Files
claude_mac/.claude/tasks/plans/2026-11-06-safety-residuals-1530.md
T

10 KiB

Bugfix plan: five safety residuals from finished phases (2026-11-06) — r3

Branch: bugfix/safety-residuals off develop (ddb04f4). One commit per item. r2 after the 3-lens challenge (correctness CONCERNS(2), robustness CONCERNS(5), simplicity CONCERNS(1); no BLOCKER). Changes vs r1 are tagged [r2: lens n].

  • BLK-043 / LRN-263 (item 1): hooks never call the claude CLI; hook-running tests stub it and cap tasks. Item 1 completes LRN-263's "future use".
  • BDR-136 (item 2): shell rules are added only once bench-proven; the bench is .claude/tasks/plans/2026-11-02-phase2-bench/.
  • LRN-074 (item 4): system awk may be mawk; no gawk-only flag.
  • BDR-045, LRN-084 (item 5): the .claude/** exemption was kept so the side-car memory commit and standalone /capitalize, /close land in place; BDR-045 rejected "narrow by size/scope" as fuzzy. Item 5 narrows by an exact path allowlist (user choice 2026-11-06), a partial revision of BDR-045, recorded as a new BDR on commit.

Safety rules for the whole run

  • Every run of a hook-running suite, of lib/gitflow-test.sh and of the bench goes through systemd-run --user --scope -q -p TasksMax=256.
  • The real claude CLI is never run, except by the bench's run.sh (claude -p --setting-sources local: no user settings, no hook).
  • The bench runs only after its own hardening (item 2 step 1) and from a cwd holding no .env (the session scratchpad).
  • Live paths: settings.json, lib/, githooks/ (~/.claude links to this tree; global core.hooksPath = ~/.claude/githooks). settings.json is written only via jq … > tmp && jq empty tmp && mv tmp settings.json. [r2: robustness 7; r3: confirm 2]
  • Rollback of an open item: git restore --staged --worktree -- <its paths> (or git stash push -- <paths>), THEN git checkout develop; the next session start in each onboarded repo re-syncs its .githooks/ from the (restored) lib. [r3: confirm 1]
  • Execution: one item at a time, in order 4, 3, 1, 2, 5 (smallest blast radius first, the machine-wide guard last); each item committed once its own suites pass, before the next starts.

1. Suites that can reach the real claude CLI

ROOT CAUSE: the LRN-263 stub is copied inline in detect-plan and session-digest and is fail-open: claude probe runs before any check that command -v claude is the stub; a failed stub write (no set -e) or a noexec TMPDIR (command -v skips a non-executable file) reaches the real CLI; a red positive control does not stop the suite. graphify-gate (:61), memory (banner() :689) and push-guard (:356) run hooks/session-start.sh with no stub. cc.test.sh (:30-38) shims claude, systemd-run, dtach with the same fail-open shape while lib/cc.sh runs systemd-run … dtach … exec claude [r2: robustness 3]. A relative work dir passes a same-cwd check but stops resolving after the cd every hook run does [r2: robustness 4].

FIX:

  • New lib/tests/claude-stub.sh (sourced):
    • stub_first_or_die <cmd> <expected-path>: command -v <cmd> must equal the path, else print <cmd> stub not first on PATH on stderr and exit 1.
    • claude_stub_install <work-dir>: refuse empty or missing dir; make it absolute (cd -P "$1" && pwd) [r2: robustness 4]; write <dir>/stub/claude (log path expanded at creation), chmod +x, each step || exit 1; prepend <dir>/stub to PATH and export it; stub_first_or_die claude <dir>/stub/claude; positive control run as a child from another cwd: (cd / && claude probe), log must hold exactly 1 line else exit 1 [r2: robustness 4]; empty the log; export GIT_CEILING_DIRECTORIES=<dir> so no hook run walks into an enclosing repo [r2: robustness 9, simplicity 3: applied uniformly].
    • claude_stub_calls: line count of the log.
  • detect-plan, session-digest: inline stub replaced by the helper.
  • graphify-gate, memory, push-guard: helper installed before the first hook run (memory: directly above banner(), never near the top: $FIX is the fixture repo itself) [r3: confirm 6]; each suite ends with no-claude-calls = 0 before its summary.
  • cc.test.sh: after writing its shims, stub_first_or_die for claude, systemd-run and dtach against $FX/bin/*. [r2: robustness 3]
  • X=$(mktemp -d) || exit 1 at the mktemp of each suite touched here (detect-plan, session-digest, graphify-gate, memory, push-guard, cc). [r2: robustness 10; no extra [ -d ], simplicity 3]
  • detect-plan: bash -c 'source "$1"; detect_plan' _ "$LIB" (both sites).
  • Regression suite lib/tests/claude-stub.test.sh: (a) fail-closed: in a subshell, chmod() { :; } shadows chmod so the stub stays 0644; PATH also holds a second dir with a fake claude that writes a sentinel; claude_stub_install must exit 1 with claude stub not first on PATH, sentinel absent. [r2: correctness 5, simplicity 2] (b) relative dir: install from a relative dir name, then (cd / && command -v claude) = the absolute stub path. (c) happy path: rc 0, claude_stub_calls = 0 after install.

