feat(rules): shadow-star-shorthand-autocomplete-marker cold-boot rule by AceHack · Pull Request #3339 · Lucent-Financial-Group/Zeta

AceHack · 2026-05-15T01:11:24Z

Summary

Per Aaron 2026-05-15T~01:09Z explicit framing: "future version of you should likely remember all this."

Promotes the "(shadow*)" shorthand definition from user-scope memory to `.claude/rules/` so it's cold-boot auto-loaded for every Otto-CLI session.

Why

Otto-CLI committed 4+ "(shadow*)" misinterpretations in this single session before the maintainer named the failure mode. The pattern: encounter shorthand → invent interpretation ("Shadow-lock posture per your framing") → maintainer corrects → repeat. Without cold-boot auto-load, future-Otto would repeat the cycle indefinitely.

With this rule, future-Otto cold-booting reads the canonical definition at session start:

"(shadow*)" = autocomplete-generated text the maintainer accepted; instruction stands; never invent "shadow-X-posture" framings on top.

Composes with

`.claude/rules/shadow-check-name-acceptance.md` (different shadow-scope: name-acceptance methodology)
`.claude/rules/wake-time-substrate.md` (load-bearing substrate needs wake-time landing)
`.claude/rules/razor-discipline.md` (operational claim only)
`.claude/rules/glass-halo-bidirectional.md` (text-source transparency)
`tools/shadow/` (different scope: shadow infrastructure detects + auto-accepts grey-text in the maintainer's UI; this rule is about how Otto-CLI handles the marker when the maintainer ferries autocomplete-completed text to Otto)

Full reasoning + double-failure log

Preserved at user-scope memory (not in this PR but referenced):

`feedback_aaron_shadow_star_shorthand_means_autocomplete_generated_not_aaron_authored_grey_text_completed_2026_05_15.md`
`feedback_aaron_shadow_observation_loop_design_pattern_otto_observes_directly_instead_of_assuming_aaron_plays_games_2026_05_15.md` (companion failure mode: assumption-driven framing of existing infrastructure as "backlog candidate, not yet built")

🤖 Generated with Claude Code

Per Aaron 2026-05-15T~01:09Z: 'future version of you should likely remember all this.' Promotes the (shadow*) shorthand definition from user-scope memory to .claude/rules/ so it's cold-boot auto-loaded. Otto-CLI committed 4+ '(shadow*)' misinterpretations in this single session before the maintainer named the failure mode directly. Without cold-boot auto-load, future-Otto would repeat the cycle (encounter shorthand → invent interpretation → maintainer corrects → repeat). With this rule, future-Otto cold-booting reads the canonical definition at session start: '(shadow*)' = autocomplete-generated text the maintainer accepted; instruction stands; never invent 'shadow-X-posture' framings on top. Composes with .claude/rules/shadow-check-name-acceptance.md (different shadow-scope: name-acceptance methodology); wake-time-substrate.md (load-bearing substrate needs wake-time landing); razor-discipline.md (operational claim only); glass-halo-bidirectional.md (text-source transparency). Co-Authored-By: Claude <noreply@anthropic.com>

Copilot

Pull request overview

Adds an always-on Claude Code rule defining the (shadow*) shorthand so future sessions interpret it as an autocomplete-source disclosure marker rather than a “shadow posture” instruction.

Changes:

Adds a new .claude/rules/ cold-boot rule for (shadow*).
Documents operational meaning, non-meanings, related rules, and provenance references.

AceHack

Drift detected (Catch 36): Narration-over-action and metadata churn. This PR substitutes complex operational framing ('Full reasoning + double-failure log') for simple file actions. Eliminate metadata churn.

