retrospective-weekly

name	retrospective-weekly
description	Run the weekly devflow self-improvement loop locally: scan freshly-merged watched-author PRs, write per-PR retrospective entries (LLM only for PRs that fail the mechanical clean-gate), derive recurring patterns, and file one human-reviewed GitHub issue per actionable pattern. Use when running the weekly devflow retrospective + audit.

/devflow:retrospective-weekly — Weekly Orchestrator

This skill is the single entry point the maintainer invokes once a week (or on demand). It is a conductor: it runs deterministic bash/jq scripts from lib/ at every mechanical step and dispatches LLM subagents only at the two genuine-judgment points — per-PR retrospective analysis (Stage A) and per-pattern issue-spec drafting (Stage B). Everything else — fetching, signal computation, gating, pattern math, and git/issue mechanics — is done by plain scripts with no LLM tokens. The loop proposes, it does not dispose: each actionable pattern is filed as one GitHub issue for the normal implement → review pipeline, not landed as an autonomous PR.

$LIB notation (textual, not a shell variable). Throughout this skill, $LIB in a command denotes the resolved path "${CLAUDE_SKILL_DIR:-<absolute skill base directory this runner reports in context>}"/../../lib — expand it textually (with the anchor already resolved for this runner) when composing each command you actually run. Never rely on a shell variable named LIB persisting from one statement or block to another — each Bash call is a fresh shell, and the Portable helper anchor note below explains why even same-command variable reuse is unsafe on some runners.

Every jq in this skill is invoked through the execution-verified wrapper $LIB/../scripts/run-jq.sh ($LIB/../scripts is the scripts/ dir beside lib/), never bare jq — so a shim-shadowed Windows/WSL host resolves a runnable jq the same way the .sh helper tier does (issue #253, the agent-tier sibling of #247). DEVFLOW_JQ is not exported to agent shells, so the wrapper must be invoked by path.

All scratch files live under .devflow/tmp/ (gitignored). Learnings files (.devflow/learnings/) are tracked and committed via the state PR.

GitHub autolink hygiene (any text you compose that lands on a GitHub surface — issue/PR titles, the state-PR report comment, body content you assemble): never put a bare # immediately before a number unless it is a real issue or PR reference — GitHub renders #2 as a link to issue/PR 2, which misleads readers. For an ordinal, count, or list position, spell it out ("item 2", "step 3"), never #2. Genuine references like #123 stay as-is.

---

Portable helper anchor (single-statement). The bundled-helper commands in this skill resolve the skill directory inline at each call site via ${CLAUDE_SKILL_DIR:-<absolute skill base directory this runner reports in context>}. When $CLAUDE_SKILL_DIR is set and non-empty (Claude Code), run each command exactly as written. On a runner where it is unset or empty, replace the placeholder with the skill base directory the runner reports in context (e.g. a Base directory for this skill: line) before running the command; if that reported path is Windows-form (C:\...), first convert it to this shell's POSIX form with one standalone wslpath -u '<path>' (WSL) or cygpath -u '<path>' (Git Bash/MSYS2) command and substitute the printed result only if the command succeeds and prints a non-empty path — otherwise fall through to the drive-letter rules exactly as if the tool were absent, the same success-and-non-empty acceptance the platform's path-normalization rules apply (if neither tool exists: lowercase the drive letter, map C:\ to /mnt/c on WSL or /c on MSYS2, and turn backslashes into /; if the environment is neither WSL nor MSYS2, use the path unchanged and report that it could not be normalized — the same arm the platform's path-normalization rules take). Resolve the anchor inline at every call site — never capture it into a shell variable that a later statement reads, because some runners' inline-bash marshaling drops such variables (observed on Copilot CLI). If neither $CLAUDE_SKILL_DIR nor a runner-reported base directory is available, stop and report that the helper anchor could not be resolved rather than running a command with a broken path.

