create-issue: write run artifacts into per-run sub-directories under .prflow/tmp/ and clean up at end of run

[The01Geek]

Problem Statement

The create-issue command writes every run artifact as a flat file directly under .prflow/tmp/ and never removes any of them. One run's working set is scattered across a directory shared with every other producer (suite shards, telemetry staging, workpad buffers, verification flights), and the directory only grows — measured at 137 MB / 1,840 entries in one checkout, 1,323 of them create-issue issue-* files dating back six weeks.

Current Behavior

Artifacts are separated only by a <slug> embedded in each filename (issue-draft-<slug>.*.staged.md, issue-audit-<slug>.md, issue-audit-state-<slug>.json, issue-body-<slug>.md, issue-fetched-<slug>.md, …), all flat under .prflow/tmp/. No step of the run deletes anything, so completed runs leave their whole working set behind indefinitely. This happens regardless of environment.

Desired Behavior

Each create-issue run writes its artifacts into its own sub-directory .prflow/tmp/create-issue/<slug>/. A run that completes successfully (issue created and reported) removes its own run directory as its final step; a run that ends any other way leaves the directory in place as its diagnostic record. Cleanup targets only the run's own recorded directory handle, never a pattern or age sweep, so concurrent runs in sibling worktrees are untouched.

User Impact

Whoever runs create-issue in this repository stops accumulating unbounded scratch files, and a failed run's evidence remains findable in one directory instead of interleaved with a thousand strangers.

Technical Context

Scope note: The files and details below are the known starting points, not the full list. Before implementing, trace the change through the codebase to find every affected call site, consumer, and layer — this issue maps the work, it does not bound it.

• Relevant Classes/Files — skills/create-issue/SKILL.md and skills/create-issue/references/*.md compose the artifact paths; scripts/issue-audit-state.py owns the audit-state file path; skills/create-issue/references/audit-prompt-template.md enumerates exact out-of-bounds paths under .prflow/tmp/ that must move with the artifacts.
• Architecture Alignment — .prflow/tmp/ stays the scratch root; this adds one namespaced level under it. The scaffolded .prflow/.gitignore already ignores tmp/ wholesale, so sub-directories need no gitignore change.
• Dependencies — none beyond the guaranteed toolset.
• Data/Schema Considerations — the audit-state JSON and staged-draft history keep their formats; only their directory changes.
• Cross-layer Impact — create-issue skill prose, its references, and scripts/issue-audit-state.py. Other commands' .prflow/tmp/ usage is untouched.

The durable record of a successful run is the created issue plus its investigation-record comment, so deleting the run directory on success destroys no citable evidence; the retained-on-failure arm preserves the audit lifecycle and staged canonical-draft history exactly when they are still needed.

Acceptance Criteria

• A create-issue run writes every artifact it produces (drafts, staged history, audit files, audit state, emitted body, fetched copies, derivation artifact) under .prflow/tmp/create-issue/<slug>/ and none as a flat file directly under .prflow/tmp/.
• A run that completes successfully leaves no files of its own under .prflow/tmp/ after it reports the created issue URL.
• A run that ends without creating the issue leaves its run directory and contents in place.
• Cleanup removes only the directory the run itself recorded creating; artifacts belonging to any other slug or producer are untouched.
• Every path literal in skills/create-issue/references/audit-prompt-template.md's out-of-bounds enumerations names the new per-run location, and lib/test/modules/create-issue-contract.sh passes.
• Pre-existing flat issue-* files under .prflow/tmp/ are left in place by this change.

