chain (multi-root): 3 issues — spectra-archive sync + cluster glob + pasted-image cache

.agents/skills/spectra-archive/SKILL.md

@@ -11,10 +11,27 @@ metadata:
 
 Archive a completed change.
 
-**Input**: Optionally specify a change name after `$spectra-archive` (e.g., `$spectra-archive add-auth`). If omitted, check if it can be inferred from conversation context. If vague or ambiguous you MUST prompt for available changes.
+**Input**: Optionally specify a change name after `/spectra-archive` (e.g., `/spectra-archive add-auth`). If omitted, check if it can be inferred from conversation context. If vague or ambiguous you MUST prompt for available changes.
 
 **Prerequisites**: This skill requires the `spectra` CLI. If any `spectra` command fails with "command not found" or similar, report the error and STOP.
 
+**Step 0: Bootstrap Stage Task List** (required)
+
+Before doing anything else, call `TaskCreate` to build a harness-level todo list for this archive run — one entry per step below. Mark each `TaskUpdate → completed` as you finish it; silent completion is a violation. This mirrors the Step 0 Bootstrap discipline every `idd-*` skill enforces, and gives per-step accountability when `/idd-close` cascades into `/spectra-archive`.
+
+```
+TaskCreate(name="prompt_change_name", description="Step 1: resolve change name — prompt via AskUserQuestion if not provided / inferable")
+TaskCreate(name="check_artifacts", description="Step 2: spectra status --json — warn + confirm if any artifact not done")
+TaskCreate(name="check_tasks", description="Step 3: scan tasks.md — warn + confirm if incomplete - [ ] tasks")
+TaskCreate(name="assess_delta_sync", description="Step 4: compare delta specs vs main specs; prompt sync now / archive without sync")
+TaskCreate(name="cleanup_tracking", description="Step 5: rm -f .spectra/touched/<name>.json")
+TaskCreate(name="run_archive_cli", description="Step 6: spectra archive <name>")
+TaskCreate(name="post_implementation_complete", description="Step 7: invoke spectra-archive-post-ic.sh to post ## Implementation Complete to the linked GitHub issue (#56)")
+TaskCreate(name="display_summary", description="Step 8: read Step 7 outcome + show archive completion summary")
+```
+
+Complete each step → `TaskUpdate → completed` immediately.
+
 **Steps**
 
 1. **If no change name provided, prompt for selection**
@@ -93,13 +110,111 @@ Archive a completed change.
 
