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

[prflow-implementer]

View run

Summary

• create-issue now writes every run artifact into a per-run sub-directory .prflow/tmp/create-issue/<slug>/ instead of as flat files directly under .prflow/tmp/, so one run's working set is no longer scattered across a directory shared with every other producer.
• A run that successfully creates its issue reaps its own run directory as a final step, keyed to the recorded slug (never a pattern or age sweep); a run that ends any other way leaves the directory in place as its diagnostic record.
• Pre-existing flat issue-* files under .prflow/tmp/ are left untouched.

Changes

create-issue skill prose: skills/create-issue/SKILL.md and its references/*.md (step-4-present-create, audit-prompt-template, step-2/3-5/3-6 references, fallback arms, issue-template, degradation-routing, revision-delta) now compose artifact paths under .prflow/tmp/create-issue/<slug>/, including the shared slug pointer at .prflow/tmp/create-issue/issue-run-slug.

Audit-state path: scripts/issue-audit-state.py relocates the audit-state file and the bound canonical-draft file to .prflow/tmp/create-issue/<slug>/, updating state_path, _bound_draft_file, and the coupled dispatch/binding docstrings and --path help.

Cleanup reaper: new scripts/cleanup-create-issue-run.sh removes a completed run's <root>/.prflow/tmp/create-issue/<slug>/ directory keyed to the recorded slug, refusing empty or path-unsafe slugs (delete nothing, exit 0), removing the shared slug pointer only when it still holds this run's slug, and iterating over one or more --root values.

Tests & coverage: the #1811 cleanup: block in lib/test/test_python_scripts.py drives the reaper via subprocess; lib/test/modules/issue-audit-state.sh and lib/test/test_python_scripts.py cover the audit-state relocation; lib/test/modules/create-issue-contract.sh covers the coupled path enumerations (AC5); a coverage-map.json row registers the reaper; lib/test/run.sh's draft-path pin and test_render_audit_prompt.py are updated to the per-run layout.

Docs: docs/internal/DEVFLOW_SYSTEM_OVERVIEW.md, cloud-allowlist.md, and create-issue-context.md are reconciled to the per-run scratch layout.

Resolves

Resolves #1811

Test Plan

• lib/test/test_python_scripts.py — the #1811 cleanup: cases pass (valid-slug reap, foreign-pointer skip, empty/unsafe-handle non-destructive, traversal + interior-slash refusal, empty --root, unexpected arg, newline-less pointer, multi-root, valueless-flag termination, absent-dir idempotent).
• lib/test/modules/issue-audit-state.sh passes with the state file at the per-run location.
• lib/test/modules/create-issue-contract.sh passes (AC5 coupled path enumerations).
• A create-issue run that creates its issue leaves no files of its own under .prflow/tmp/; a run that ends otherwise leaves its run directory in place.

Visual Changes

N/A

Breaking Changes

None

Prompt-surface size

Derived at 8b4e74ccc359180cd34001a366acbb6c8783cc52 against merge-base 66eec8a8a2aee14677cd27359929722760691fe3 (origin/main). Covered: tracked *.md under skills/, agents/, .prflow/prompt-extensions/.

Path	Before	After	Δ bytes	Δ %
.prflow/prompt-extensions/create-issue.md	25,111	25,131	+20	+0.1%
skills/create-issue/SKILL.md	29,262	29,461	+199	+0.7%
skills/create-issue/references/audit-prompt-template.md	27,490	28,010	+520	+1.9%
skills/create-issue/references/degradation-routing.md	11,166	11,186	+20	+0.2%
skills/create-issue/references/fallback-audit-dispatch-arms.md	5,607	5,807	+200	+3.6%
skills/create-issue/references/fallback-no-task-tool.md	3,756	3,789	+33	+0.9%
skills/create-issue/references/fallback-read-only-sandbox.md	7,673	7,786	+113	+1.5%
skills/create-issue/references/issue-template.md	36,289	36,369	+80	+0.2%
skills/create-issue/references/revision-delta.md	6,290	6,310	+20	+0.3%
skills/create-issue/references/step-2-clarify.md	29,949	30,082	+133	+0.4%
skills/create-issue/references/step-3-5-steelman.md	34,508	34,588	+80	+0.2%
skills/create-issue/references/step-3-6-audit-adjudication.md	27,926	28,006	+80	+0.3%
skills/create-issue/references/step-3-6-audit-dispatch.md	18,890	19,110	+220	+1.2%
skills/create-issue/references/step-3-6-audit-shared.md	18,532	18,632	+100	+0.5%
skills/create-issue/references/step-3-6-audit.md	6,106	6,126	+20	+0.3%
skills/create-issue/references/step-4-present-create.md	54,998	56,828	+1,830	+3.3%
Whole covered surface	2,117,268	2,120,936	+3,668	+0.2%

Generated via /prflow:implement (v2.34.14, claude-opus-4-8, low)

[prflow-implementer]

/prflow:review

[prflow-reviewer]

Verdict: APPROVE with notes (clean mechanical scratch-relocation; strong test coverage, all required CI checks green; only low-severity Suggestions)

Issue Compliance

Reviewed against issue #1811: "create-issue: write run artifacts into per-run sub-directories under .prflow/tmp/ and clean up at end of run" — criteria from the /prflow:implement workpad comment; scope unchanged. The change satisfies every stated requirement: run artifacts now compose under .prflow/tmp/create-issue/<slug>/ (verified in scripts/issue-audit-state.py state_path/_bound_draft_file and across the skill prose); a successful run reaps its own run dir keyed to the recorded slug via the new scripts/cleanup-create-issue-run.sh, invoked at a success-only site (sub-step 5d of references/step-4-present-create.md); a run ending any other way leaves the dir in place; and pre-existing flat issue-* files are structurally untouched (the reaper only removes create-issue/<slug>/ and the pointer when it still holds this run's slug).

