[bug] idd-issue Step 1 does not immediately persist Claude Code image-cache → cache evicted before Step 4 upload (pasted-image source)

[kiki830621]

Problem

Incident verbatim (2026-05-20 session, downstream caller kiki830621/ai_martech_global_scripts):
User pasted a screenshot into a /idd-issue invocation. The prompt contained a [Image: source: /Users/che/.claude/image-cache/<session-id>/1.png] annotation. I read the path in Step 1, did Step 2 AskUserQuestion (priority + clarification), Step 3 created the issue (#788 in caller repo), then in Step 4 tried to cp and gh release upload the screenshot — and the entire ~/.claude/image-cache/ was empty. Cache evicted between turns. The data-preservation hard rule was then unfulfillable without asking the user to re-provide.

The current idd-issue SKILL.md Step 1 / Step 4 split assumes the file at [Image: source: <path>] persists from Step 1 (read source) through Step 4 (upload to release). For Claude Code's ~/.claude/image-cache/<session-id>/ this assumption is fragile:

• The cache directory is per-session (path includes session id).
• It can be cleared by context compaction, session lifecycle housekeeping, or session-id rollover (e.g., when a session is continued from a previous compacted session and the new agent loop runs under a fresh id).
• Skill flow has at least one AskUserQuestion turn + several Bash calls between Step 1 and Step 4 — easily long enough for the cache to be evicted by Claude Code internals.

When this hits, the user sees: skill happily created the issue body referencing an attachment in Step 1, then later Step 4 fails to upload because ls of the staged path returns nothing. The fallback is "ask user to re-provide", which violates the spirit of Step 1's data-preservation hard rule (the user's intent was "attach this image", not "attach if cache happens to survive").

Type

bug — Step 1 → Step 4 lifecycle gap for the "pasted image" source type

Expected

When the source is a pasted image (Claude Code [Image: source: <path>] annotation), Step 1 immediately persists the image to a stable staging path (/tmp/idd-issue-attachments/issue_pending_<n>.png or similar) — within the same tool turn that read the annotation. Step 4's gh release upload then references the staged path, not the image-cache path.

In other words: Step 1's "read source" already includes "persist to /tmp", same way docx Step 1 already does export_image(image_id, output_path) instead of relying on the docx file staying mounted in some volatile location.

Actual

Step 1 just notes the image-cache path in conversation and moves on. Step 4's gh release upload <image-cache path> is the first time the file is touched after Step 1, by which point the cache may be gone. The "Pasted image" / "Mixed (text + pasted images)" rows of the Source Type Adapter table in SKILL.md don't mention immediate persistence; they only describe the upload at Step 4.

Suggested fix (incremental, low-risk)

In idd-issue/SKILL.md Step 1 → Source Type Adapter table:

1. Either replace, or add a new row, for the pasted image source type with an explicit cp [image-cache path] /tmp/idd-issue-attachments/issue_pending_<idx>.png (or equivalent Read-then-Write via tools, since Bash cp is the obvious choice) — done in the same tool turn that first sees the [Image: source: ...] annotation, NOT deferred to Step 4.
2. The "Mixed (text + pasted images)" row gets the same treatment — every pasted image gets cp'd to staging immediately.
3. Optional: add an explicit Bash snippet to the SKILL.md (similar to the Telegram fallback template at the bottom of Step 1) showing:

   mkdir -p /tmp/idd-issue-attachments
   for src in "${PASTED_IMAGES[@]}"; do
     cp "$src" "/tmp/idd-issue-attachments/issue_pending_$(date +%s)_$RANDOM.png"
   done

4. Step 4's gh release upload then reads from /tmp/idd-issue-attachments/, not from ~/.claude/image-cache/.

The fallback (image-cache already gone before Step 1 even runs — rare, but possible if there's a long pre-Step-1 turn for some reason) keeps the existing "ask user to re-provide / re-paste / give path" template, just like Telegram MCP no-download fallback.

Impact

• Severity: P3 — the fallback (ask user) works, just costs a re-paste turn and breaks the "every attachment is preserved without nag" feel the data-preservation hard rule was reaching for.
• Frequency: Anytime a session spans multiple turns between Step 1 and Step 4 (which is almost always: AskUserQuestion in Step 2 + Step 2.5/2.6 mention resolve + Step 3 gh issue create + Step 4 upload). The compaction-resumed-session scenario (this incident) is the highest-risk variant.
• Workaround: Eyeball — the user re-pastes. But that's exactly what Step 1's hard rule was supposed to make unnecessary.