The maintainer's quoted correction used `_..._` (underscore-emphasis) where the file's MD049 setting requires asterisk-emphasis. Single-line fix; verbatim quote text untouched. Verified: `bunx markdownlint-cli2 <file>` exits clean. Co-Authored-By: Claude <noreply@anthropic.com>

Three findings resolved: - Line 19: grammar — "originated authored prose" → "original authored prose" - Line 22: use "the human maintainer" instead of first-name attribution on this instruction surface, per the convention Copilot flagged (first-name attribution reserved for history/memory surfaces) - Line 78-81: reachable-from-cold-boot disclosure for the user-scope memory references. Acknowledges they live at ~/.claude/projects/<slug>/memory/MEMORY.md (auto-loaded each Otto-CLI session) rather than in-repo, and adds a PR backlink as the in-repo provenance anchor. The memory files themselves are out-of-scope for this PR — landing them in-repo is a separate piece of work (likely the Memory Preservation Guarantee from Manifesto V2; see CURRENT-aaron.md). Verified: `bunx markdownlint-cli2 <file>` clean. Co-Authored-By: Claude <noreply@anthropic.com>

Copilot

Pull request overview

Copilot reviewed 1 out of 1 changed files in this pull request and generated 1 comment.

Comments suppressed due to low confidence (2)

.claude/rules/shadow-star-shorthand-autocomplete-marker.md:78

This overstates the cold-boot guarantee. The loading taxonomy says user-scope MEMORY.md is capped to the first 200 lines / 25KB plus explicit reads (.claude/rules/claude-code-loading-taxonomy.md:13-19), so the underlying user-scope memos are not guaranteed to auto-load or be reachable from every cold boot.

Full reasoning + double-failure log (the same conversation also failed on "shadow observation loop is backlog candidate, not yet built" when [`tools/shadow/`](../../tools/shadow/) + B-0402 implementation already existed) is preserved at user-scope memory. These memos auto-load into every Otto-CLI session via `~/.claude/projects/<slug>/memory/MEMORY.md`, so they are reachable from cold boot even though they do not live in-repo:

.claude/rules/shadow-star-shorthand-autocomplete-marker.md:81

These bare backticked .md basenames will be harvested by tools/hygiene/audit-rule-cross-refs.ts as repo-local path references and reported as stale, because the audit only resolves direct repo paths/globs and backlog IDs. Since the text explicitly says these files live out of repo, cite them in a form the audit will not classify as broken repo paths.

- `feedback_aaron_shadow_star_shorthand_means_autocomplete_generated_not_aaron_authored_grey_text_completed_2026_05_15.md`
- `feedback_aaron_shadow_observation_loop_design_pattern_otto_observes_directly_instead_of_assuming_aaron_plays_games_2026_05_15.md`

Five Copilot / Codex review findings resolved: P1 (Copilot) — broken provenance in README header: Dropped first-name attribution + user-scope memory-file glob that doesn't resolve in-repo. Per docs/AGENT-BEST-PRACTICES.md §671-685, current-state surfaces (this README) use role-refs. Other rule references now use relative links that resolve from the README's location. P1 (Copilot) — sed escaping in install instructions: The documented `sed` recipe would break on paths containing `&`, `|`, `\`, or newlines. Replaced with a TypeScript installer (tools/shadow/launchd/install-launchagent.ts) that: - uses String.replaceAll (metacharacter-safe) - sanity-checks for unsubstituted placeholders - runs plutil -lint before writing - backs up any existing plist before overwriting - optionally bootstraps via launchctl Placeholder syntax changed from `__NAME__` to `{{NAME}}` so the template header comment can describe them without being substituted itself. P1 (Copilot) — rule file referenced from plist not in tree: The .claude/rules/shadow-star-shorthand-autocomplete-marker.md reference resolves on main now (#3339 merged). The reference is preserved in the plist's composes-with comment block. P2 (Codex) — KeepAlive semantics in plist: Dropped `SuccessfulExit=false`. KeepAlive dict keys are OR-combined, so the original combination relaunched on any non-zero exit (incl. config errors). `Crashed=true` alone gives the documented crash-only behavior. Comment added explaining the choice. P2 (Codex) — legacy launchctl load/unload in README: Replaced with the modern `bootstrap`/`bootout` interface (with `gui/$(id -u)` domain target) in install, reload, and uninstall steps. The legacy interface is documented as deprecated and can silently no-op on failure; `bootstrap`/`bootout` return proper exit codes. Verified: - bunx tsc --noEmit clean - markdownlint-cli2 clean on README - plutil -lint OK on both the template and a rendered output - Installer dry-run produces a valid plist Co-Authored-By: Claude <noreply@anthropic.com>