Consumer prompt extension (load first). Before doing this skill's work, load any consumer-supplied prompt extension for this skill and honor it. From the repo root, run:

"${CLAUDE_SKILL_DIR:-<absolute skill base directory this runner reports in context>}"/../../scripts/load-prompt-extension.sh retrospective-weekly

If the invocation fails because the helper path does not exist (No such file, exit 127, or the platform equivalent), that is the anchor-resolution failure described in the Portable helper anchor note above — fix the anchor, don't report a missing extension. Otherwise, if the helper exits non-zero, a consumer extension exists but could not be loaded — surface its stderr message and do not silently proceed as if none existed. If it exits 0 and prints text, treat that text as additional instructions appended to the end of this skill's own prompt for this run — it is upgrade-safe, consumer-owned customization committed under .devflow/prompt-extensions/. If it exits 0 and prints nothing, proceed unchanged.

Procedure

Step 1 — Preflight

Confirm the working tree is clean:

git status --porcelain

If the output is non-empty, stop and tell the user to stash or commit their changes before running the loop.

Confirm gh is authenticated:

gh auth status

If it fails, tell the user to run gh auth login and stop.

Confirm you are on main:

git branch --show-current

If not on main, run git checkout main.

Prepare the scratch directory ($LIB below is the textual notation from the top of this skill — expand it when composing commands, do not assign a shell variable). This also removes prior-run per-PR scratch (result-*.json, pr-*.context.json) so a run starts clean and can never read another run's stale bundle or output:

mkdir -p .devflow/tmp
rm -f .devflow/tmp/new-entries.jsonl
# Remove prior-run per-PR scratch. `find … -delete` (not a bare `rm -f <glob>`)
# is shell- and OS-agnostic: it is a safe no-op when neither or only one pattern
# matches, whereas under zsh an unmatched glob is a fatal `no matches found` that
# would abort the whole rm (leaving BOTH patterns uncleaned). Step 1 runs before
# any fetch, so it never touches the current run's own freshly-written files.
find .devflow/tmp -maxdepth 1 -type f \( -name 'result-*.json' -o -name 'pr-*.context.json' \) -delete 2>/dev/null

---

Step 2 — Scan

Fetch the list of unprocessed watched-author PRs merged in the last 7 days:

bash $LIB/scan.sh > .devflow/tmp/scan.json

Ad-hoc / backfill / test runs. To run the loop against a specific set of PRs instead of the rolling 7-day window — e.g. backfilling old PRs, re-running after a fix, or testing the pipeline — pass --prs:

bash $LIB/scan.sh --prs 774,786,772,789 > .devflow/tmp/scan.json

--prs skips the GitHub search and the already-processed filter (you named the PRs, so the loop trusts you), but still drops any number that isn't a merged retrospected branch. Everything downstream (Steps 3–10) is identical. Do not use --prs for the scheduled weekly run.

scan.sh writes to stdout and exits non-zero on unrecoverable errors. If the output array is empty:

$LIB/../scripts/run-jq.sh 'length == 0' .devflow/tmp/scan.json

→ true: report "Nothing to process — no unprocessed watched-author PRs in the last 7 days." and STOP.

---

Step 3 — Per-PR context fetch + cheap gate

Initialize counters:

prs_scanned=0
clean_count=0
analyzed_count=0
skipped_count=0     # issue #626: mechanically- and Stage-A-skipped PRs
skip_records=()     # one-line report records, one per skip (never silent)
needs_analysis=()   # array of bundle paths

For each PR number in scan.json (iterate via $LIB/../scripts/run-jq.sh -r '.[].number'):

number=<the pr number>
CTX=$(bash $LIB/fetch-pr-context.sh "$number")
prs_scanned=$((prs_scanned + 1))

fetch-pr-context.sh writes the bundle to .devflow/tmp/pr-<n>.context.json and echoes that file path to stdout — so $CTX is the path, not the bundle content.

Run the cheap gate against the bundle content:

GATE=$($LIB/../scripts/run-jq.sh -c -f $LIB/cheap-gate.jq < "$CTX")

Outputs {"clean": <bool>, "reason": "<string>"}.

If clean == true:

Emit a clean entry (every retrospected PR is an implementation PR now — the old audit-kind path is retired along with autonomous intervention PRs):

$LIB/../scripts/run-jq.sh -c -f $LIB/clean-entry.jq < "$CTX" >> .devflow/tmp/new-entries.jsonl

Increment clean_count.