Repro outline (for someone with the plugin source repo open)

1. Run /idd-issue with a pasted screenshot from a Claude Code session whose ~/.claude/image-cache/ is fresh.
2. Let Step 1 read the [Image: source: ~/.claude/image-cache/<session-id>/1.png] annotation.
3. Run AskUserQuestion (e.g., priority clarification) — answer it.
4. Run a couple more Bash tool calls (anything).
5. In Step 4, try cp or gh release upload from the original image-cache path.

In a compaction-resumed session (/compact ran since the image was pasted) the cache directory is gone before Step 4 even tries.

Source

In-session incident 2026-05-20 during /idd-issue invocation from kiki830621/ai_martech_global_scripts l4_enterprise workspace. Caller-side issue (the BrandEdge zig-zag bug that triggered this invocation) is kiki830621/ai_martech_global_scripts#788, currently paused pending re-upload of the same screenshot — the upload step is exactly the place where this bug surfaces.

Not cross-referencing #788 directly here (different repo, the cross-repo link wouldn't auto-render anyway); the link is in this body for plugin-author context.

---

Current Status

• Phase: closed
• Complexity: Plan
• Diagnosed at: 2026-05-20

[kiki830621]

Diagnosis

Type

bug (Step 1 → Step 4 lifecycle gap for pasted-image source type — cache eviction between turns)

Root Cause / Analysis

idd-issue Step 1 / Step 4 split assumes the file at [Image: source: <path>] persists from Step 1 (read source) through Step 4 (upload to release). For Claude Code's ~/.claude/image-cache/<session-id>/ this assumption is fragile:

• The cache directory is per-session (path includes session id)
• Context compaction / session lifecycle housekeeping / session-id rollover (continued session under fresh id) can clear it
• Skill flow has at least one AskUserQuestion (Step 2 priority) + Step 2.5/2.6 mention resolve + Step 3 gh issue create + Step 4 upload between read and use — easily long enough for cache eviction

When this hits, Step 4 gh release upload <image-cache-path> fails because ls of the staged path returns nothing. Fallback is "ask user to re-provide" — violates the spirit of Step 1's data-preservation hard rule.

Incident verbatim (2026-05-20 session, downstream caller kiki830621/ai_martech_global_scripts):

User pasted a screenshot into a /idd-issue invocation. The prompt contained a [Image: source: /Users/che/.claude/image-cache/<session-id>/1.png] annotation. I read the path in Step 1, did Step 2 AskUserQuestion (priority + clarification), Step 3 created the issue (#788 in caller repo), then in Step 4 tried to cp and gh release upload the screenshot — and the entire ~/.claude/image-cache/ was empty. Cache evicted between turns.

Impact

• plugins/issue-driven-dev/skills/idd-issue/SKILL.md Step 1 Source Type Adapter table — specifically the "Pasted image" + "Mixed (text + pasted images)" rows
• 0 backward compat impact for sources that already self-persist (docx via export_image, file paths in stable locations) — change is purely additive for the pasted-image path
• Reduces re-paste friction for the common multi-turn invocation pattern

Strategy

• Locate Step 1 Source Type Adapter table in idd-issue/SKILL.md (around line 290-352 per existing SKILL.md structure)
• Update "Pasted image" row to include explicit "persist immediately" semantics:
  • In the same tool turn that first sees [Image: source: ...] annotation: mkdir -p /tmp/idd-issue-attachments + cp <image-cache-path> /tmp/idd-issue-attachments/issue_pending_$(date +%s)_$RANDOM.png
  • Track staged paths in an array PASTED_IMAGES_STAGED=() for Step 4 to reference
• Update "Mixed (text + pasted images)" row with same treatment
• Add explicit Bash snippet showing the staging pattern (mirrors the existing Telegram fallback template at bottom of Step 1)
• Update Step 4 gh release upload to read from /tmp/idd-issue-attachments/ staged paths, not from raw ~/.claude/image-cache/
• Cleanup policy: /tmp/idd-issue-attachments/issue_pending_*.png files staying around indefinitely is OK (cron-cleaned by /tmp housekeeping); optionally add Step 5 cleanup rm -f /tmp/idd-issue-attachments/issue_pending_*.png after successful upload
• Document the eviction failure mode in the SKILL.md prose so future debuggers don't reinvent the diagnosis

Complexity

Plan (multi-touch in same skill + decision-heavy on staging path / cleanup / failure handling)

