task-done: parse .info file via awk instead of source (defends against opaque non-gh identifiers)

[martin-conur]

Context

Follow-up from PR #145 (closes #144) — reviewer's clean-with-nits finding.

After #144 lands, non-gh loadouts (claude-jira / claude-notion / claude-local) can pass any opaque string as the spec identifier. task-reviewer writes it unquoted into the worktree's .info file:

printf 'ISSUE_NUMBER=%s\n' "$ISSUE_NUMBER"

Producing lines like:

ISSUE_NUMBER=PROJ-$(date)

task-done then reads the file via source "$INFO_FILE" (claude-gh/bin/task-done:54). The shell re-parses the value — $(...) executes, $VAR expands, backticks execute.

Practical risk: zero

• Real Jira keys: PROJ-123 — no metacharacters.
• Real Notion URLs: https://notion.so/...?p=... — %-encoded special chars.
• Real local slugs: kebab-case-string — no metacharacters.

The vector requires explicitly crafted identifiers and would be a self-inflicted local operation. Filed for completeness and code-hygiene, not because exploitation is realistic.

Why fix anyway

• source of user-influenced data is a code-smell that ages badly — future contributors adding fields to .info may not realize the value gets eval'd.
• The fix is small and matches an existing pattern in the same file.
• After Radio session corruption: unregister cascade → re-seed loses TAB_ID + LOADOUT (breaks --auto wake-up) #140 (radio session corruption) and task-done --remove-worktree leaks reviewer branch + state; blocks re-dispatch of task-reviewer #148 (task-done branch leak), task-done's cleanup path is getting attention anyway — bundling the parse-hardening keeps task-done polished.

Proposed fix

Two options:

Option A — quote on write (defensive at source):

- printf 'ISSUE_NUMBER=%s\n' "$ISSUE_NUMBER"
+ printf 'ISSUE_NUMBER=%q\n' "$ISSUE_NUMBER"

…in task-reviewer. task-done's source then sees a quoted assignment that bash parses safely.

Option B — replace source with awk parser in task-done (defensive at read):

task-done already uses an awk-based read for TAB_ID (claude-gh/bin/task-done:116, 121):

_WORKER_TAB_ID=$(awk -F= '/^TAB_ID=/ { print $2; exit }' "$INFO_FILE")

Extend the same pattern to every field task-done currently relies on from $INFO_FILE:

• BASE_BRANCH
• SLUG
• GH_URL
• PR_NUMBER (post-task-done --remove-worktree leaks reviewer branch + state; blocks re-dispatch of task-reviewer #148)
• ISSUE_NUMBER

Drop the source "$INFO_FILE" line. Each field gets a parser-side awk read; nothing in the file is ever eval'd.

Recommended: Option B. Defends regardless of write-side discipline (a future loadout adding a new field can't accidentally introduce the vector). Mirrors the existing TAB_ID pattern.

Files (expected touchpoints)

File	Action
claude-gh/bin/task-done (+ all 4 claude + 3 kiro variants in drift groups task-done-std / task-done-local)	MODIFY — replace source with field-wise awk reads
tests/task_done.bats	EXTEND — add a test that passes a .info containing ISSUE_NUMBER=PROJ-$(date) and asserts task-done completes without $(date) evaluating
claude-gh/bin/task-reviewer (and 3 sibling drift members)	(Option A only) — write printf %q quoted value

tools/check-drift.sh should stay green — the changes are inside existing confirm-and-cleanup / worktree-context regions.

Verification

1. bats tests/task_done.bats — green, including the new injection-attempt test.
2. tools/check-drift.sh — green.
3. End-to-end: PM dispatches task-reviewer 42 'PROJ-$(date)' in a claude-jira scratch project; later task-done --remove-worktree in that reviewer's worktree completes without $(date) being substituted.

Why P2 / no milestone

• Zero practical risk for real-world identifier formats.
• Adjacent to Radio session corruption: unregister cascade → re-seed loses TAB_ID + LOADOUT (breaks --auto wake-up) #140 / task-done --remove-worktree leaks reviewer branch + state; blocks re-dispatch of task-reviewer #148 (also task-done hardening); reasonable to bundle in a task-done polish pass after v0.3.0 ships.

References

• PR task-reviewer: tracker-aware spec-issue cross-check (claude-jira/notion/local) (#144) #145 reviewer round-5 finding (clean-with-nits)
• claude-gh/bin/task-done:54 — the source line
• claude-gh/bin/task-done:116,121 — the existing awk -F= pattern to mirror
• claude-gh/bin/task-reviewer — the write-side
• Radio session corruption: unregister cascade → re-seed loses TAB_ID + LOADOUT (breaks --auto wake-up) #140 / task-done --remove-worktree leaks reviewer branch + state; blocks re-dispatch of task-reviewer #148 — sibling task-done issues

[martin-conur]

Closed into #180 (grooming pass, 2026-09-27) — the work is still wanted, folded rather than dropped.

#149, #152 and #180 are all defects in resolving a tab identity or reading a .info file, across the same few call sites. #180 is now retitled to cover all three, because landing them separately means three PRs re-reading the same code — and the fixes share the identity-check pattern #167 already established.

The ordering matters and is recorded there: #152 first. A staleness guard (#180) on top of a resolver that can return the wrong id (#152) only narrows the window, since the guard verifies that an id still names the tab it named — useless if it named the wrong tab to begin with.

Also noted there: #188 and #218 made .info load-bearing in more paths than when #149 was filed, which raises the value of parsing it without source.

Track the work at #180.