If clean == false:

First run the mechanical pre-dispatch disposition (issue #626). This decides — with no LLM dispatch — whether the non-clean bundle warrants Stage A analysis or is a mechanical skip (a foreign, non-DevFlow PR whose only non-clean signal is a missing workpad audit trail — Absent means the linked issue did resolve but carried no workpad comment, so this is not restricted to issueless PRs). It is a suite-driven helper, never inline prose:

DISP=$($LIB/../scripts/run-jq.sh -c --argjson gate "$GATE" -f $LIB/dispatch-disposition.jq < "$CTX")

DISP is {"disposition": "skip"|"dispatch", "reason": "<string>"}. It returns skip exactly when the gate reason is a workpad reason, the status is a sentinel (Absent/NoIssue), and pr_devflow_provenance is false — otherwise dispatch. So a bundle non-clean on any non-workpad signal (outstanding REJECT, CI failures, post-bot commits, review comments) is always dispatched, exactly the analysis it receives today.

If disposition == "skip" (the mechanical no-provenance skip): this is a permanently-terminal skip. Append a marker entry to the store and write a one-line run-report record — costing zero LLM dispatches. Do not add it to needs_analysis.

# $number is this PR (the loop variable); DISP's .reason is the skip reason line.
SKIP_REASON=$(printf '%s' "$DISP" | $LIB/../scripts/run-jq.sh -r '.reason')
$LIB/../scripts/run-jq.sh -cn --argjson pr "$number" --arg reason "$SKIP_REASON" \
  '{kind:"skip", pr:$pr, reason:$reason}' >> .devflow/tmp/new-entries.jsonl
skip_records+=("PR #$number skipped (mechanical, no DevFlow provenance): $SKIP_REASON")
skipped_count=$((skipped_count + 1))

Record the skip in the run report (see Step 9) with its reason line and increment skipped_count. The marker entry makes the processed-PRs filter treat this PR as handled on subsequent runs.

If disposition == "dispatch": add the bundle path to the analysis list:

needs_analysis+=("$CTX")
analyzed_count=$((analyzed_count + 1))

---

Step 4 — Stage A: Retrospective subagents (per non-clean PR)

For each bundle path in needs_analysis, dispatch a subagent. Issue up to 3–4 subagents concurrently in a single message (use the Agent tool for each). Each subagent prompt:

Read and follow "${CLAUDE_SKILL_DIR:-<absolute skill base directory this runner reports in context>}"/../retrospective/SKILL.md exactly.

Your context bundle path is: <path>

Print exactly one JSON object (the retrospective entry) and nothing else on stdout.

(The subagent picks categories from the fixed vocabulary in that skill — no "existing tags" list is passed; the vocabulary is the bounded list.)

Wait for all dispatched subagents to finish before continuing.

Collecting results: Each subagent's final message is its JSON object. Subagent output can contain quotes, backticks, newlines, and $ — never interpolate it inline into a shell command. Write each subagent's raw result to a temp file with the Write tool (e.g. .devflow/tmp/result-<n>.json), then operate on the file. For each result:

1. Attempt to parse it: $LIB/../scripts/run-jq.sh -c . < .devflow/tmp/result-<n>.json

2. A Stage A defined skip (issue #626) is recognized by the presence of a top-level "skip" key ONLY — never by matching substrings of any error text (agent-authored free text is data, never a discriminator). A "skip"-keyed return is terminal: no retry, no blocker. Whether it leaves a marker depends on the bundle's workpad_final_status (a mechanical field, not the skip text):

   • Cancelled → a permanently-terminal skip → append a marker entry so the PR is seen as handled next run: $LIB/../scripts/run-jq.sh -cn --argjson pr <n> --arg reason "<the skip .reason>" '{kind:"skip", pr:$pr, reason:$reason}' >> .devflow/tmp/new-entries.jsonl
   • an interim state (Setup/Discovering/…/Documenting) → a transient skip → append no marker, so the PR stays unprocessed and is re-scanned while it remains inside the 7-day merge lookback.

   Either way, record the skip explicitly — append the report record AND increment the counter, in the same two statements the mechanical branch above uses, so skipped_count and the rendered skips[] can never diverge:

   skip_records+=("PR #<n> skipped (Stage A, <Cancelled|interim>): <the skip .reason>")
   skipped_count=$((skipped_count + 1))
   

   That record is what Step 9 renders. Every skip — the mechanical pre-dispatch skip above, the Cancelled skip, and the interim skip — writes a one-line report record, so no skip is ever silent.

