthesize` also moves from `one_success` to `all_done` — defensive rather than causal, but `one_success` permits a synthesis to stand on a subset and present it as the whole review.
📝 WalkthroughWalkthroughThe PR adds issue-comment authority rules and one-level linked-issue review. It also updates the experimental workflow with dependency preparation, typed outputs, implementation and PR guards, and fail-safe conditional review orchestration. ChangesIssue-aware workflow
Estimated code review effort: 4 (Complex) | ~45 minutes Sequence Diagram(s)sequenceDiagram
participant Operator
participant Workflow
participant Worktree
participant GitHub
Operator->>Workflow: Submit issue request
Workflow->>Worktree: Install locked dependencies
Workflow->>Worktree: Implement issue
Workflow->>Worktree: Verify changes or commits
Workflow->>GitHub: Create draft PR
Workflow->>GitHub: Capture PR number and URL
Workflow->>Workflow: Run smoke validation and review gate
Possibly related PRs
🚥 Pre-merge checks | ✅ 5✅ Passed checks (5 passed)
✨ Finishing Touches🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 5
🧹 Nitpick comments (1)
.archon/workflows/experimental/archon-fix-github-issue-experimental.yaml (1)
409-435: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick winSuperseded advisory PR-number writing in
create-prshould be simplified.
capture-pr-numbernow deterministically overwrites$ARTIFACTS_DIR/.pr-numberand validates the result, which fully mitigates the failure modes this comment block documents.create-pr's own Step 7 (unchanged, outside this diff) still instructs the model to compute and write$ARTIFACTS_DIR/.pr-numberitself — the exact advisory mechanism this new node was added to replace. Sincecapture-pr-numberalways runs afterward and unconditionally overwrites the file, Step 7 is now dead weight that only adds confusion about which node is authoritative for PR identity.Consider removing or simplifying create-pr's Step 7 to avoid the redundant, now-superseded instruction.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In @.archon/workflows/experimental/archon-fix-github-issue-experimental.yaml around lines 409 - 435, Remove or simplify Step 7 in the create-pr workflow so it no longer instructs the model to compute or write $ARTIFACTS_DIR/.pr-number. Keep capture-pr-number as the sole authoritative PR-number capture and validation step, preserving its existing behavior.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In @.archon/commands/defaults/archon-create-plan.md:
- Around line 39-45: Update the input classification table in
archon-create-plan.md to add GitHub issue number and URL cases before the
Free-form text branch. Match the number and URL detection patterns and routing
used in archon-investigate-issue.md, ensuring these inputs reach the
comment-fetch path before Phase 0.3.
- Around line 47-52: The comment-authority rules in
.archon/commands/defaults/archon-create-plan.md lines 47-52 and
.archon/commands/defaults/archon-investigate-issue.md lines 54-62 must require
verified maintainer authorship before treating a comment as authoritative, and
must state that the latest verified maintainer decision takes precedence. Update
both command documents consistently while preserving the existing planning and
investigation guidance.
In @.archon/commands/defaults/archon-investigate-issue.md:
- Around line 63-68: Update the linked-issue guidance in
.archon/commands/defaults/archon-investigate-issue.md lines 63-68 and
.archon/commands/defaults/archon-create-plan.md lines 53-55 to preserve the
owner/repo parsed from full GitHub issue URLs and demonstrate gh issue view with
--repo owner/repo for URL references; explicitly distinguish this from bare
`#1234` references, which should use the current repository.
In @.archon/workflows/experimental/archon-fix-github-issue-experimental.yaml:
- Around line 309-346: Update the rationale comment above assert-implemented to
accurately state that the guard checks both uncommitted working-tree changes
outside .archon/ and commits ahead of origin/$BASE_BRANCH via the commits count.
Remove or revise the claim that only the working tree is inspected, and explain
the commit-count fallback’s purpose without changing the bash logic.