2. rtk read with a leading flag or file on a root-level .env

ROOT CAUSE: the rtk read deny twins were generated as rtk read X* and rtk read */X*. A root-level target after a flag or another file (rtk read -n .env, rtk read a.txt .env) matches neither; rtk rewrite rc 3 passes it through (60ee4a9). Targets under a */ prefix are already covered: the * right after rtk read absorbs -n sub or a.txt sub (matcher * spans spaces, bench A1, LRN-194(c)). [r2: simplicity 1, correctness 6]

FIX:

  • Step 1, harden the bench before any run [r2: robustness 5]: bench.sh W="$(mktemp -d)" || exit 1; run.sh cd "${4:?}" || exit 1, with run.sh's usage line and README.md making dir mandatory. [r3: confirm 7]
  • Step 2, bench rows (cases.tsv), run from the scratchpad, contained:
    • R3 deny Bash(rtk read * .env*), cmd rtk read -n .env → DENIED.
    • R4 deny Bash(rtk read */.env*) (the EXISTING rule), cmd rtk read -n sub/.env → DENIED (proves no * */X twin is needed).
    • R5 deny Bash(rtk read * .env*), cmd rtk read -n README.md → RAN.
    • R6 deny Bash(*/rtk read * .env*), cmd /home/bchanot/.local/bin/rtk read -n .env → DENIED. [r2: robustness 8]
    • R7 deny Bash(*/rtk read */.env*) (EXISTING), cmd /home/bchanot/.local/bin/rtk read -n sub/.env → DENIED. [r3: confirm 5] A NO_TOOL_CALL verdict, or a sent column that differs from the cases.tsv cmd, is rerun once, not counted as a rule failure. [r2: correctness 7; r3: confirm 4]
  • Step 3, only if R3, R4, R6, R7 DENIED (each with sent = its cmd), R5 RAN and every control RAN: add Bash(rtk read * .env*) and Bash(*/rtk read * .env*) to permissions.deny (.env.* is contained in .env*), pin both in push-guard's T63 list. Any other outcome → item 2 stops, reported, nothing added. If R4 is not DENIED, item 2 stops too (the */ twins then need their own bench).

3. session-start memory.sh check not time-bounded

ROOT CAUSE: hooks/session-start.sh:319 runs bash "$_mem_lib" check .claude/memory unbounded; the digest (:377) uses timeout 2 when present.

FIX: same shape as the digest (tmo=(), command -v timeout && tmo=(timeout 2), "${tmo[@]}" bash …, bash 3.2 safe as the digest is). Regression in session-digest: a memory.sh that sleeps 10 s on check → hook returns in < 3.5 s, rc 0 (mirror SD12).

4. awk reads an assignment-shaped path operand as an assignment

ROOT CAUSE: awk treats an operand matching ^[A-Za-z_][A-Za-z0-9_]*= as a variable assignment and reads stdin instead. memory.sh passes user paths as awk operands: --body-file to _mem_fence_open, _mem_body_heading and _mem_entry_text (the entry text itself is copied by awk, so today the entry gets stdin's content, not the file's) [r2: correctness 4], and the check / upgrade dir argument.

FIX: _mem_path_safe <p>: prints ./<p> only when p matches the assignment shape, else p unchanged; empty in, empty out [r2: correctness 2+3, robustness 1]. Applied to the body file in memory_add and to the check / upgrade dir AFTER the ${1:-.claude/memory} default and the usage guard. Every existing output stays byte-identical (T7f unchanged). Regression in memory.test.sh, every call </dev/null [r2: correctness 4, robustness 11]: add … --body-file x=1.md whose body holds an unfenced ## line → rc 2; add … --body-file x=1.md with a clean body → the entry holds that body; check a=b reads a=b/*.md.

5. Pre-commit .claude/** exemption too wide

ROOT CAUSE: lib/gitflow.sh emit (:490) whitelists any staged set under .claude/ or .githooks/, listed with git diff --cached --name-only, which shows only a rename's destination [r2: correctness 8, robustness 2].

FIX (user choice 2026-11-06: memory, tasks, audits, deploy; settings.json NOT exempt):

  • Emitter: exempt only when every staged path is under .claude/memory/, .claude/tasks/, .claude/audits/, .claude/deploy/ or .githooks/; list with git diff --cached --name-only --no-renames; BLOCKED hint names the allowlist.
  • Order: emitter edit → tests (lib/gitflow-test.sh, contained) green on the emitter → only then regenerate. [r3: confirm 2]
  • Regenerate both hooks in the same commit: .githooks/pre-commit (bash lib/gitflow.sh install-hook) and githooks/pre-commit (bash lib/gitflow.sh global-hooks "$PWD/githooks" '~/.claude/githooks'); drift proof T19a, T19d, G5. [r2: correctness 1, robustness 6]
  • Tests in lib/gitflow-test.sh: relabel T3; add to T14/T21 on develop: .claude/rules/x.md only → BLOCKED; .claude/settings.json only → BLOCKED; git mv .claude/rules/x.md .claude/tasks/x.md → BLOCKED; memory-only, tasks-only, deploy-only, audits-only → allowed. [r3: confirm 3] [r2: simplicity 4]
  • Citers reworded to the allowlist: CLAUDE.global.md:160, :172, :174; settings.json:606 (environment prose); lib/gitflow-aiguillage.md:35. Grep \.claude/\*\* again after the edit. [r2: simplicity 4]

RISK: medium. Item 5 changes a guard every repo runs: committing .claude/settings.json or rules on develop now needs a branch (1 such commit in develop's history, 858e7a7). Item 2 depends on the bench. Items 1, 3, 4 are local to tests and libs.