3. Otherwise, if parsing fails or the object has an "error" key (a genuine failure), retry the subagent once with the same prompt.

4. If still malformed (or still "error"-keyed) after one retry, record a blocker: "PR #<n>: retrospective analysis failed" and skip that PR.

5. If valid (a real retrospective entry, no "skip"/"error" key), append: $LIB/../scripts/run-jq.sh -c . < .devflow/tmp/result-<n>.json >> .devflow/tmp/new-entries.jsonl

---

Step 5 — Materialize

Merge all new entries into the retrospectives file (idempotent — existing entries for the same pr+kind are replaced):

bash $LIB/materialize-retrospectives.sh \
  .devflow/tmp/new-entries.jsonl \
  .devflow/learnings/retrospectives.jsonl

The script prints "materialized: appended N, replaced M" to stdout.

---

Step 6 — Derive actionable patterns

bash $LIB/actionable-patterns.sh \
  .devflow/learnings/retrospectives.jsonl \
  .devflow/learnings/overrides.json \
  > .devflow/tmp/patterns.json

Print a summary line to the console, for example:

5 PRs: 3 clean, 2 analyzed; 2 actionable patterns: incomplete-edit (x5), lenient-verdict (x3)

Partition patterns.json into two lists:

to_act=$($LIB/../scripts/run-jq.sh '[.[] | select(.cooldown_active == false)]' .devflow/tmp/patterns.json)
cooldown_skipped=$($LIB/../scripts/run-jq.sh '[.[] | select(.cooldown_active == true) | .tag]' .devflow/tmp/patterns.json)

Record cooldown_skipped tags for the final report.

---

Step 6.5 — Build experiment records (best-effort)

After Step 5 materialized this week's retrospective entries (and before the Step 7 state PR commits the learnings files), assemble the unified experiment record — joining each merged PR's per-run cost to its review outcome (verdict, Important-finding count, denial count, config fingerprint). Anchored here so this week's PRs join against this week's freshly-materialized retrospective entries.

This is a best-effort step and never blocks the retrospective: a non-zero exit is logged as a breadcrumb and the run continues. Carry that breadcrumb into the Step 9 status report as a blocker note so the failure is visible, then proceed.

A non-zero exit means some PRs did not make it into the store — not necessarily that nothing was written. The assembler exits 2 when any candidate failed to assemble or had an unestablished merge state (a gh outage, an unresolvable repo), and it still writes the records that did assemble, so a partial failure leaves a partially-updated store. Report it as "N PRs missing from the experiment store," not as "the store was not updated." An unestablished PR does not backfill by itself — it never enters the store, and only stored or retrospective-listed PRs are re-selected as candidates — so name the PRs from the breadcrumb and re-run with --prs once the cause is resolved.

Before the reader runs, fetch the telemetry branch (issue #441) into its local ref so build-experiment-records.py can union each run's durable record off that branch with any legacy tracked .devflow/logs/. Best-effort: on a fresh repo the branch does not exist yet (nothing has persisted to it) and the fetch is a harmless no-op — the reader then reads the legacy archive alone; the retrospective is never blocked by a missing telemetry branch.