Implementation Notes

• Approach — introduce the per-run directory at the point the slug is first derived, thread it through every path composition, and add a final delete-on-success step keyed to that recorded handle. Backfill of the ~1,300 existing flat artifacts is out of scope; a human sweeps them once by hand.
• Relevant files — this likely touches skills/create-issue/SKILL.md, the references/*.md files that compose artifact paths (step-4-present-create.md, audit-prompt-template.md, issue-template.md's posting block), and scripts/issue-audit-state.py; plausibly lib/test/modules/create-issue-contract.sh fixtures.
• Code Patterns — mirror the parallel coordinator's run-root convention (.prflow/tmp/parallel-suite/run-<pid>-<n>), which already namespaces per-run scratch under .prflow/tmp/ and reaps by its own recorded handle.
• Testing Strategy — lib/test/modules/create-issue-contract.sh covers the coupled path enumerations; a focused case drives scripts/issue-audit-state.py with the new state-file location and asserts it reads and writes there. One residual-risk case: a cleanup invoked with an empty or unset directory handle deletes nothing and exits non-destructively.
• Documentation Needed — none. The artifact layout is internal scratch; the coupled prose that names paths is updated as part of the change itself.
• Potential Gotchas — the audit prompt template enumerates exact .prflow/tmp/ paths in two places and skills/create-issue/** plus scripts/issue-audit-state.py compose them independently: this is a coupled edit across all of them, not a one-line change. Editing skills/** bodies goes through the writing-skills discipline with a Writing-skills evidence: marker. The Write tool refuses main-root paths from a worktree, so path composition must keep resolving through resolve-main-root.sh as today.

[The01Geek]

/devflow:implement 1811

[github-actions]

PRFlow Workpad — Issue #1811

Status: 🎉 Complete
Branch: issue-1811-create-issue-write-run-artifacts-into-per-run-sub
Run: View run
PR: #1957
Last updated: 2026-08-25 20:01 UTC