    **If archive fails** with "already exists" error, suggest renaming existing archive.
 
-7. **Display summary**
+7. **Post `## Implementation Complete` to linked GitHub issue (v1.3+, PsychQuant/issue-driven-development#56)**
+
+   **Purpose**: ensures `/idd-close` Step 0 supersession gate triggers for Spectra-path issues, removing the need for manual retroactive Implementation Complete synthesis.
+
+   **Delegated to executable helper script** `.claude/scripts/spectra-archive-post-ic.sh` (with unit tests at `.claude/scripts/tests/spectra-archive-post-ic/`). The script is the source of truth — this skill calls it and reads the outcome. Behavior contract (detection / idempotent guard / safe body composition / multi-candidate handling) lives in the script + its tests, not in skill prose. This separation was introduced after R2 verify found that prose-with-illustrative-bash had structural defects (variable persistence across Bash tool calls, Python3 RCE via shell-string interpolation, etc.) — see PsychQuant/issue-driven-development#56 R2 verify report.
+
+   **Required inputs from caller (agent)**: before invoking Step 7, the agent MUST have these in scope (from earlier skill steps):
+   - `$CHANGE_NAME` — the Spectra change name (slug; same value passed to `spectra archive`). **This exact value must also be reused, byte-identical, in Step 8** — the outcome file path is derived from it, so any drift (typo / whitespace / case) makes Step 8 read the wrong path.
+   - `$SPEC_DELTAS` (optional) — comma-separated capability names from `spectra archive` stdout in Step 6 (e.g., `"idd-all-chain, idd-spawn-manifest"`); defaults to placeholder if absent
+
+   **Invocation**:
+
+   ```bash
+   # Resolve repo root to make the script path cwd-independent — the skill may be
+   # invoked from a subdirectory (cd openspec && /spectra-archive ...). Relative
+   # path .claude/scripts/... breaks; absolute path via git-root prefix doesn't.
+   REPO_ROOT=$(git rev-parse --show-toplevel 2>/dev/null || pwd)
+
+   ARCHIVE_DIR="${REPO_ROOT}/openspec/changes/archive/$(date +%Y-%m-%d)-${CHANGE_NAME}"
+   SPEC_DELTAS="${SPEC_DELTAS:-(see archived change directory)}"
+
+   # The helper script DERIVES the outcome file path internally from
+   # --change-name: /tmp/spectra-archive-ic-outcome-<change-name>.txt
+   # (--change-name is allowlist-validated inside the script before being used
+   # in the path, so the derived path is always traversal-safe). The formula
+   # has a single source of truth — the script — so this skill does NOT pass
+   # --outcome-file. Step 8 recomputes the same path from the same $CHANGE_NAME
+   # to read the outcome across separate Bash tool calls.
+   # Why deterministic, not $$-$(date +%s): a random suffix cannot be recomputed
+   # in a second shell, which breaks the Step 7 → Step 8 handoff (R4 verify
+   # R4-S1 finding). The change name is the skill's primary input — always known
+   # to Step 8 — so the derived path IS recoverable.
+   bash "${REPO_ROOT}/.claude/scripts/spectra-archive-post-ic.sh" \
+       --change-name "$CHANGE_NAME" \
+       --archive-dir "$ARCHIVE_DIR" \
+       --spec-deltas "$SPEC_DELTAS"
+   POST_IC_EXIT=$?
+   ```
+
+   **Exit codes**:
+
+   | Exit | Meaning | Agent action |
+   |------|---------|--------------|
+   | `0` | Success or normal skip (posted / none / idempotent / unsafe-name / generic failure) | Read outcome from the derived outcome file (see Step 8), proceed to Step 8 |
+   | `2` | Usage error (missing args, or unsafe `--outcome-file` path) | Skill bug — fix invocation |
+   | `64` | Dependency missing (python3) | Surface to user; archive itself succeeded; manual retry after install |