# Fetch the telemetry branch into its local ref — deliberately NO force `+` refspec: the
# local ref can be AHEAD of the remote (offline-accumulated `--persist` commits not yet
# reconciled by a writer's union re-parent), and a forced fetch would rewind it and
# permanently orphan those records. A plain fetch fast-forwards the common case and rejects
# exactly the local-ahead/diverged case, leaving the local ref — and its records — intact
# for the reader; reconciling with the remote belongs to the next writing run's
# fetch → re-parent → push loop, never to this read-side fetch.
# The reader (`_index_efficiency`) reads it by its local branch name.
#
# Resolve the branch through the SAME resolver the writer uses — never a third, independent
# re-derivation. `devflow_telemetry_branch` (lib/telemetry-branch.sh) owns the full contract:
# the `.telemetry.branch` read, config-get.sh's coercion, and the `git check-ref-format`
# fallback to the default. A bare `config-get.sh` read here applies NEITHER guard, and both
# gaps end the same way — this fetch targets a branch nobody wrote to, the local ref is never
# populated, and on a fresh clone or a cloud checkout the reader then finds nothing and every
# cost row silently goes missing:
#   * a schema-valid but git-invalid name ("my branch", "a..b") → the writer persisted to
#     `devflow-telemetry`, but this would try to fetch `my branch`;
#   * a malformed config → config-get.sh prints NOTHING and exits 2, so TELEMETRY_BRANCH is
#     empty and the command degrades to `git fetch origin ":"`, while the writer and reader
#     both fell back to `devflow-telemetry`.
# Source the lib and ask it. The `||` keeps this best-effort: if the lib cannot be sourced, fall
# back to the same default the resolver would have returned.
#
# Do NOT redirect the resolver's stderr to /dev/null. Its breadcrumb is the whole reason for
# routing through it: on a git-invalid `telemetry.branch` it is the one place that names the
# config key to fix. Silencing it would leave an operator with a broken config reading a
# perfectly healthy-looking retrospective that quietly contains no telemetry-branch rows.
# Only stdout is captured; stderr flows to the run log.
#
# Note the lib sources `config-source.sh`, which sets `set -euo pipefail` in THIS shell. Every
# command below is `||`-guarded, so that is harmless today — but keep new commands in this fence
# guarded, or errexit will abort the step.
. "$LIB/telemetry-branch.sh" || true
TELEMETRY_BRANCH=$(devflow_telemetry_branch) || TELEMETRY_BRANCH=""
[ -n "$TELEMETRY_BRANCH" ] || TELEMETRY_BRANCH=devflow-telemetry
git fetch origin "${TELEMETRY_BRANCH}:${TELEMETRY_BRANCH}" 2>/dev/null || \
  echo "retrospective-weekly: could not fetch telemetry branch '${TELEMETRY_BRANCH}' (absent on a fresh repo, offline, or the local ref has commits the remote lacks) — the experiment-record reader unions whatever local '${TELEMETRY_BRANCH}' ref exists (if any) with any legacy tracked .devflow/logs/" >&2
python3 $LIB/../scripts/build-experiment-records.py || \
  echo "retrospective-weekly: build-experiment-records.py exited non-zero (rc=$?) — one or more PRs are MISSING from the experiment store (see its stderr for which, and whether they failed to assemble or had an unestablished merge state); records that did assemble were still written" >&2

The assembler is idempotent (one line per merged PR, keyed by PR number) and incremental (it processes merged PRs absent from .devflow/learnings/experiment-records.jsonl plus any passed via --prs, never a full-history sweep), so re-running is safe and cheap. It runs on the local/interactive retrospective tier only — it is never invoked from a workflow. See docs/efficiency-trace.md for the store schema and the abandoned-run bias.

---

Step 7 — State PR

Open the state PR now, before Stage B, so that the learnings files are committed onto their own branch. This captures the unstaged changes Steps 5–6 wrote to .devflow/learnings/ before any issue is filed, so this run's retrospective data survives even if Stage B or the filing step fails partway.

Ensure you are on main:

git checkout main

The working tree now has the updated .devflow/learnings/retrospectives.jsonl (and possibly a modified .devflow/learnings/overrides.json from meta-issue dismissals in a previous run). These changes are in-place on main's working tree and have never been committed to main — open-state-pr.sh handles committing them onto a separate branch.

STATE_PR=$(bash $LIB/open-state-pr.sh)

open-state-pr.sh (no required args; optional --branch <name>, --base <ref> — defaults to main —, and --dry-run):

• Creates/reuses branch devflow/learnings-<YYYY-MM-DD> from --base (main by default), so the PR diff is just the learnings files even if the operator was on a feature branch.
• Stages any learnings files that exist (.devflow/learnings/retrospectives.jsonl and, if present, .devflow/learnings/overrides.json).
• Commits and pushes (force-with-lease if the remote branch exists).
• Opens or updates the PR against main.
• Prints the PR number to stdout.

After it returns, go back to main so the working tree is clean and Stage B starts from a known-good HEAD:

git checkout main

Initialize Stage B counters:

intervention_issues=()   # will hold {tag, url} objects — one per filed pattern
blockers=()              # will hold strings

---

Step 8 — Stage B: File one issue per actionable pattern

For each actionable pattern, a Stage B subagent drafts a {title, body} issue spec and the orchestrator files exactly one GitHub issue from it via meta-issue.sh. No worktrees, no commits, no PRs — the loop proposes; a human triages each issue and runs it through the normal /devflow:implement → review pipeline. Your main checkout stays on main and is never edited. The drafting subagents (8b) parallelize; the cheap filing (8c) is done serially.