Test evidence

CI ground truth for the reviewed head a9efc1d7ca13758470d9e78c8b891e89d8b82f7a (read from the GitHub API): lib + python tests = success, lint (shellcheck + actionlint + ruff) = success, and every shard (…) = success. The suite was not re-run in-review; these API conclusions are the authoritative test evidence.

Verification

This engine-self-modifying diff (touches skills/**, lib/**, .prflow/prompt-extensions/*.md) was reviewed with the full Phase 3 specialist roster plus manual verification of the two code files. has_new_types is false, so type-design-analyzer was not dispatched. All five review passes returned clean:

• code-reviewer — no Critical/Important. Slug regex ^[A-Za-z0-9][A-Za-z0-9._-]*$ byte-matches the state_path guard in Python; rm -rf is keyed to a validated single path segment (no glob/age sweep); arg-parse terminates on a trailing valueless flag; set -u + "${roots[@]:-}" is safe; no stale flat-path references remain.
• silent-failure-hunter — no genuine silent failures. The reaper's suppressions are documented best-effort that fail non-destructively; the relocated Python write paths all create the deeper create-issue/<slug>/ parent (makedirs(exist_ok=True) / mkdir(parents=True)) and surface real I/O errors loudly.
• comment-analyzer — no findings. All relocated docstrings (7 sites) accurate; every reaper comment names a concrete wrong change it prevents and stays within the 3-line cap.
• pr-test-analyzer — coverage adequate. All destructive-safety branches (unsafe/empty/traversal/interior-slash refusal, foreign-pointer skip, valueless-flag non-spin, newline-less pointer, multi-root, idempotent absent-dir) are directly and behaviorally covered.
• final-pass (requesting-code-review) — no Critical/Important; path references consistent, reaper at a success-only site, slug-keyed not sweep.

Intra-diff contradiction scan: no contradiction found. Truthfulness sweep: no finding promoted.

Repo prompt-extension gates

• Prompt-surface edit routing evidence gate — DISCHARGED. The diff touches trigger globs (skills/create-issue/SKILL.md, .prflow/prompt-extensions/create-issue.md). Issue #1811's workpad carries two Writing-skills evidence: markers, each with all four dispositions recorded: skill-loaded=yes guidance-applied=yes pressure-scenario=no micro-tests=no. All slots present, so the gate is discharged (a no never draws a finding on its own).
• Wording-only pin review policy — clean. The lib/test/run.sh change is an update of an existing draft-save-message pin (carries a valid # structural-pin-ok: cross-file-phase-contract -- <rationale> declaration), with its lib/test/pin-corpus-adjudications.tsv literal-hash row updated in lockstep. No new wording-only presence pin was added; the added tests are behavioral (subprocess drives the reaper, asserts filesystem effects).
• Shipped-prose cost audit — the ~3.7 KB net prose growth under skills/**/.prflow/prompt-extensions/ is dominated by the mechanical create-issue/<slug>/ path-segment insertion into existing sentences plus one new reaper-invocation step (5d) carrying a genuine instruction. No gratuitous additions warranting an Important finding.
• Verification-evidence advisory — the workpad carries Verification evidence: markers; advisory is silent.

Code Review Findings

🟡 Suggestion / Minor

1. scripts/cleanup-create-issue-run.sh — the pointer removal (rm -f -- "$ptr" && printf …) is asymmetric with the target-dir removal above it: a failed pointer rm -f emits no warning: could not remove breadcrumb (the && short-circuits the success print with no else), whereas the directory removal warns on failure. Within the best-effort/exit-0 contract this is not a defect; an else warning for parity would make a stuck pointer visible. (raised by 1/5 agents)
2. docs/internal/advisory-adjudication-calibration.md:80,85 — still names the pre-move flat globs .prflow/tmp/issue-audit-state-*.json / .prflow/tmp/issue-audit-*.md. This file is not touched by the PR and is a dated (2026-07-23) frozen inventory snapshot, so it is correctly left under the past-time-snapshot exemption; noted only so it is a conscious decision. (raised by 2/5 agents)
3. scripts/cleanup-create-issue-run.sh — the arg-parse and pointer-read comment blocks sit at the 3-line cap; they satisfy §2.3 comment discipline but could be tightened if the file is edited later. (raised by 1/5 agents)

Verdict Criteria

• No verification FAIL/INCONCLUSIVE.
• No self-contradicting-diff finding.
• No finding at or above the configured verdict threshold (critical).
• Only Suggestion-level findings present → APPROVE with notes.