Layer P triggers:

• Decision-heavy with 2+ valid approaches — staging path could be:
  • /tmp/idd-issue-attachments/issue_pending_<timestamp>_<rand>.png (POSIX-safe, anonymous)
  • /tmp/idd-issue-attachments/<issue-number-pending>.png (issue-tied, but issue number unknown at Step 1)
  • .claude/.idd/issue-pending/<id>.png (in-repo, version-controlled by .gitignore)
  • Each has different cleanup ergonomics + survival across compactions
• 2+ files have sequential dependency — Step 1 staging + Step 4 upload must agree on staged path naming convention
• Risk-sensitive boundary — failure modes during staging (disk full / cache already gone before Step 1 / Bash cp permission denied) need fallback handling matching existing Telegram MCP no-download fallback pattern

Layer 2 check: skill internals — no published API change. Not Spectra.

Risks

• /tmp/idd-issue-attachments/ may accumulate stale files (mitigated by /tmp housekeeping on most OSes)
• Cache eviction can hit between paste-time and first tool turn (rare — issue body says "highest-risk variant is compaction-resumed session"); fallback to ask-user-to-re-provide stays
• macOS cp preserves permissions / xattrs; doesn't preserve in-cache atomic-write semantics — but the cache is already user-readable, so this isn't a security concern

Residue

Staging path naming convention is the central decision point for Plan tier — anonymous /tmp vs in-repo .claude/.idd/issue-pending/ has materially different ergonomics (cleanup automation vs version-control visibility). Default recommendation per lead-minimal is /tmp/idd-issue-attachments/issue_pending_<timestamp>_<rand>.png — POSIX-safe, anonymous, doesn't pollute the repo, relies on system cleanup. But user may prefer in-repo staging for audit-trail symmetry with .claude/.idd/attachments/ post-create staging convention.

Vagueness Pre-check

• V1: 2 — Strategy bullets enumerate concrete files + ordered investigation/implementation steps
• V4: 2 — Issue body's "## Expected" / "## Suggested fix" gives explicit completion criteria
• Triggered: no — both axes ≤ 3

[kiki830621]

Implementation Complete

Checklist

• Strategy bullets shipped
• Commit: 0351d57 (cluster branch idd/chain-multi-4afc6a91-refactor-spectra-archive-skill-md-exists)
• PR opened: chain (multi-root): 3 issues — spectra-archive sync + cluster glob + pasted-image cache #115 (cluster, all 3 sister issues)
• CHANGELOG v2.70.0 entry + plugin.json version bump

Scope shipped

NEW Pasted image row in skills/idd-issue/SKILL.md Step 1 Source Type Adapter + NEW Pasted-image immediate-persistence subsection with bash snippet showing cp loop + staged-path tracking + Step 4 reference contract. Staging path /tmp/idd-issue-attachments/issue_pending__.png (anonymous, POSIX-safe, system-cleanup-friendly). Updated Mixed (text + 圖片貼上) row to reference the immediate-persistence rule.

Notes

• Per cluster discipline: per-issue verify findings will be partitioned in the cluster verify report at PR chain (multi-root): 3 issues — spectra-archive sync + cluster glob + pasted-image cache #115

[kiki830621]

Verify (PR mode, cluster) — verify-gated PASS for this issue's portion of PR #115. 0 blocking findings post-fix; 3-AI + Codex ensemble PASS (requirements process gap recorded — coverage substituted via Codex + DA). Cluster ready to merge → /idd-close per issue.

[kiki830621]

Closing Summary

Problem

idd-issue Step 1 / Step 4 split assumed the file at [Image: source: <path>] persists from Step 1 (read source) through Step 4 (upload to release). For Claude Code's ~/.claude/image-cache/<session-id>/ this assumption is fragile:

• Per-session cache (path includes session id)
• Cleared by context compaction / session lifecycle housekeeping / session-id rollover

Step 1 → Step 4 flow spans AskUserQuestion + Step 2.5/2.6 mention resolve + 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: Step 4 gh release upload failed because ~/.claude/image-cache/ was empty; user had to re-paste, violating Step 1's data-preservation hard rule.

Root Cause

Step 1 only noted the image-cache path in conversation and moved on. Step 4's gh release upload <image-cache path> was the first time the file was touched after Step 1, by which point the cache could be gone.