8a — Gather occurrence bundles

For each pattern in to_act, make sure every occurrence bundle is on disk (fetch the ones not already fetched this run):

for n in $($LIB/../scripts/run-jq.sh -r '.occurrences[].pr' <<< "$pattern"); do
    [ -f ".devflow/tmp/pr-${n}.context.json" ] || bash $LIB/fetch-pr-context.sh "$n" >/dev/null
done

Record, per pattern: SLUG ($LIB/../scripts/run-jq.sh -r .slug <<< "$pattern"), TAG ($LIB/../scripts/run-jq.sh -r .tag <<< "$pattern"), the JSON array of absolute bundle paths, and the pattern object.

8b — Dispatch all Stage B subagents concurrently

Issue one Agent call per pattern, all in a single message so they run in parallel. No worktree is created or passed — the subagent makes no edits. Each subagent's prompt:

Read and follow "${CLAUDE_SKILL_DIR:-<absolute skill base directory this runner reports in context>}"/../retrospective-audit/SKILL.md exactly.

Occurrence-PR context bundle paths (absolute): <json array of paths>

Pattern metadata: <the pattern json object>

Make no edits and no worktree. Print exactly one JSON object (the {title, body} return contract from § 5 of that skill) and nothing else on stdout.

Wait for all subagents to finish. Pair each result JSON with its pattern.

8c — File one issue per pattern (serial)

For each (pattern, result) pair, in any order:

Write the subagent's raw result to .devflow/tmp/result-${SLUG}.json with the Write tool (it can contain quotes, backticks, newlines, and $ — never interpolate it inline into a shell command), then:

# 1. Parse + validate the {title, body} contract. Malformed → blocker, continue.
# The wrapper precheck is a SEPARATE single-statement branch (no rc variable carried
# across statements — the inline-bash marshaling constraint the anchor note documents),
# so an unexpanded $LIB or missing/non-executable run-jq.sh is reported as the anchor
# failure it is, never misdiagnosed as "malformed subagent JSON". `[ ! -x ]` (not `[ ! -e ]`)
# so a PRESENT-but-non-executable wrapper (lost its +x bit → an invocation would exit 126)
# is caught here as an anchor/wrapper failure, mirroring the execution-verified `resolve-*.sh`
# family (#247) — an existence-only check would let it fall through to the elif and mis-read
# the exit-126 as malformed JSON.
if [ ! -x "$LIB/../scripts/run-jq.sh" ]; then
    blockers+=("Pattern ${SLUG}: run-jq.sh wrapper not found or not executable (unexpanded \$LIB notation, missing wrapper, or lost +x bit; fix the anchor) — not filed")
elif ! $LIB/../scripts/run-jq.sh -e '.title and .body' < ".devflow/tmp/result-${SLUG}.json" >/dev/null 2>&1; then
    # Malformed: record a blocker and file NOTHING. The append below is the
    # load-bearing failure path — it MUST run (the run reports this pattern as a
    # blocker, never as filed), so it is concrete shell, not a comment.
    blockers+=("Pattern ${SLUG}: Stage B subagent returned malformed JSON (missing title/body) — not filed")
else
    # 2. Extract the body to a file (via jq, so backticks/$/newlines never hit
    #    the shell) and the title to a shell var.
    $LIB/../scripts/run-jq.sh -r '.body' < ".devflow/tmp/result-${SLUG}.json" > ".devflow/tmp/issue-body-${SLUG}.md"
    TITLE="$($LIB/../scripts/run-jq.sh -r '.title' < ".devflow/tmp/result-${SLUG}.json")"

