Claude Code subagents can't write to graphify-out/ (skill.md): chunks silently lost during /graphify --update #808

Summary

When /graphify --update dispatches general-purpose Claude Code subagents for semantic extraction (Step B2), each subagent successfully extracts the JSON but cannot persist it because the bundled skill.md instructs them to write to graphify-out/.graphify_chunk_NN.json — a path outside Claude Code's general-purpose subagent sandbox allowlist.

The subagent does the work, then fails with "Write tool is denied" for that path. Step B3 then can't find the chunks, the merge gets empty input, and docs/images are silently dropped from the graph.

Reproducer

1. uv tool install graphifyy (v0.6.7 confirmed)
2. graphify install --platform claude in a project
3. In Claude Code, run /graphify --update against a project with mixed code + docs/images
4. Step B2 dispatches general-purpose subagents per skill.md lines 351–368
5. Each subagent attempts Write to <project>/graphify-out/.graphify_chunk_NN.json
6. Observed: subagent reports "Write tool is denied". It returns gracefully after burning ~3 min of inference per chunk.
7. Step B3 reads graphify-out/.graphify_chunk_*.json → empty → .graphify_semantic_new.json has 0 nodes added → docs/images effectively skipped from the update.

Root cause

Claude Code's general-purpose subagent runs in a sandbox that does NOT inherit the orchestrator's filesystem allowlist. As of 2026-05 the writable paths for a typical project layout are:

• The specific app/component subdirs the user has open (e.g. apps/mobile, src/)
• <project>/.claude/
• /tmp

Notably NOT writable from a general-purpose subagent:

• <project>/graphify-out/
• The project root itself

The bundled skill.md assumes the orchestrator's permissions propagate. They don't.

Cost data point

A /graphify --update on a ~400-file Expo monorepo on 2026-05-08 lost ~16 min and ~290k tokens across two retry cycles before the issue was diagnosed by reading the subagent's error output. Both subagents reported "Write tool is denied" after completing the extraction work in memory.

Proposed fix

Route subagent chunk writes through /tmp/graphify_chunk_NN.json — universally writable across every Claude Code sandbox configuration I've tested. The orchestrator's merge step copies them into graphify-out/ before the existing glob+merge logic, so cache semantics + the rest of the pipeline are unchanged.

Four small edits to graphify/skill.md (1437 lines → 1437 lines, net):

1. Step B2 prompt template (~line 354) — replace "Output exactly this JSON" with explicit "Write to /tmp/graphify_chunk_NN.json" instruction + sandbox rationale comment.
2. Step B3 success-signal check (~line 361) — change path from graphify-out/.graphify_chunk_NN.json to /tmp/graphify_chunk_NN.json.
3. Step B3 merge step (~line 374) — insert 4 lines that shutil.copy /tmp/graphify_chunk_*.json → graphify-out/ before the existing glob (200+ lines unchanged).
4. Step 9 cleanup (~line 838) — also rm -f /tmp/graphify_chunk_*.json.

PR coming next with the exact diff.

Tradeoffs of /tmp

• Universally writable across every Claude Code sandbox I've seen
• Auto-cleared on reboot (cleanup-friendly)
• Theoretical concurrency collision: two /graphify runs from different projects could clash on filenames. Mitigation: subdirectory-per-run if it ever bites in practice (it hasn't for me across ~30 update runs).
• Doesn't survive reboot mid-run. Acceptable — /graphify --update completes in <10 min.

Why not project-level additionalDirectories

Per-project busywork that doesn't help users running graphify across multiple projects, and would silently regress for any user who hasn't applied it. The /tmp path is portable and works everywhere by default.

Other affected platforms?

The bundled package has 12 platform-specific skill-*.md files. Only skill.md (Claude Code default) is confirmed affected by me. The fix likely applies to other platforms with similar subagent sandbox models — needs verification per-platform. I can scope this PR to skill.md only and let other contributors validate the parallel files.