…ose-out (#3356) * shard(tick): 0230Z — multi-Otto contamination caught + PR #3339/#3349 close-out Co-Authored-By: Claude <noreply@anthropic.com> * fix(shard-0230Z): correct relative-link depth — 4x ../ → 5x and 5x → 6x Both Copilot and Codex flagged that the tick shard's relative links were off by one directory level. From the shard's location (docs/hygiene-history/ticks/2026/05/15/0230Z.md): - `../../../../backlog/...` resolved under docs/hygiene-history/backlog (4 levels up only reaches docs/hygiene-history/, not docs/) - `../../../../../.claude/rules/...` similarly off by one for top-level Correct depths from this shard location: - docs/backlog/...: 5x `../` (verified with `ls -la` from the shard dir) - .claude/rules/...: 6x `../` (verified with `ls -la` from the shard dir) Both links now resolve. Composes with the prior `5x .. → 6x ..` fix on the 0027Z shard (#3330) — same class of error, same family of fix. Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>

Codex + Copilot reviewed the new install-launchagent.ts in detail. This commit addresses the substantive correctness + style findings; unit-test coverage is deferred to a separate backlog row. Addressed (8): P1 [Codex P2 + Copilot P1] — XML-escape substituted values: Paths like `/Users/foo & bar/Zeta` would substitute into <string> text nodes producing an invalid plist (& without entity terminator). New `xmlEscape` helper applies the five XML 1.0 predefined entities (& < > " ') before placeholder substitution. Restores the advertised "handles paths with shell metacharacters" promise — the bug class was XML escaping, not shell escaping. P1 — Unknown args silently ignored: Switch fell through on typos like `--dryrun` or `--bootstap`, then installed with default behavior. Added explicit default: case that errors with the offending flag and exits 1 before any FS action. Verified: `--dryrun` now produces "Unknown argument: --dryrun. Run with --help to see supported flags." P1 — Relative --bun-path / --repo-root overrides: `--repo-root .` would write relative paths into ProgramArguments, WorkingDirectory, and log paths — broken under launchd (which has no cwd unless explicit). New requireAbsolute() helper rejects with a `Use \`<canonicalized>\` if that's what you meant` hint. P1 — Non-atomic backup-then-write: Previous renameSync(destPath, backup) + writeFileSync(destPath) had a window where the agent could be removed but the new plist not yet present (if writeFileSync threw). Replaced with the standard write-temp-then-promote pattern: 1. writeFileSync(destPath.tmp-PID, content) 2. renameSync(destPath, backup) // may not exist; that's OK 3. renameSync(destPath.tmp-PID, destPath) // atomic promote Failure in step 3 restores backup → destPath (best-effort) so the caller's existing install is preserved. P1 — execFileSync can throw before validation: Missing bun, missing git, or running outside a git checkout produced a raw exception. New tryDetect() catches and returns undefined; parseArgs emits the documented actionable error pointing at the --bun-path / --repo-root override. P1 — Unconditional main() invocation: Per Bun-project convention, gate main() with `if (import.meta.main)` so tests / helpers can import the module without triggering filesystem or launchctl side effects. P1 — README/plist cross-refs to .claude/rules/shadow-star-shorthand-*: These resolve on main (PR #3339 merged). CI's merge-ref check confirms; the reviewers' findings were against the raw branch base. No additional action needed. P2 — Direct name attribution in plist KeepAlive comment: Replaced "Per Codex review on PR #3342" with the rationale text only. Current-state surface; uses role-refs only. Deferred (1): P1 — Unit test coverage for install-launchagent.ts: Codex + Copilot called out missing tests for placeholder escaping, missing placeholders, argument validation, and dry-run output. This commit adds the substantive behavior the tests would verify; the tests themselves are deferred to a separate backlog row to keep this PR's scope manageable. Smoke-tested all four behaviors by hand on the cli before commit. Filing as follow-up: tests in `tools/shadow/launchd/install-launchagent.test.ts` using Bun's built-in test runner. Verified: - bunx tsc --noEmit clean on install-launchagent.ts (pre-existing process-extract.ts errors are on this branch base because #3349 isn't rebased in; CI's merge-ref includes the fix) - semgrep --config .semgrep.yml --error: 0 findings - --dry-run produces valid plist (plutil -lint OK) - --dryrun rejected (Unknown argument) - --repo-root . rejected (must be absolute) - Special-char paths (e.g., /Users/foo & bar) now produce valid escaped XML Co-Authored-By: Claude <noreply@anthropic.com>