Progress

• Setup — branch & workpad
  • prompt extension resolved: implement
  • 05:22:15 — /prflow:implement run started
  • 05:22:57 — Claude job setup complete; invoking agent
  • 05:25:14 — agent entered Phase 1 setup; workpad triage passed
  • 05:25:21 — agent initialized; Phase 1 workpad hydrated
  • 05:25:32 — resume-kind: fresh
  • 05:25:53 — Phase 1.3.5: early declared-dependency preflight PROCEED
  • 05:27:53 — issue-claim audit (projection): every Desired Behavior obligation is represented in the resolved acceptance criteria — pass complete
  • 05:27:55 — issue-claim audit (count): no count or enumeration claims found — pass complete
  • 05:28:00 — issue-claim audit (negative-scope): issue excluded other-commands tmp usage and documentation; trace confirms no impact - no new helper invocations added, so cloud-tier workflow allowlists are unaffected
  • 05:28:02 — issue-claim audit (policy): no policy-referencing AC claims found — pass complete
  • 05:28:06 — issue-claim audit (execution-capability): no workflow-resident acceptance criteria found — pass complete
  • 05:28:08 — issue-claim audit (verified-premise): no Verified: bullets and no ungraded claims found in the issue body — pass complete
  • 06:15:34 — run died: fail-exhausted — https://github.com/The01Geek/prflow/actions/runs/32816077331
  • 13:22:58 — trigger accepted; existing workpad adopted
  • 13:23:40 — Claude job setup complete; invoking agent
  • 13:26:38 — /prflow:implement new run initialized from terminal workpad; Phase 1 workpad hydrated
  • 13:26:45 — resume-kind: terminal-re-trigger
  • 13:26:45 — classification: non-bug — Feature/refactor: requests per-run scratch sub-directories and cleanup; no malfunction described, no bug label
  • 13:27:17 — Phase 1 (branch setup, dependency preflight, issue-claim audit) completed in prior attempt; branch current with base (UP_TO_DATE); resuming into Phase 2 verification
  • 13:36:41 — test-first (resumed): cleanup-create-issue-run.sh had no tests when adopted; added 15 behavioral subprocess cases in lib/test/test_python_scripts.py (valid-slug reap, foreign-pointer skip, empty-handle non-destructive, unsafe-slug traversal refusal, multi-root, idempotent absent-dir), all PASS; copy-based mutation check removed the slug guard and confirmed the traversal victim IS deleted, proving the refusal test is non-vacuous. Registered coverage-map row for the helper (owner test_python_scripts.py).
  • 13:36:41 — Phase 2.3 sweeps over branch delta (merge-base origin/main -> worktree). Shape: relocation of create-issue scratch path literal from flat .prflow/tmp/issue-* to .prflow/tmp/create-issue// across 15 skill+reference files and 2 python composition fns (state_path, _bound_draft_file); new best-effort reaper helper; test updates. 2.3.0 changed-contract/relocated-literal: grep of create-issue surface + issue-audit-state.py + tests for surviving flat .prflow/tmp/issue- citations found only the unrelated implement issue-body/ cache (different subsystem) — relocation complete; audit-prompt-template out-of-bounds enumerations (8 and 10 paths) all name the new per-run location (AC5). 2.3.0a/2.3.0c: helper guards (empty-handle non-destructive, anchored slug charset regex, per-root, pointer removed only when it holds this run slug) each tested; state_path slug-safety peer unchanged. 2.3.3 convention: shellcheck clean on the new helper. 2.3.4 boundary: resolve-main-root.sh confirmed present; no external-tool-output matcher added. 2.3.4a self-authored-claim: updated docstrings/comments in issue-audit-state.py match the shipped composed paths. 2.3.4b coverage-claim: stale-prose-lint recognized ~14 pre-existing CU/R3 coverage+count lines in the delta; every one is a path-prefix-only relocation with count and membership unchanged and the quantifier range enumerated inline or grounded by verified helper behavior — no newly-authored unclosed universal. 2.3.5/2.3.6: helper is a best-effort reaper with per-branch stderr breadcrumbs and fails safe (unreadable pointer -> empty -> not removed). Tests: python-scripts 5633/0, create-issue-contract 397/0, render-audit-prompt 105 OK.
• Implement
  • code + sweeps
  • 05:28:48 — entered Phase 2; §2.0 resume gate did not fire (resume-kind fresh)
  • 05:36:24 — Path B (complex): coupled prompt-surface + Python change. Design: insert one namespaced dir level 'create-issue//' after '.prflow/tmp/' at every artifact-path composition, keeping filenames identical (uniform, minimal-churn prefix insertion). Slug pointer moves to .prflow/tmp/create-issue/issue-run-slug. Cleanup-on-success centralized in a new tested helper scripts/cleanup-create-issue-run.sh keyed to the recorded slug + anchor roots.