Solution

Commit 0351d57 + 8b15784 + d63575a (PR #115):

skills/idd-issue/SKILL.md Step 1 Source Type Adapter:

1. NEW "Pasted image" row in the adapter table — explicit immediate-persistence rule
2. Updated "Mixed (text + 圖片貼上)" row — references the immediate-persistence rule for every pasted image
3. NEW ### Pasted-image immediate-persistence (v2.70.0+, #112) subsection between Source Type Adapter and MCP pre-flight:

# Run in the SAME tool turn that first sees the [Image: source: ...] annotation.
mkdir -p /tmp/idd-issue-attachments

declare -a PASTED_IMAGES_STAGED=()
for src_path in "${PASTED_IMAGE_PATHS[@]}"; do
  if [ ! -f "$src_path" ]; then
    # Cache already evicted before Step 1 ran — fallback to user re-paste
    echo "⚠ Pasted-image source $src_path no longer exists (cache evicted)." >&2
    continue
  fi
  # mktemp guarantees uniqueness atomically (collision-safe per #112-COLLISION
  # logic finding — 0.4% collision rate at 1000 tight-loop trials with $RANDOM)
  staged_base=$(mktemp /tmp/idd-issue-attachments/issue_pending_XXXXXX) || {
    echo "✗ mktemp failed — skipping $src_path" >&2
    continue
  }
  staged_path="${staged_base}.png"
  mv "$staged_base" "$staged_path"
  cp "$src_path" "$staged_path"
  PASTED_IMAGES_STAGED+=("$staged_path")
done

NEW prose paragraphs explicitly document:

• PASTED_IMAGE_PATHS source contract (DA-4 qualified): agent populates the array from [Image: source: ...] annotations; the annotation is currently the only authoritative source per v2.70.0+ Claude Code prompt format
• Step 4 hand-off contract: ATTACHMENT_PATHS+=("${PASTED_IMAGES_STAGED[@]}") concatenation in Step 4's upload loop

Verify-fix commit 8b15784 dispatched 3 MEDIUM/LOW logic findings:

• Collision-safe via mktemp (logic finding [bug] idd-issue Step 1 does not immediately persist Claude Code image-cache → cache evicted before Step 4 upload (pasted-image source) #112-COLLISION)
• Explicit PASTED_IMAGE_PATHS contract (logic finding [bug] idd-issue Step 1 does not immediately persist Claude Code image-cache → cache evicted before Step 4 upload (pasted-image source) #112-CONTRACT)
• Explicit Step 4 hand-off (logic finding [bug] idd-issue Step 1 does not immediately persist Claude Code image-cache → cache evicted before Step 4 upload (pasted-image source) #112-STEP4)

Verify-fix commit d63575a dispatched DA findings:

• DA-1 MEDIUM (CHANGELOG honesty — fixed under [refactor] spectra-archive SKILL.md exists in 3 copies (.claude/, .agents/, plugins/) — divergence risk (follow-up from #56) #93 scope)
• DA-2 LOW (cluster mode: feature-branch direct-commit tension + cluster-detection glob looseness (follow-up from #96) #100 backward-compat note)
• DA-4 LOW (PASTED_IMAGE_PATHS qualifier added)
• Codex nit (table row + subsection wording updated to mktemp form)

Verification

3-AI verify ensemble (PR #115 cluster mode) + Codex — verify-gated PASS post-fix.

Codex empirically verified mktemp portability across macOS BSD + GNU; 100-parallel-invocation stress test produced 100 unique files. Devil's Advocate empirically verified declare -a works on macOS bash 3.2.57 (default shell).

Changes

• Squash merge: f5afb93 (PR chain (multi-root): 3 issues — spectra-archive sync + cluster glob + pasted-image cache #115)
• plugins/issue-driven-dev/skills/idd-issue/SKILL.md — Step 1 Source Type Adapter table + NEW Pasted-image immediate-persistence subsection
• Shared CHANGELOG.md v2.70.0 + plugin.json 2.69.0 → 2.70.0

Residue

DA-3 LOW (staged_base.png stale-overwrite) — documented as out-of-scope per IDD threat model. mktemp guarantees fresh uniqueness; stale-overwrite only possible under hostile pre-creation by attacker, and there is no user-controlled write path to /tmp/idd-issue-attachments/. If a future attacker model surfaces this, add cp -n (noclobber) flag.

Future Claude Code prompt format extension — current PASTED_IMAGE_PATHS contract assumes [Image: source: ...] annotation is the only authoritative source. If future Claude Code versions add drag-drop or other attachment annotation forms, this contract may need extension. Qualifier added to subsection prose (DA-4 fix).