- Around line 58-75: Update the prep-worktree prompt so a failed locked
dependency installation deterministically fails the node rather than merely
being reported in text. In the prep-worktree instructions, require the model to
stop and explicitly fail/exit when the install command returns non-zero, or add
a follow-up deterministic shell check that propagates the install exit status;
preserve the existing manager detection and successful-install reporting
behavior.
---
Nitpick comments:
In @.archon/workflows/experimental/archon-fix-github-issue-experimental.yaml:
- Around line 409-435: Remove or simplify Step 7 in the create-pr workflow so it
no longer instructs the model to compute or write $ARTIFACTS_DIR/.pr-number.
Keep capture-pr-number as the sole authoritative PR-number capture and
validation step, preserving its existing behavior.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro Plus
Run ID: 99f4bbfd-3015-4276-8cb2-0c6a67179c66
⛔ Files ignored due to path filters (1)
packages/workflows/src/defaults/bundled-defaults.generated.tsis excluded by!**/*.generated.*
📒 Files selected for processing (3)
.archon/commands/defaults/archon-create-plan.md.archon/commands/defaults/archon-investigate-issue.md.archon/workflows/experimental/archon-fix-github-issue-experimental.yaml
| ### 0.1a If the input is a GitHub issue: comments outrank the body | ||
|
|
||
| Fetch the issue **with its comments**, and treat them as authoritative: | ||
|
|
||
| ```bash | ||
| gh issue view {number} --json title,body,labels,comments,state,url,author | ||
| ``` |
There was a problem hiding this comment.
🎯 Functional Correctness | 🟠 Major | ⚡ Quick win
Classify GitHub issue inputs before the free-form branch.
The input table above lists only PRD, document, and free-form inputs. It has no GitHub issue number or URL case. Therefore, 1234 or a GitHub issue URL can enter Phase 0.3 and skip this new comment-fetch path. Add number and URL cases before Free-form text, matching .archon/commands/defaults/archon-investigate-issue.md Lines 30-32.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In @.archon/commands/defaults/archon-create-plan.md around lines 39 - 45, Update
the input classification table in archon-create-plan.md to add GitHub issue
number and URL cases before the Free-form text branch. Match the number and URL
detection patterns and routing used in archon-investigate-issue.md, ensuring
these inputs reach the comment-fetch path before Phase 0.3.
| - **Read every comment before planning.** A later maintainer comment supersedes | ||
| the body wherever they disagree — the body is the opening position, not the | ||
| specification. | ||
| - **A comment stating a decision IS the specification.** If one settles a shape, | ||
| a field list, an interface, or an explicit "do X, not Y", plan that. Do not | ||
| re-derive the choice from the body and do not compromise between them. |
There was a problem hiding this comment.
🔒 Security & Privacy | 🟡 Minor | ⚡ Quick win
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
set -euo pipefail
printf 'Files:\n'
for f in \
".archon/commands/defaults/archon-create-plan.md" \
".archon/commands/defaults/archon-investigate-issue.md" \
".archon/commands/defaults/archon-triage-stale.md"
do
if [ -f "$f" ]; then
echo "== $f =="
wc -l "$f"
ast-grep outline "$f" || true
fi
done
printf '\nRelevant excerpts:\n'
rg -n -C 3 "verified|maintainer|specification|decision|latest|supersedes|comment" .archon/commands/defaults/archon-create-plan.md .archon/commands/defaults.archon-investigate-issue.md 2>/dev/null || trueRepository: coleam00/Archon
Length of output: 4251
🏁 Script executed:
#!/bin/bash
set -euo pipefail
printf '== .archon/commands/defaults/archon-investigate-issue.md excerpts ==\n'
sed -n '1,110p' .archon/commands/defaults/archon-investigate-issue.md
sed -n '210,360p' .archon/commands/defaults/archon-investigate-issue.md
sed -n '530,602p' .archon/commands/defaults/archon-investigate-issue.md || true
printf '\n= maintainers/verification checks =\n'
rg -n "verification|trusted|maintainer|team|collaborator|read access|write access|GH\\(|gh issue|issue view|GitHub\\(|GitHub" .archon/commands/defaults/archon-create-plan.md .archon/commands/defaults/archon-investigate-issue.md || trueRepository: coleam00/Archon
Length of output: 12406
Clarify comment-authority in both planning commands.