  • 05:36:24 — 2.2.4 Reuse & Altitude gate: reuse search — grepped skills/create-issue and scripts/issue-audit-state.py for '.prflow/tmp' path-composition sites; no reusable per-run-dir helper exists (the parallel-suite run-root convention in lib/test/run-parallel.sh _try_run_root is shell-internal, not importable into skill prose or issue-audit-state.py). New composition is a one-level dir insert; cleanup is a new helper. Altitude: uniform namespaced level, not a special case — good depth.
  • 05:36:24 — 2.2.7 coupled-site map (searches run: Grep '.prflow/tmp/(issue-|steelman-|finding-|create-issue)' over skills/create-issue → ~13 reference/SKILL files; Grep over scripts/issue-audit-state.py → state_path L1026 + _bound_draft_file L2782 only; Grep over lib/test/modules/create-issue-contract.sh → NO .prflow/tmp literals; Grep over test_python_scripts.py → ~40 sites; test_render_audit_prompt.py → L34 + SCOPE_GLOB/DISPATCH_FILE/RECORD_FILE L1728-1730). Coupled: skill prose ↔ _bound_draft_file (record-dispatch --write-path mismatch guard requires identical composed draft path).
  • 06:04:02 — Writing-skills evidence: skills/create-issue/SKILL.md mode=subagent skill-loaded=yes (writing-skills present in available-skills list, loaded via Skill tool, RED/GREEN discipline followed) guidance-applied=yes (Reference-type mechanical path relocation; create-issue-contract module passed 397/0) pressure-scenario=no (factual path relocation adds and relaxes no rule, no discipline behavior to pressure-test) micro-tests=no (no behavior-shaping wording changed, a no-guidance control would exhibit no failure)
  • 06:12:56 — trigger accepted; existing workpad adopted
  • 06:13:35 — Claude job setup complete; invoking agent
  • 06:14:27 — trigger accepted; existing workpad adopted
  • 06:15:03 — Claude job setup complete; invoking agent
• Review
  • prompt extension resolved: review engine
  • prompt extension resolved: fix loop
  • prompt extension resolved: code-review reception
  • Classify diff (Phase 0.5)
  • Generate verification checklist (Phase 1)
  • Verify checklist (Phase 2)
  • Review agents (Phase 3)
  • Aggregate & verdict (Phase 4)
  • Run complete — everything this run owed
  • /simplify
  • review-and-fix
  • acceptance-criteria gate
  • 13:39:07 — versioning: added .changeset/issue-1811-create-issue-per-run-scratch.md (bump: patch, type: Changed) after draft PR create-issue: write run artifacts into per-run sub-directories under .prflow/tmp/ and clean up at end of run #1957, before review pass; patch chosen (issue did not authorize minor/major). Engine surface touched (skills/create-issue, scripts/issue-audit-state.py, new scripts/cleanup-create-issue-run.sh) so a changeset is owed.
  • 13:39:07 — Phase 2.3.4b changeset coverage-claim: stale-prose-lint on the committed changeset branch delta produced 0 rows at exit 0 with hunks present (clean pass); the every-run-artifact universal is grounded by the executed 2.3.0 relocation enumeration and scoped to this change.
  • 13:43:03 — /simplify (4 cleanup agents): APPLIED 1 finding — altitude agent caught a real dropped-site regression from the relocation: step-2-clarify.md wrote to .prflow/tmp/create-issue// but its mkdir only made .prflow/tmp (the derivation write would fail on a clean fs and mis-route to the read-only-sandbox arm); deepened the mkdir to match SKILL.md:153. Confirms AC1 completeness. Module still 397/0. SKIPPED (recorded rationale): reuse/altitude _run_dir helper extraction across issue-audit-state.py state_path/_bound_draft_file (marginal — 2 sites, one module, same-commit coupling, low value vs churn on a green core file); reaper-as-python-subcommand and the reaper's slug-regex 'duplication' (deliberate standalone dependency-light testable bash design per plan; bash cannot cheaply reuse the python validator). Efficiency+simplification agents: clean.
  • 13:43:23 — Phase 3.3: pre-loop iter snapshot empty (fresh review, no prior iter-*.json in this ephemeral checkout); invoking review-and-fix in PR mode with PR create-issue: write run artifacts into per-run sub-directories under .prflow/tmp/ and clean up at end of run #1957
  • 13:43:23 — mid-phase resume point recorded
  • 13:57:49 — Writing-skills evidence: trigger=.prflow/prompt-extensions/create-issue.md mode=subagent skill-loaded=yes (superpowers:writing-skills present in available-skills and invoked via Skill tool before editing) guidance-applied=yes (Match the Form to the Failure — pure reference-fact path correction, no rule reworded) pressure-scenario=no (a corrected path literal shapes no agent behavior, no discipline failure to pressure-test) micro-tests=no (no behavior-shaping wording authored, no no-guidance control applies)