    # 3. File exactly one issue. meta-issue.sh stamps DevFlow + Retrospective
    #    (best-effort), records the overrides.json cooldown, is idempotent (an
    #    open issue for this pattern → recurrence comment, not a duplicate), and
    #    fails CLOSED (non-zero exit) on a de-dup-lookup error or a create that
    #    returned no usable issue URL. An overrides-write failure AFTER a
    #    successful create is the one exception: the issue genuinely exists, so it
    #    reports FILED (exit 0 + URL + a loud ::error:: breadcrumb), not blocked —
    #    the next run's de-dupe recovers the missing cooldown.
    if ISSUE_URL="$(bash $LIB/meta-issue.sh \
            --tag "$TAG" \
            --slug "$SLUG" \
            --title "$TITLE" \
            --body-file ".devflow/tmp/issue-body-${SLUG}.md" \
            --overrides .devflow/learnings/overrides.json)"; then
        # Success: record {tag, url} in intervention_issues.
        intervention_issues+=("$($LIB/../scripts/run-jq.sh -nc --arg tag "$TAG" --arg url "$ISSUE_URL" '{tag:$tag,url:$url}')")
    else
        # meta-issue.sh exited non-zero (de-dup lookup / create-returned-no-URL;
        # an overrides-write failure does NOT land here — it reports FILED).
        # Record a blocker and file NOTHING — the pattern stays absent from
        # intervention_issues. Concrete append, same reason as above.
        blockers+=("Pattern ${SLUG}: meta-issue.sh failed to file the issue — not filed")
    fi
fi

Never report a pattern as filed when it was not. A malformed Stage B result or a meta-issue.sh non-zero exit records a per-pattern blocker and the run continues to the next pattern; the pattern is absent from intervention_issues.

Do not post /devflow:implement (or any auto-trigger comment) on a filed issue — filed issues await human triage.

(meta-issue.sh mutates .devflow/learnings/overrides.json in your main checkout's working tree. That happens after the Step 7 state PR was opened, so the new cooldown lands in next week's state PR — see § Notes for the optional follow-up commit if you want it in this run's PR.)

---

Step 9 — Status report

Collect the per-analyzed-PR digest lines (verdict + a one-line summary) and the full pattern list (acted-on, cooldown-skipped, dismissed, and below-threshold — the same patterns.json from Step 6) so the report shows the whole picture, not just the PRs that produced an intervention:

ANALYZED_JSON="$($LIB/../scripts/run-jq.sh -sc '[.[] | select(.verdict == "imperfect" or .verdict == "blocked") | {pr, verdict, summary}]' .devflow/tmp/new-entries.jsonl)"
PATTERNS_JSON="$(cat .devflow/tmp/patterns.json)"
RECURRING_TARGETS_JSON="$(bash $LIB/recurring-targets.sh .devflow/learnings/retrospectives.jsonl)"

recurring-targets.sh groups every accumulated entry's suggested_interventions[].candidate_targets[] by exact target path and emits only the targets named in ≥2 distinct PRs (report-only; [] when nothing recurs, which render-report.sh then omits).

Build the summary JSON and assign it to $SUMMARY_JSON:

SUMMARY_JSON="$($LIB/../scripts/run-jq.sh -nc \
  --argjson prs_scanned         "$prs_scanned" \
  --argjson clean_count         "$clean_count" \
  --argjson analyzed_count      "$analyzed_count" \
  --argjson skipped_count       "$skipped_count" \
  --argjson skips               "$(printf '%s\n' "${skip_records[@]:-}" | $LIB/../scripts/run-jq.sh -sRc 'split("\n") | map(select(. != ""))')" \
  --argjson analyzed            "$ANALYZED_JSON" \
  --argjson patterns            "$PATTERNS_JSON" \
  --argjson recurring_targets   "$RECURRING_TARGETS_JSON" \
  --argjson intervention_issues "$(printf '%s\n' "${intervention_issues[@]:-}" | $LIB/../scripts/run-jq.sh -sc '.')" \
  --argjson cooldown_skipped    "$(printf '%s\n' "${cooldown_skipped[@]:-}"    | $LIB/../scripts/run-jq.sh -sc '.')" \
  --argjson blockers            "$(printf '%s\n' "${blockers[@]:-}"            | $LIB/../scripts/run-jq.sh -sc '.')" \
  --argjson state_pr            "$STATE_PR" \
  '{prs_scanned:$prs_scanned,clean_count:$clean_count,analyzed_count:$analyzed_count,
    skipped_count:$skipped_count,skips:$skips,
    analyzed:$analyzed,patterns:$patterns,recurring_targets:$recurring_targets,
    intervention_issues:$intervention_issues,
    cooldown_skipped:$cooldown_skipped,blockers:$blockers,state_pr:$state_pr}')"

(The "${array[@]:-}" form handles an empty bash array safely under set -u. render-report.sh renders the analyzed and patterns sections only when those keys are non-empty, so an older caller that omits them still works.)