Both commands say any maintainer comment outranks the issue body, but they do not require verifying that the comment author is an authorized maintainer. Restrict the specification rule to verified maintainer decisions and state latest-decision precedence.
📍 Affects 2 files
.archon/commands/defaults/archon-create-plan.md#L47-L52(this comment).archon/commands/defaults/archon-investigate-issue.md#L54-L62
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In @.archon/commands/defaults/archon-create-plan.md around lines 47 - 52, The
comment-authority rules in .archon/commands/defaults/archon-create-plan.md lines
47-52 and .archon/commands/defaults/archon-investigate-issue.md lines 54-62 must
require verified maintainer authorship before treating a comment as
authoritative, and must state that the latest verified maintainer decision takes
precedence. Update both command documents consistently while preserving the
existing planning and investigation guidance.
| - **Follow linked issues.** When the body or a comment points at another issue | ||
| (`#1234`, or a full URL) for a decision, design, or contract, fetch it and read | ||
| its comments too: | ||
| ```bash | ||
| gh issue view 1234 --json title,body,comments,state,url | ||
| ``` |
There was a problem hiding this comment.
🎯 Functional Correctness | 🟠 Major | ⚡ Quick win
🧩 Analysis chain
🏁 Script executed:
#!/bin/bash
set -euo pipefail
echo "== changed files/stat =="
git diff --stat || true
echo "== target snippets =="
for f in .archon/commands/defaults/archon-investigate-issue.md .archon/commands/defaults/archon-create-plan.md; do
if [ -f "$f" ]; then
echo "--- $f"
nl -ba "$f" | sed -n '45,80p'
else
echo "missing $f"
fi
done
echo "== gh issue view occurrences =="
rg -n "gh issue view|url parsing|repo " .archon/commands/defaults/archon-*.md || true
echo "== all linked-issue guidance occurrences =="
rg -n "Follow linked issues|follow the link|gh issue view" .archon/commands/defaults || trueRepository: coleam00/Archon
Length of output: 299
🌐 Web query:
GitHub CLI gh issue view syntax --repo owner/repo full URL support
💡 Result:
The GitHub CLI (gh) issue view command supports both issue numbers and full URLs as valid arguments [1]. Syntax: gh issue view { | } [flags] - You can pass an issue number (e.g., gh issue view 123) or a direct URL to the issue (e.g., gh issue view https://github.com/owner/repo/issues/123) [1]. - When working with a repository other than the current directory's local repo, use the --repo (or -R) flag with the [HOST/]OWNER/REPO format [1][2]. Note on syntax constraints: - The command expects the issue number as an integer. Using the # prefix (e.g., gh issue view #123``) is generally not supported, as many shells interpret the # character as the start of a comment and strip it before the command receives the argument [3]. - While there have been community requests to support `owner/repo#number` shorthand syntax, it is not currently integrated into the standard `gh issue view` command interface [4]. When viewing issues, you can use the `--web` flag to open the issue directly in your default web browser rather than displaying it in the terminal [1][2].