  • 15:24:48 — review coverage recorded (coverage=full, dispatch=attempted, roster=complete, checklist=complete)
  • 15:25:07 — review engine + fix-loop + code-review-reception extensions all resolved content-present via load-prompt-extension.sh at loop entry and re-loaded each iteration; the review engine ran inline (its ladder runs inside the loop's own turn sequence, reached by file read) so there is no separate Skill-tool return to observe.
  • 15:32:27 — Phase 3.4: all 6 ACs verified (two-verifier reconciled satisfied, all_satisfied=true); evidence and dispositions in this run's AC dispositions note
• Documentation
  • 15:35:22 — Phase 4.1: no documentation changes to commit (docs subagent ran clean / made no changes — the three coupled internal docs were already reconciled on-branch; external docs and release notes correctly no-op for internal scratch hygiene; docs extension present-empty)
  • 15:39:05 — 4.2 claim audit (behavioral): checked and confirmed true — traced reaper + issue-audit-state.py claims into the shipped code; body rewritten from the diff, no false behavioral claim survived
  • 15:39:05 — 4.2 claim audit (verification): every verification row bound to a named test in the diff (create-issue: write run artifacts into per-run sub-directories under .prflow/tmp/ and clean up at end of run #1811 cleanup block in test_python_scripts.py; issue-audit-state.sh module; create-issue-contract.sh for AC5; coverage-map.json) — no over-broad row
  • 15:39:05 — 4.2 claim audit (artifact-existence): no follow-up-issue claim made (nothing deferred); only linked artifact is the existing changeset — pass complete
  • 15:39:05 — 4.2: PR body updated from its placeholder (REST PATCH); pr-description extension content-present (inserted prompt-surface-growth size table); no code-level fix made
  • 15:39:57 — Phase 4.3 checkpoint 4: merged origin/main and pushed (was behind by 33; new HEAD 8b64184; ort auto-merge, no conflicts). Re-verifying merged tree and re-running the 2.3.0 create-issue relocation sweep against newly-arrived sites before the completion flight.
  • 16:03:54 — run died: fail-exhausted — https://github.com/The01Geek/prflow/actions/runs/32852950004
  • 18:37:00 — trigger accepted; existing workpad adopted
  • 18:37:42 — Claude job setup complete; invoking agent
  • 18:41:08 — resume-kind: terminal-re-trigger
  • 18:41:08 — classification: non-bug — Feature/refactor: per-run scratch sub-directories + cleanup; no malfunction described, no bug label
  • 18:41:08 — Backstop resume (run 32884751790): prior attempt died fail-exhausted mid-Phase-4.3 completion flight after merging origin/main (HEAD 8b64184) and a follow-up pin-corpus fix (31e78b8), both pushed. Phases 1-3 complete, all 6 ACs verified/ticked, PR create-issue: write run artifacts into per-run sub-directories under .prflow/tmp/ and clean up at end of run #1957 up, docs+PR-body done. Resuming to finish Phase 4.3: base-update checkpoint, re-verify merged tree, completion whole-suite gate, publish, finalize.
  • 19:22:23 — Phase 4.3 checkpoint 4: base was 141 behind; merged origin/main. One conflict in lib/test/test_python_scripts.py (both sides appended independent end-of-file test blocks: create-issue: write run artifacts into per-run sub-directories under .prflow/tmp/ and clean up at end of run #1811 cleanup and Require a stated disposition per chartered step across the implement-phase agents #1740 validator) resolved by keeping both; focused python-scripts run 5744/0. Merge concluded, pushed (HEAD d26700b).
  • 19:22:23 — Phase 4.3: whole-suite run found 1 failure — Prevent wording-only pin tests from regrowing after the corpus cleanup #810 mutation-routing gate on the relocated create-issue draft-path pin. Fixed in two parts: (1) authorized the pin adjudication delta with a new manifest .prflow/logs/pin-corpus-adjudication-changes/2026-08-25-issue-1811/adjudication-delta.tsv (84e2059b removed, a5220bed added); (2) added the required structural-pin-ok cross-file-phase-contract declaration at the pin site so Step 2 of the routing ladder (declared AND ledger-boundary) passes. monolith shard now 9839/0.
  • 19:43:35 — checkpoint 4: observed token UPDATED (merged 23 base commits, clean ort, HEAD a9efc1d) — clean, proceeding to publish
  • 20:00:46 — completion verification recorded (flight 9585528a554c…, validated)
• PR marked ready
  • 20:01:14 — /prflow:implement run finished, PR published (gh pr ready): create-issue: write run artifacts into per-run sub-directories under .prflow/tmp/ and clean up at end of run #1957