+   | `75` | Multi-candidate detected | **AskUserQuestion required** — read `/tmp/spectra-archive-candidates.txt` for the candidate list, prompt user to pick canonical issue (show `#N + gh issue title` for each), then re-invoke script with `--linked-issue <chosen>` |
+
+   **Stdout**: a single line that is either the IC comment URL, or one of the documented status messages (`(none — ...)`, `(skipped — ...)`, `(pending — ...)`, `(failed — ...)`). The same line is also written to the derived outcome file `/tmp/spectra-archive-ic-outcome-${CHANGE_NAME}.txt` for cross-Bash-call persistence (Step 8 reads from there).
+
+   **Multi-candidate flow (agent responsibility)**:
+
+   ```bash
+   if [ "$POST_IC_EXIT" = "75" ]; then
+     # Read candidates + AskUserQuestion + re-invoke (script re-derives the outcome path from --change-name)
+     CANDIDATES=$(cat /tmp/spectra-archive-candidates.txt)
+     # For each candidate, fetch title via `gh issue view <N> --json title -q .title`
+     # Then AskUserQuestion: "Multi-candidate detected: which is canonical?"
+     # User picks → CHOSEN_ISSUE=<N>
+     bash "${REPO_ROOT}/.claude/scripts/spectra-archive-post-ic.sh" \
+         --change-name "$CHANGE_NAME" \
+         --archive-dir "$ARCHIVE_DIR" \
+         --spec-deltas "$SPEC_DELTAS" \
+         --linked-issue "$CHOSEN_ISSUE"
+   fi
+   ```
+
+   The agent (LLM-driven) handles the AskUserQuestion step — bash cannot prompt. The script validates `$CHOSEN_ISSUE` against the original candidate set on re-invoke.
+
+   **Failure semantics**: any failure in Step 7 (gh auth lost, network, body too large, etc.) is recorded in the outcome file but does NOT abort the overall archive operation — the archive itself (Step 6) has already succeeded, and the archived change directory + main spec deltas are the canonical record. The GitHub comment is the convenience anchor for `/idd-close` supersession.
+
+   **Testing**: run `.claude/scripts/tests/spectra-archive-post-ic/test.sh` to validate the script against fixture archive directories (covers explicit-marker / Refs-fallback / no-marker / multi-candidate / malicious-tasks.md / missing-tasks.md / unsafe-change-name / linked-issue-resolved / linked-issue-invalid / outcome-path-derivation / unsafe-outcome-file). All 11 fixtures pass.
+
+8. **Display summary**
+
+   Read the outcome from Step 7. Recompute the outcome file path the same way the
+   script derived it — from the **identical** `$CHANGE_NAME` slug passed to
+   `spectra archive` in Step 6. This works whether Step 7 + Step 8 ran in the same
+   Bash invocation (var still in scope) OR in separate Bash calls (the formula is
+   deterministic). **The `$CHANGE_NAME` here MUST be byte-identical to Step 7's —
+   no typo, trailing whitespace, or case difference** — otherwise Step 8 reads a
+   different path:
+
+   ```bash
+   # Recompute the same path the script derived in Step 7.
+   # MUST use the identical $CHANGE_NAME slug — see contract note in Step 7.
+   OUTCOME_FILE="/tmp/spectra-archive-ic-outcome-${CHANGE_NAME}.txt"
+
+   if [ -f "$OUTCOME_FILE" ]; then
+       IMPLEMENTATION_COMPLETE_POSTED=$(cat "$OUTCOME_FILE")
+   else
+       # Loud failure — NOT a quiet "(unknown)". A missing outcome file means
+       # either Step 7 never ran, or $CHANGE_NAME drifted between Step 7 and
+       # Step 8. Both are real errors the user must see.
+       IMPLEMENTATION_COMPLETE_POSTED="⚠️  ERROR — outcome file not found: $OUTCOME_FILE (Step 7 did not run, or \$CHANGE_NAME drifted between Step 7 and Step 8)"
+       echo "WARNING: spectra-archive Step 8 — outcome file missing: $OUTCOME_FILE" >&2
+   fi
+   ```
 