Citations:
- 1: https://cli.github.com/manual/gh_issue_view
- 2: https://man.archlinux.org/man/extra/github-cli/gh-issue-view.1.en
- 3: Clarify error message when someone uses #<num> instead of <num> for pr and issue commands cli/cli#315
- 4: auto-resolve issue/PR shorthand syntax cli/cli#11967
🏁 Script executed:
#!/bin/bash
set -euo pipefail
echo "== changed files/stat =="
git diff --stat || true
for f in .archon/commands/defaults/archon-investigate-issue.md .archon/commands/defaults/archon-create-plan.md; do
if [ -f "$f" ]; then
echo "--- $f"
awk '{printf "%6d %s\n", NR, $0}' "$f" | sed -n '45,80p'
else
echo "missing $f"
fi
done
echo "== relevant occurrences in target files =="
grep -nE "Follow linked issues|gh issue view|owner/repo|`#1234`|issue URL|URL" .archon/commands/defaults/archon-investigate-issue.md .archon/commands/defaults/archon-create-plan.md || true
echo "== broader linked-issue guidance locations =="
grep -RInE "Follow linked issues|follow the link|gh issue view|owner/repo|`#1234`" .archon/commands/defaults || trueRepository: coleam00/Archon
Length of output: 6593
Preserve the linked issue repository when the reference is a full URL.
gh issue view accepts a full GitHub issue URL, so the #1234 example does not cover cross-repository linked issues. In both command docs, make URL parsing preserve owner/repo and show that URL references include --repo owner/repo, while bare #1234 uses the current repository.
📍 Affects 2 files
.archon/commands/defaults/archon-investigate-issue.md#L63-L68(this comment).archon/commands/defaults/archon-create-plan.md#L53-L55
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In @.archon/commands/defaults/archon-investigate-issue.md around lines 63 - 68,
Update the linked-issue guidance in
.archon/commands/defaults/archon-investigate-issue.md lines 63-68 and
.archon/commands/defaults/archon-create-plan.md lines 53-55 to preserve the
owner/repo parsed from full GitHub issue URLs and demonstrate gh issue view with
--repo owner/repo for URL references; explicitly distinguish this from bare
`#1234` references, which should use the current repository.
| - id: prep-worktree | ||
| prompt: | | ||
| Install this project's dependencies so builds, type-checks and tests can run | ||
| in this fresh worktree. | ||
|
|
||
| Follow this order: | ||
|
|
||
| 1. Locate the package-manager markers actually present in the repo — lockfiles | ||
| and manifests. Decide from evidence, never from convention. | ||
| 2. Read the manifest head to confirm the manager and any pinned version. | ||
| 3. Verify that manager responds to a version check before relying on it. | ||
| 4. Install using the manager's locked mode — the variant that reproduces the | ||
| lockfile exactly and fails rather than updating it. The install must not | ||
| modify the lockfile. | ||
| 5. Report one line: the manager you detected, and whether the install succeeded. | ||
|
|
||
| Bound every read; never print a whole file. | ||
| model: small |
There was a problem hiding this comment.
🩺 Stability & Availability | 🟠 Major | ⚡ Quick win
Add a deterministic failure check to prep-worktree.
This node only asks the model to "Report one line: the manager you detected, and whether the install succeeded." A prompt-based report is advisory: if the model's turn completes normally (even after the install command exits non-zero), the node itself is not forced to fail. implement (Line 304) then proceeds against a broken or partially-installed dependency tree, and later build/type-check/test failures get misattributed to the implementation instead of the environment.
This is the same failure mode the PR explicitly calls out and fixes elsewhere in this file: capture-pr-number's rationale states create-pr's advisory reporting is unreliable because "Being a prompt instruction, it is advisory," and assert-implemented's rationale exists specifically because a node "can decline to change anything ... and still report success." Apply the same principle here: instruct the model to explicitly fail/exit the turn (or add a follow-up deterministic bash check of the install command's exit status) when the locked install fails, instead of relying on narrative text.
Proposed prompt addition
4. Install using the manager's locked mode — the variant that reproduces the
lockfile exactly and fails rather than updating it. The install must not
modify the lockfile.
- 5. Report one line: the manager you detected, and whether the install succeeded.
+ 5. If the install command exits non-zero, treat this as a hard failure of this
+ task — do not continue as if it succeeded.