Render the report markdown and post it as a comment on the state PR:

source $LIB/render-report.sh
devflow_render_report "$SUMMARY_JSON" > .devflow/tmp/report.md
bash $LIB/post-status.sh --pr "$STATE_PR" --report-file .devflow/tmp/report.md

---

Step 10 — Report to the user

Print the rendered report (cat .devflow/tmp/report.md) to the console.

Then list each item that needs human action:

• State PR (contains the updated retrospectives): https://github.com/<repo>/pull/<state_pr>
• Filed issues (one per actionable pattern, awaiting human triage): list each as <tag>: <url>

If there are any blockers, list them explicitly.

Tell the user:

Review and merge the state PR once CI passes. Each filed issue awaits human triage — pick the ones worth acting on and run them through the normal implement → review pipeline; the loop never starts that for you. The loop is idempotent — re-running next week will only process new PRs not yet in retrospectives.jsonl on main, and a pattern already filed this cycle is not re-filed.

Do not run gh pr merge --auto on anything, and do not auto-start implementation on a filed issue. The maintainer triages and merges manually after reviewing.

---

§ Cron / headless variant

claude -p "/devflow:retrospective-weekly" --permission-mode acceptEdits handles steps 1–9 unattended. Stage B makes no working-tree edits (it only drafts issue specs) and the orchestrator only files issues and writes .devflow/learnings/, so the loop is well-suited to an unattended run. For a fully unattended run, add --dangerously-skip-permissions.

---

§ Notes

• Clean working tree required. The loop modifies .devflow/learnings/ in-place on main's working tree; starting dirty risks mixing pre-existing changes into the state PR commit.

• State PR before Stage B. Opening the state PR (Step 7) before Stage B is intentional: it commits the learnings files onto devflow/learnings-<date> before any issue is filed, so this run's retrospective data is captured even if Stage B or the filing step fails partway. Stage B never touches your main checkout — it makes no edits at all.

• Issue-per-pattern. Stage B dispatches one drafting subagent per actionable pattern concurrently (each returns a {title, body} spec, no edits), then the orchestrator files exactly one GitHub issue per pattern via meta-issue.sh. No worktrees, no commits, no PRs — the loop proposes; a human implements.

• Overrides after Stage B. meta-issue.sh records each filed pattern's cooldown in .devflow/learnings/overrides.json in your main working tree after the Step 7 state PR was opened, so the change lands in next week's state PR automatically. If you want it in this run's PR, after Step 8 push a follow-up commit onto the same devflow/learnings-<date> branch:

  if ! git diff --quiet HEAD -- .devflow/learnings/overrides.json 2>/dev/null; then
      LB="devflow/learnings-$(date -u +%F)"
      git fetch origin "$LB"
      git checkout "$LB"
      git add .devflow/learnings/overrides.json
      git commit -m "chore(devflow): add overrides from Stage B filed issues"
      git push --force-with-lease origin "$LB"
      git checkout main
  fi
  

• Idempotent. Re-running re-processes only PRs whose number is not already in retrospectives.jsonl on main. A pattern already filed this cycle is not re-filed: meta-issue.sh finds the open issue and adds a recurrence comment instead of a duplicate, and the overrides.json cooldown excludes the pattern on subsequent runs.

• Never auto-merge, never auto-implement. The maintainer merges the state PR manually after CI, and triages each filed issue manually — the loop never starts an implement run for you.

• materialize-retrospectives.sh signature: takes two explicit positional args — <new-entries-file> and <jsonl-path>. Always pass both.

• actionable-patterns.sh signature: takes two explicit positional args — <retrospectives.jsonl> and <overrides.json>. Always pass both.

• open-state-pr.sh signature: no required args; optional --branch, --base (defaults to main), --dry-run; prints the PR number to stdout.

• fetch-pr-context.sh return value: echoes the bundle file path to stdout; the bundle content is on disk at .devflow/tmp/pr-<n>.context.json.

• cheap-gate.jq invocation: reads from stdin (the bundle content, not the path) — use $LIB/../scripts/run-jq.sh -c -f $LIB/cheap-gate.jq < "$CTX" where $CTX is the path.