Plan

• Add scripts/cleanup-create-issue-run.sh — takes a run slug and one or more anchor roots; validates the slug is a safe path segment ([A-Za-z0-9][A-Za-z0-9._-]*); for each root removes <root>/.prflow/tmp/create-issue/<slug>/ and removes <root>/.prflow/tmp/create-issue/issue-run-slug only when it still holds this run's slug; empty/unsafe slug deletes nothing and exits non-destructively; best-effort stderr breadcrumbs. RED-first unit test in test_python_scripts.py.
• Change scripts/issue-audit-state.py::state_path() to compose <root>/.prflow/tmp/create-issue/<slug>/issue-audit-state-<slug>.json (insert create-issue/<slug>/); update docstring/comment.
• Change scripts/issue-audit-state.py::_bound_draft_file() to compose <root>/.prflow/tmp/create-issue/<slug>/issue-draft-<slug>.md; update docstring; update the accuracy of comments at L5530/L7517/L9324 that quote the draft path.
• Update skills/create-issue/references/audit-prompt-template.md — the three out-of-bounds enumerations (di, file/inline, embed arms) to name .prflow/tmp/create-issue/<slug>/<file> for every listed artifact.
• Update skills/create-issue/SKILL.md — Step 1 slug pointer moves to .prflow/tmp/create-issue/issue-run-slug; mkdir -p .prflow/tmp/create-issue; every path reference (issue-step1, issue-derivation, issue-audit, pointer) and the Step 4 ls -lL listing rewritten to the per-run location.
• Update skills/create-issue/references/step-4-present-create.md — draft, steelman-projection, issue-record, issue-fetched paths to the per-run dir; add the final delete-on-success cleanup step (after 5c, before/at 6) invoking cleanup-create-issue-run.sh with the recorded slug and both anchor roots; guard on empty slug.
• Update the remaining create-issue reference files that compose/cite .prflow/tmp/issue-* paths: step-3-6-audit-shared.md, step-3-6-audit-dispatch.md (incl. its out-of-bounds mirror), step-3-6-audit-adjudication.md, step-3-5-steelman.md, step-2-clarify.md, issue-template.md (emit-body → issue-body), step-3-6-audit.md, degradation-routing.md, revision-delta.md, fallback-read-only-sandbox.md, fallback-no-task-tool.md, fallback-audit-dispatch-arms.md.
• Update lib/test/test_python_scripts.py — every site that composes .prflow/tmp/issue-audit-state-<slug>.json / issue-draft-<slug>.md / staged paths to the per-run create-issue/<slug>/ layout, plus the state_path escape tests; add the cleanup helper's RED-first cases (incl. the empty-handle non-destructive case).
• Update lib/test/test_render_audit_prompt.py — the out-of-bounds path assertions (FILE_ARM_OOB prefix, SCOPE_GLOB/DISPATCH_FILE/RECORD_FILE) to the per-run location.
• Verify lib/test/modules/create-issue-contract.sh still passes (no .prflow/tmp literals pinned there per grep) and add a changeset.
• Add .changeset/issue-1811-create-issue-per-run-scratch.md (patch) after the draft PR exists.
• Run the create-issue-contract module, the render-audit-prompt test, and the python-scripts suite; iterate to green; final whole-suite CI reading.

Acceptance Criteria

• A create-issue run writes every artifact it produces (drafts, staged history, audit files, audit state, emitted body, fetched copies, derivation artifact) under .prflow/tmp/create-issue/<slug>/ and none as a flat file directly under .prflow/tmp/.
• A run that completes successfully leaves no files of its own under .prflow/tmp/ after it reports the created issue URL.
• A run that ends without creating the issue leaves its run directory and contents in place.
• Cleanup removes only the directory the run itself recorded creating; artifacts belonging to any other slug or producer are untouched.
• Every path literal in skills/create-issue/references/audit-prompt-template.md's out-of-bounds enumerations names the new per-run location, and lib/test/modules/create-issue-contract.sh passes.
• Pre-existing flat issue-* files under .prflow/tmp/ are left in place by this change.

Devflow Reflection

Devflow Reflection (click to expand)

💡 Improvements

• 💡 Dogfood (mutation-routing + base-merge coupling): the prior attempt's fix for the relocated create-issue draft-path prose pin moved only the adjudication key in lib/test/pin-corpus-adjudications.tsv, which is necessary but not sufficient. A relocated prose-presence pin is a NEW static-pin candidate (merge_base..HEAD), so the Prevent wording-only pin tests from regrowing after the corpus cleanup #810 mutation-routing gate requires BOTH an authorized adjudication-delta manifest under .prflow/logs/pin-corpus-adjudication-changes// AND a site # structural-pin-ok: declaration (Step 2 of the issue-948 routing ladder = declared AND ledger-boundary). Because that gate reads from HEAD blobs and only runs in the monolith shard / full suite, the incomplete fix passed no focused check and only surfaced at the completion-evidence whole-suite pass — after which the prior attempt died fail-exhausted. Improvement: the relocation-sweep guidance (2.2.7 coupled-site map) that already flags run.sh pins should also flag that a relocated prose-presence pin needs the full mutation-routing authorization pair, not just an adjudication-key move.