+ 6. Report one line: the manager you detected, and whether the install succeeded.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In @.archon/workflows/experimental/archon-fix-github-issue-experimental.yaml
around lines 58 - 75, Update the prep-worktree prompt so a failed locked
dependency installation deterministically fails the node rather than merely
being reported in text. In the prep-worktree instructions, require the model to
stop and explicitly fail/exit when the install command returns non-zero, or add
a follow-up deterministic shell check that propagates the install exit status;
preserve the existing manager detection and successful-install reporting
behavior.
| # | ||
| # `implement` can decline to change anything — a blocked precondition, an unclear | ||
| # plan — and still report success, so the run continued through create-pr and the | ||
| # whole review fleet on an empty diff, then reported `completed`. Nine nodes ran | ||
| # after the work had already failed. This asks the only question that matters | ||
| # deterministically: is there a change to review? | ||
| # | ||
| # Exiting non-zero fails the node, and therefore the run, before any reviewer or | ||
| # PR node spends anything. A run with nothing to show should fail, not go green. | ||
| # `.archon/` is excluded deliberately. Archon copies the operator's workflow and | ||
| # command edits into every run worktree — that is the feature that lets a workflow | ||
| # be iterated on before it is committed — so those files are present BEFORE | ||
| # `implement` runs and are not its output. A first version of this guard counted | ||
| # them and passed three consecutive runs on an empty implementation, which is the | ||
| # exact failure it exists to catch. | ||
| # | ||
| # Only the working tree is inspected: `implement` leaves its changes uncommitted | ||
| # (committing is `create-pr`'s job). If that ever stops being true this guard fails | ||
| # a run that did have work — visible and annoying — rather than passing one that | ||
| # did not, which is silent. That is the safe direction for it to be wrong in. | ||
| - id: assert-implemented | ||
| bash: | | ||
| base=$BASE_BRANCH | ||
| tracked=$(git diff --name-only HEAD -- ':(exclude).archon' | head -1) | ||
| untracked=$(git ls-files --others --exclude-standard -- ':(exclude).archon' | head -1) | ||
| if commits=$(git rev-list --count "origin/$base..HEAD" 2>/dev/null); then :; else commits=unknown; fi | ||
|
|
||
| if [ -n "$tracked" ] || [ -n "$untracked" ]; then | ||
| git --no-pager diff --stat HEAD -- ':(exclude).archon' | tail -1 | ||
| elif [ "$commits" = "unknown" ]; then | ||
| echo "WARNING: cannot compare against origin/$base — not failing on an unknown" | ||
| elif [ "$commits" != "0" ]; then | ||
| echo "$commits commit(s) ahead of origin/$base" | ||
| else | ||
| echo "implement produced neither a commit nor a working-tree change outside .archon/ — failing before the review phase" >&2 | ||
| exit 1 | ||
| fi | ||
| depends_on: [implement] |
There was a problem hiding this comment.
📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win
Update the stale rationale comment to match the guard's actual logic.
The comment block states: "Only the working tree is inspected: implement leaves its changes uncommitted (committing is create-pr's job). If that ever stops being true this guard fails a run that did have work — visible and annoying — rather than passing one that did not, which is silent."
The bash below it does not match this description: it also computes commits via git rev-list --count "origin/$base..HEAD" (Line 334) and treats a non-zero commit count as evidence of implementation (Lines 340-341). So the guard checks both the working tree and commits ahead of the base branch, not "only the working tree." A future maintainer reading only the comment could reasonably remove the commit-count branch, believing it is dead weight, and reintroduce the exact silent-pass failure this guard exists to prevent.
Update the comment to describe both checks, or explain why the commit-count branch exists despite the stated assumption that implement never commits.
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In @.archon/workflows/experimental/archon-fix-github-issue-experimental.yaml
around lines 309 - 346, Update the rationale comment above assert-implemented to
accurately state that the guard checks both uncommitted working-tree changes
outside .archon/ and commits ahead of origin/$BASE_BRANCH via the commits count.
Remove or revise the claim that only the working tree is inspected, and explain
the commit-count fallback’s purpose without changing the bash logic.