    Show archive completion summary including:
    - Change name
    - Schema that was used
    - Archive location
    - Spec sync status (synced / sync skipped / no delta specs)
+   - **Implementation Complete posted to:** `$IMPLEMENTATION_COMPLETE_POSTED`
    - Note about any warnings (incomplete artifacts/tasks)
 
 **Output On Success**
@@ -172,3 +287,5 @@ Target archive directory already exists.
 - If sync is requested, use the Skill tool to invoke `spectra-sync-specs` (agent-driven)
 - If delta specs exist, always run the sync assessment and show the combined summary before prompting
 - If **AskUserQuestion tool** is not available, ask the same questions as plain text and wait for the user's response
+- **Step 8 multi-candidate disambiguation**: if linked-issue detection (Fallback 1 explicit marker) returns multiple distinct `#N` values, MUST prompt user via AskUserQuestion to pick the canonical one — never auto-pick to avoid posting to the wrong issue
+- **Step 8 silent skip on no linked issue**: do not warn or prompt; not all archives have GitHub trackers (legacy archives, design-only changes). Reflect skip reason in Step 7 summary line

plugins/issue-driven-dev/CHANGELOG.md

@@ -7,6 +7,25 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [2.70.0] - 2026-05-20
+
+### Fixed
+
+- **`idd-issue` Step 1 pasted-image immediate-persistence** ([#112](https://github.com/PsychQuant/issue-driven-development/issues/112)): Claude Code's `~/.claude/image-cache/<session-id>/` is per-session + cleared by context compaction / session lifecycle / session-id rollover. Step 1 → Step 4 separation (read annotation in Step 1, upload in Step 4) spans `AskUserQuestion` + Step 2.5/2.6 + Step 3 `gh issue create` + Step 4 upload — easily long enough for cache eviction. 2026-05-20 downstream incident (`kiki830621/ai_martech_global_scripts#788`) hit exactly this failure mode. NEW immediate-persistence rule: when Step 1 encounters `[Image: source: <path>]` annotation, `cp` to `/tmp/idd-issue-attachments/issue_pending_<ts>_<rand>.png` in the **same tool turn** that reads the annotation; Step 4 references the staged path, not the original cache path. Anonymous `/tmp` staging (POSIX-safe, system-cleanup-friendly, no repo pollution) per `feedback_lead_minimal`. Fallback for already-evicted source: warn + continue without that attachment.
+
+### Refactored
+
+- **`spectra-archive` skill `.agents/` ↔ `.claude/` sync** ([#93](https://github.com/PsychQuant/issue-driven-development/issues/93)): #93 surfaced 3-copy divergence between `.claude/skills/`, `.agents/skills/`, and `plugins/.../references/spectra-skills/`. Investigation refuted the diagnose-time recommendation to delete `.agents/` — 4 openspec specs reference `.agents/skills/spectra-*/SKILL.md` as Spectra-tier dependencies (the path is LIVE, not legacy). Revised disposition: sync `.agents/skills/spectra-archive/SKILL.md` from `.claude/` so both LIVE load paths carry the v1.3+ Implementation Complete auto-post feature (#56). `plugins/.../references/spectra-skills/spectra-archive/` left as historical snapshot (no markdown cross-refs found; low cleanup ROI per `lead-minimal`). **Sister-skill divergences out of scope**: 7 other spectra-* skills (audit / discuss / propose / apply / ingest / debug / commit) also diverge between `.claude/` and `.agents/` — audit comment on #93 documents the drift matrix. **Sister issues NOT auto-filed in this PR** per `feedback_lead_minimal` — drift documented as observation, separate issues will be filed if specific divergence causes user-visible friction. (Original wording "filed for separate follow-up as needed" was misleading per #115 DA finding DA-1 — no issues actually filed.) Drift-prevention CI hook deferred until drift recurs naturally.
+
+- **`idd-implement` cluster detection glob hardening + Option A-final doc** ([#100](https://github.com/PsychQuant/issue-driven-development/issues/100)): two non-blocking findings from PR #99 (#96) verify rounds.
+  - **Finding 1 (design)** — Option A (cluster mode unconditionally forces PR regardless of branch context) confirmed final. NEW `### Feature-branch + cluster + direct-commit — rejected case` subsection in `references/pr-flow.md` § Cluster mode override documenting the rejected Option B (branch-context-gated cluster direct-commit) with comparison table + rationale. Contract simplicity wins; feature-branch direct-commit workflow remains viable for single-issue `--no-pr` invocations.
+  - **Finding 2 (refactor)** — `idd-implement` Step 0.5 cluster detection bash hardened. Previous glob `\#[0-9]*` over-matched (`#42abc` counted, `#34 #34` over-counted as 2). Replaced with strict integer check (`[[ "$arg_num" =~ ^[0-9]+$ ]]`) + associative-array dedup matching the documented `^#\d+$` form in `batch-and-cluster.md`. 0 behavior change for well-formed distinct invocations. **Quiet behavior change for malformed tokens** (per #115 DA finding DA-2): pre-v2.70.0, `#42abc` was counted as a cluster member (causing later failures when used as issue number); post-v2.70.0 it's silently skipped from the count. Users invoking with typo'd tokens get cluster-mode evaluation based on well-formed tokens only — failure modes shifted from "fail mid-loop on bad number" to "treat as if token not present".
+
+### Notes
+
+- Plugin v2.70.0 is a **minor** bump (over v2.69.0) covering 3 issues across `idd-issue` + `idd-implement` + `pr-flow.md` + `.agents/skills/spectra-archive/SKILL.md`. All changes additive (#112 immediate-persistence + #93 sync + #100 glob hardening + Option A-final documentation). Cluster PR for review surface — verify ensemble runs over the cumulative diff.
+- Marketplace.json sync deferred to `/idd-close` Step 6.5 chain (per repo precedent).
+
 ## [2.69.0] - 2026-05-20
 
 ### Fixed

plugins/issue-driven-dev/references/pr-flow.md

@@ -65,6 +65,23 @@ Then proceeds as PR path. **No abort, no silent ignore** — the flag is acknowl
 
 **Single-issue invocation behavior is unchanged** — the cluster carve-out only fires on ≥2 `#N`. Backward compatibility preserved.
 