ℹ️ Notes

• ℹ️ skill-directory anchor resolved from the runner-reported base directory
• ℹ️ resume-precheck: both open-PR queries ran and returned none for this issue; workpad Branch line absent; no prior attempt to adopt
• ℹ️ fix-delta gate (iteration 1): operand-contract and adversarial-input-shape checks CLEAN on the reaper arg-loop fix; added-assertion attribution flagged one vacuous test assertion (the reaper never writes a pointer, so asserting its non-existence under root B could not flip) — refixed to a namespace-untouched assertion. Outcome: refixed.
• ℹ️ fix-delta gate (iteration 2): CLEAN on all three checks — the reaper newline-less-pointer fix is fail-safe on every pointer shape, the ~33-site issue-audit-state.sh relocation is internally consistent (every manual pre-command write gained a matching mkdir; no corruption-sim passes vacuously against a mismatched path), and the new newline-less-pointer assertion flips correctly. Outcome: clean.
• ℹ️ Coupled-site miss caught by the review loop (dogfood): the Phase 2 §2.3.0 relocation sweep grepped skills/create-issue, issue-audit-state.py, test_python_scripts.py and test_render_audit_prompt.py but NOT lib/test/modules/issue-audit-state.sh (a registered module) nor lib/test/run.sh's monolith-shard draft-path pin — both compose flat create-issue paths independently and both went RED. The create-issue-contract module (397/0) covered neither. Improvement signal: the 2.2.7 coupled-site map should enumerate registered test modules and run.sh pins that compose the relocated literal, not only the skill/script surface.
• ℹ️ Step 2.6 shadow (convergence, iteration 3): FULL coverage (5 reviewers: code-reviewer, silent-failure-hunter, comment-analyzer, pr-test-analyzer, requesting-code-review; type-design-analyzer gated out — no new types). shadow_verdict REJECT on ONE new Important finding: coupled-mirror stale docs — docs/internal/DEVFLOW_SYSTEM_OVERVIEW.md (sec 11) and create-issue-context.md described the old flat .prflow/tmp/issue-* paths as current. Per CLAUDE.md's coupled-invariant discipline and the issue's own statement that coupled path prose is updated in-change, this promoted to iteration 4 and was FIXED (13 path literals reconciled to create-issue//). Over-grade note: an internal-doc path staleness borders the behavior-inert-prose cap (Suggestion), but the coupled-invariant discipline + the issue mandate make same-change reconciliation required, so it was fixed rather than parked. A tree-wide re-sweep then confirmed no flat create-issue path literal remains outside the intentional run.sh absence-pins, historical snapshots, and the unrelated issue-body cache.
• ℹ️ Step 2.6 shadow (convergence, iteration 4): FULL coverage (5 reviewers), shadow engine verdict APPROVE, but surfaced one more new Important (non-blocking) coupled-mirror: docs/internal/cloud-allowlist.md documented create-issue's emit-body redirect at the old flat /.prflow/tmp/issue-body-.md while issue-template.md now composes the nested create-issue// form. Promoted to iteration 5 and FIXED. An exhaustive tree-wide grep (all create-issue artifact stems, excluding the retrospective loop's own issue-body scratch, the implement issue-body/ cache, historical snapshots, and the correct run.sh absence pins) then confirmed NO flat create-issue path literal remains anywhere. Dogfood improvement: three separate convergence-shadow rounds each found a distinct coupled-mirror site the Phase 2 §2.3.0 sweep missed (issue-audit-state.sh module, run.sh pin, then docs). The relocation-sweep guidance should enumerate ALL sites that compose OR document a relocated literal — registered test modules, run.sh pins, and docs/internal — not just the skill/script producer surface.
• ℹ️ park-calibration gate: 1 parking preserved on evidence equivalence. The iteration-2 advisory-parked untested-reaper-branch gap was re-raised by the iteration-5 shadow at Suggestion (at/below its parked severity, same evidentiary basis) — equivalent, so parking preserved, not a mis-grade. No other parked finding matched a shadow re-raise.
• ℹ️ over-grade calibration gate: 1 finding calibrated. The iteration-5 shadow's code-reviewer graded skills/create-issue/references/step-3-6-audit-shared.md's stale bootstrap-mkdir mirror Important; it was calibrated to Minor under the behavior-inert-prose cap (the staged-write helper mkdirs parents, so the draft still lands; sole observable impact is the prose/claim itself). It was fixed anyway per the coupled-invariant discipline, so no promotion was needed. Loop verdict: CLEAN-FULL APPROVE (converged after 5 iterations; final full-coverage shadow APPROVE, post-shadow edit delta-review clean).
• ℹ️ Phase 3.4 acceptance-criteria gate — two-verifier reconciled record (reconcile-ac-verifiers.py: all_satisfied=true, blocking=[]). All six criteria reconciled satisfied, every disposition slot discharged. - AC1 (all artifacts under create-issue//, none flat) — satisfied (both). Evidence: issue-audit-state.py state_path/_bound_draft_file + every skill write-site compose create-issue//; corroborated by test_python_scripts.py create-issue: write run artifacts into per-run sub-directories under .prflow/tmp/ and clean up at end of run #1811 block (5644/0). Slots: evidence type-decided/command-run/single-flight/evidence-recorded all discharged; claim claim-traced/command-source-read/evidence-recorded all discharged. undischarged: none. - AC2 (successful run leaves no files under .prflow/tmp/) — satisfied (both). Evidence: reaper removes run dir + matching pointer on valid slug (test PASS), invoked at step-4 sub-step 5d after URL reported. Slots all discharged; undischarged: none. - AC3 (run ending without creation leaves its dir) — satisfied (both). Evidence: reaper's sole call-site gated behind successful gh issue create (sub-steps 5-5c); absent-dir/empty-handle tests confirm nothing deleted otherwise. Slots all discharged; undischarged: none. - AC4 (cleanup removes only recorded dir, never a pattern/age sweep) — satisfied (both). Evidence: deletion keyed to $base/$slug + pointer only when it holds this slug; foreign-slug/traversal/interior-slash/multi-root tests pass; no glob/mtime/listing sweep. Slots all discharged; undischarged: none. - AC5 (audit-prompt-template OOB enumerations name new location; create-issue-contract.sh passes) — satisfied (both). Evidence: audit-prompt-template.md:55,116,120 name only create-issue// paths; lib/test/run-module.sh create-issue-contract = 397/0. Slots all discharged; undischarged: none. - AC6 (pre-existing flat issue-* files left in place) — satisfied (both). Evidence: reaper's deletion surface scoped to create-issue/ + create-issue/issue-run-slug pointer; never touches .prflow/tmp/ directly, so a flat issue-* file is structurally unreachable (cleanup-create-issue-run.sh:41-63). Slots all discharged; undischarged: none. evidence_status_reported = satisfied and claim_status_reported = satisfied on all six; evidence_source = both.
• ℹ️ Verification evidence: whole-suite lib/test/run-parallel.sh on HEAD a9efc1d — aggregate CLEAN, 21277 passed / 0 failed / 0 skipped, all 5 required shards covered; flight 9585528a passed and satisfies_verification=true; retained logs under .prflow/tmp/parallel-suite/run-1022321-0/logs; elapsed 988s.