1 participant
Summary
review-gatefails safe toward more review; the PR number is captured deterministically instead of by the model.The three fixes
1. Planning ignored decisions recorded in comments (bundled commands — affects all users)
fetch-issuealready pullscomments, andarchon-investigate-issuealready listed them under "Extract Context". The gap was authority, not availability.On 2026-08-03, a run on #2359 implemented the issue body while a maintainer comment on that same issue specified a different design. The comment was posted at 07:54:13Z, the run started at 07:54:29Z, and the contract text (
startedAt,--events,replaces,first appearance) was confirmed present in the run'sfetch-issueoutput. The resulting PR (#2399) was closed.Both planning entry points —
archon-investigate-issue(bugs) andarchon-create-plan(everything else) — now state:#1234or URL) one level and read their comments tooThe incident is written into both files. A rule with its origin attached survives; an abstract instruction gets skimmed.
Note: the workflow cannot inject this per-node.
commandNodeSchemais{ command: string }with no input channel, so acommand:node renders the shared file and the only way to reach the planner is the prompt itself.2. A failed gate silently reduced review coverage (experimental workflow)
review-gateread$smoke-validate.output.claims_accuratevia strict field access. Whensmoke-validatetimed out and emitted nothing, the strict ref failed the gate, every conditional reviewer skipped viatrigger_rule, and the run completed successfully with one reviewer instead of the intended set. Observed on run7f392748.Now: reads the whole output and greps, runs under
trigger_rule: all_doneso a failed producer still reaches the branch, and suppresses extra reviewers only when scope is small AND the claims were positively confirmed. Missing output, timeout, or failure all bias toward more review.3. The PR number was recorded by the model, and was usually wrong (experimental workflow)
create-prasked the model to rungh pr view --json numberand echo the result into$ARTIFACTS_DIR/.pr-number. Being a prompt instruction, it is advisory. Measured across four parallel runs:.pr-numberwritten3b03cb58c77ed1bc500fa086bd0db0cdreview-scopethen resolved a nonexistent PR, correctly refused to fabricate ascope.md, andreview-classifydeclined every specialist for lack of evidence. Three of four PRs received one generalist reviewer while the runs reported success.Replaced with a deterministic
capture-pr-numberbash node that resolves the PR from the branch and exits non-zero if it cannot. Verified in the exact worktree that failed: the node returns 2388 where the model had written 2323.Per the workflow constitution this is the reliability carve-out, not a style preference — a check with no judgment content whose silent failure guts the review phase belongs in a node that cannot decline to fire.
synthesizealso movesone_success→all_done. Defensive rather than causal, butone_successpermits a synthesis to stand on a subset and report it as the whole review.Validation Evidence
The single remaining validator warning is pre-existing on
fetch-issueand unrelated to this change.End-to-end evidence for fix 3 — run
abd3370con #2344 with the deterministic node in place:.pr-numberscope.mdcode_reviewonly (3 of 4)Fix 1 verification is in flight — a re-run of #2359 with the corrected planning prompts was still executing when this PR was opened. If it produces the contract shape (
nodesreplacingevents, a--eventsflag,startedAt), the instruction works. If it reproduces the old shape, the prompt is insufficient and this PR should not merge on the strength of reasoning alone.Security Impact
capture-pr-numberusesgh pr view, whichcreate-pralready invoked at the same pointCompatibility / Migration
.archon/commands/archon-*.mdstill overrides the bundled default by name.Blast radius
The two command changes are bundled defaults, so they affect every workflow invoking those commands — the whole fix-github-issue family. They only add instruction; no behaviour is removed. The failure mode if the wording is wrong is a planner that over-weights a comment, which is far cheaper than the current failure mode of ignoring a decision entirely.
The workflow changes are confined to
.archon/workflows/experimental/.Rollback
git revert— prompts and one workflow file, no state, no migration. Operators can override either command locally without waiting for a release.Risks and Mitigations
capture-pr-numberfails the run where the old behaviour limped on.Summary by CodeRabbit
New Features
Bug Fixes