+#### Feature-branch + cluster + direct-commit — rejected case (v2.70.0+, #100 Finding 1)
+
+PR #99 (#96 implementation, Option A) chose to unconditionally force PR for cluster mode, regardless of the starting branch. Devil's Advocate during verify flagged that the "force PR" rationale (stacked half-isolated changes on default branch) **only holds when the user starts from the default branch**. On a non-default feature branch, cluster direct-commit just stacks N `Refs #N` commits on that feature branch — a legitimate workflow (one local feature tracking N issues, shipped as one PR later).
+
+The alternative was **Option B** (branch-context-gated cluster direct-commit): if current branch != default branch, honor `--no-pr` / `pr_policy=never`; otherwise force PR as today. This issue confirms **Option A is final**:
+
+| Aspect | Option A (current) | Option B (rejected) |
+|--------|--------------------|--------------------|
+| Contract simplicity | Cluster → PR. Uniform regardless of branch context. | Cluster → PR if default branch, else direct-commit. Two paths. |
+| Override notice | One wording, mirrors fork detection | Two wordings depending on branch context |
+| `git symbolic-ref` dependency | None | Required (detached HEAD / merge-state edge cases) |
+| Cluster-on-feature-branch frequency | Rare (most cluster invocations are explicit `--pr`) | Same rare frequency, but now requires branch-context check overhead |
+
+**Recommendation**: keep Option A. The feature-branch direct-commit workflow remains viable for **single-issue** invocations (which honor `--no-pr` / `pr_policy=never`). Users who want cluster-on-feature-branch direct-commit pattern can: (a) run cluster as PR + cherry-pick or rebase to feature branch post-merge, or (b) run N atomic single-issue `--no-pr` invocations on the feature branch.
+
+If cluster-on-feature-branch direct-commit becomes a common pattern (not anticipated based on current usage), revisit Option B in a future issue. Until then, contract simplicity wins.
+
 Cross-reference: full cluster semantics in [batch-and-cluster.md](batch-and-cluster.md).
 
 ### `pr_policy` config field

plugins/issue-driven-dev/skills/idd-implement/SKILL.md

@@ -110,14 +110,30 @@ TaskCreate(name="sister_bug_sweep", description="Step 5.7: review session log +
 ```
 
 ```bash
-# 1. Parse flag + count issue args (cluster mode = ≥2 #N)
+# 1. Parse flag + count DISTINCT well-formed issue args (cluster mode = ≥2 #N)
+#
+# v2.70.0+ #100 Finding 2 — glob hardening:
+#   - Previous glob `\#[0-9]*` matched any `#<digit><anything>` including
+#     malformed tokens like `#42abc` (the `*` absorbed the trailing letters)
+#   - Duplicate args like `#34 #34` over-counted as 2, tripping CLUSTER_MODE
+#     on a single distinct issue
+#   - Strict regex `^#[0-9]+$` (matching batch-and-cluster.md documented form)
+#     + associative-array dedup closes both issues
 PR_FLAG=""  # "pr" / "no-pr" / ""
+declare -A SEEN_ISSUES=()
 ISSUE_COUNT=0
 for arg in "$@"; do
     case "$arg" in
         --pr)     PR_FLAG="pr" ;;
         --no-pr)  PR_FLAG="no-pr" ;;
-        \#[0-9]*) ISSUE_COUNT=$((ISSUE_COUNT + 1)) ;;
+        \#*)
+            # Strict integer check + dedup (v2.70.0+ #100)
+            arg_num="${arg#\#}"
+            if [[ "$arg_num" =~ ^[0-9]+$ ]] && [ -z "${SEEN_ISSUES[$arg_num]:-}" ]; then
+                SEEN_ISSUES[$arg_num]=1
+                ISSUE_COUNT=$((ISSUE_COUNT + 1))
+            fi
+            ;;
     esac
 done
 CLUSTER_MODE="false"