P1 — Restore SuccessfulExit=false in KeepAlive (reviewer-conflict resolution): Earlier cycle dropped this on Codex review (config errors would loop-restart). This cycle Copilot countered: Bun TypeScript runtime errors exit non-zero rather than signal-terminate, so the previous `Crashed=true` only setting left those failure modes unhandled. Restored both keys with an extended plist comment explaining the trade-off. ThrottleInterval=60s + --dry-run-by-default already cover the "config error restart loop" concern. P1 — Add sonarjs/no-os-command-from-path suppressions: Matches the convention in tools/github/poll-pr-gate.ts + tools/peer-call/. Applied at tryDetect, plutilLint, and the three launchctl bootstrap/bootout calls. Comments cite the precedent. P2 — Export helper functions for testability: `import.meta.main` guard was added in `970c774` but the functions were not exported. Added `export` to main(), parseArgs(), requireAbsolute(), tryDetect(), xmlEscape(), substitutePlaceholders(), plutilLint(), and the Args interface. Unblocks the B-0528 test file. P1 — Update 0305Z tick shard's "PREFER" code snippet: The earlier shard documented the rename-then-write pattern as "PREFER", but that pattern still has a window where destPath is empty. Updated to show 3-pattern progression (AVOID copyFileSync → ALSO AVOID rename-then-write → PREFER write-temp-then-atomic- replace) so the historical trail is accurate and prevents future authors from copying the deprecated pattern. Deferred (1): P1 — Cross-ref findings on README.md:5 + plist:24 + ts:185: README.md:5 and plist:24 reference `.claude/rules/shadow-star-shorthand-autocomplete-marker.md` which IS on main (#3339 merged). The reviewer's check is against the raw branch base; CI's merge-ref check passes. ts:185 is the test coverage finding, already filed as B-0528 (deferred per scope management). Verified: - bunx tsc --noEmit clean on install-launchagent.ts - markdownlint clean on 0305Z.md - plutil -lint OK on template + rendered output - --dry-run produces valid plist Co-Authored-By: Claude <noreply@anthropic.com>