[prflow-implementer]

DevFlow stall backstop — this cloud run ended while the workpad Status was still in-progress (interim 🚀 Planning). Auto-resume attempt 1 of 2:

Resume note: invoke bundled helpers as .prflow/vendor/prflow/scripts/… (and .prflow/vendor/prflow/lib/…) with that path as the leading token — never an absolute path, never repo-root scripts/…, never behind a VAR= prefix or bash <path> wrapper (those forms are silently denied).

Headless note: this is a headless run — ending the turn ends the process, with no re-invocation. Never end the turn while any dispatched agent has not returned; poll to keep the turn alive; treat ScheduleWakeup and any future task-notification as unavailable.

/prflow:implement 1811

[prflow-implementer]

DevFlow stall backstop — this cloud run ended while the workpad Status was still in-progress (interim 🚀 Planning). Auto-resume attempt 2 of 2:

Resume note: invoke bundled helpers as .prflow/vendor/prflow/scripts/… (and .prflow/vendor/prflow/lib/…) with that path as the leading token — never an absolute path, never repo-root scripts/…, never behind a VAR= prefix or bash <path> wrapper (those forms are silently denied).

Headless note: this is a headless run — ending the turn ends the process, with no re-invocation. Never end the turn while any dispatched agent has not returned; poll to keep the turn alive; treat ScheduleWakeup and any future task-notification as unavailable.

/prflow:implement 1811

[prflow-implementer]

DevFlow stall backstop — this cloud run ended while the workpad Status was still in-progress (interim 🚀 Planning), and the auto-resume cap (max_resume_attempts=2, attempts already made: 2) is exhausted. The job has been failed so the stall is visible; to retry, post a fresh implement trigger comment naming this issue number.

[The01Geek]

/prflow:implement 1811