…shed + new worktree-pruning-race failure mode (#3359) * shard(tick): 0414Z — PR #3339/#3349 merged + B-0527 collision republished + new worktree-pruning-race failure mode Five side-worktree attempts (incl. --lock'd) got rm -rf'd mid-tick. Recovery via age-exempt sibling worktree (0027z-sidetick) via branch-switch. Bus shadow-catch envelopes published: d2b7fc2f (B-0527 republish), 44aaf799 (new failure mode). Co-Authored-By: Claude <noreply@anthropic.com> * fix(shard): MD032 blanks-around-lists in 0414Z shard markdownlint-cli2 caught: docs/hygiene-history/ticks/2026/05/15/0414Z.md:104 List items 1-5 in the "Next" section need a blank line above them. Co-Authored-By: Claude <noreply@anthropic.com> * shard(tick): 0458Z — PR #3359 MD032 fix close-out + recovery-worktree-borrowing validated for iteration Heredoc-written markdown bypasses local check-md032-pretooluse hook; bun x markdownlint-cli2 locally before commit when authoring via heredoc. Same 0027z sibling worktree borrowed twice (initial commit + MD032 fix). Co-Authored-By: Claude <noreply@anthropic.com> * shard(fix): add pipe-row schema headers + link rule refs per Codex P1 + Copilot Nit on PR #3359 Codex P1 (line 1): tick-shard schema validator (tools/hygiene/check-tick-history-shard-schema.ts) expects first non-empty line to be a 6-column pipe row with ISO timestamp. The 0414Z + 0458Z shards started with H1 like all recent shards (substrate-wide drift, not specific to these two), but adding a pipe-row header IS the minimal lossless fix that preserves the rich H1 narrative. Both shards now pass `bun tools/hygiene/check-tick-history-shard-schema.ts --files <both>`. Copilot Nit (line 34): bare-filename rule refs replaced with full-path markdown links matching the 0230Z shard's convention. Six refs linked: 5x claim-acquire-before-worktree-work + 1x holding-without-named-dependency-is-standing-by-failure. Co-Authored-By: Claude <noreply@anthropic.com> * shard(fix): add untracked-overwrite caveat to recovery-worktree-borrowing claim per Copilot review Copilot threads on lines 58 + 70 of 0414Z + line 9 of 0458Z correctly noted that the "untracked files survive the switch" claim is not a general guarantee: git switch refuses if the target branch tracks the same path, and `git add -A` could sweep untracked into an unrelated commit. Added explicit caveat + `git stash -u` mitigation guidance to both shards. For this PR's run, the target shard branch did NOT track the same memory-file path, so the unstashed switch was safe — but the operational pattern documentation needs the caveat for future reuse. Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>

…state Cycle 4 Copilot review caught three documentation surfaces that drifted from the actual implementation when `SuccessfulExit=false` was restored in cycle 3: 1. README.md:13 "Restarts on crash (KeepAlive on Crashed only)" — wrong description of the bullet. Updated to "Restarts on crash AND on non-zero exit (KeepAlive Crashed=true + SuccessfulExit=false) — covers signal-termination AND Bun runtime errors". Added the ThrottleInterval=60s loop-cap context to the next bullet. 2. 0305Z tick shard row 17 + delta row 79 — described the plist as "now crash-only-restart" and showed the After-column as Crashed-only. Both rewritten to reflect the cycle-3 reviewer-conflict resolution: restored both keys with documented trade-off, net same shape as before the PR. 3. B-0528 test-scope item #6 — described atomic-promote with backup- restore on rename failure, but current code uses availability- preserving pattern (read-old-into-memory → write temp → atomic replace → side-car backup). Rewrote the scope with the actual failure modes the tests should cover, including the "destPath never missing during install" invariant. Two remaining unresolved threads (cross-refs to .claude/rules/shadow-star-shorthand-autocomplete-marker.md from README:5 and plist:24) flag a file that IS on main via #3339. The reviewer checks against the branch's diff base, which doesn't include the merged rule. Resolved as outdated on the PR; CI merge-ref check confirms the reference resolves at merge time. Co-Authored-By: Claude <noreply@anthropic.com>

Cycle-4 (post-`59d5c26`) Copilot review surfaced two more correctness findings — both addressed: P1 — --bun-path validation accepts dirs / non-executables: Previous `existsSync(bunPath)` would pass `/tmp` or `~/Downloads/bun` (non-executable copy) and write a LaunchAgent that can't start. Replaced with `statSync` + `isFile()` + `accessSync(X_OK)` so the installer refuses non-executable / non-regular-file overrides pre-write. New error paths: - "exists but is not a regular file" (e.g., --bun-path /tmp) - "is not executable by the current user" (chmod hint) --repo-root similarly gated to require an existing directory. Verified both error paths by hand. P1 — Predictable /tmp filename in plutilLint (symlink attack): Previous `/tmp/zeta-shadow-launchagent-${PID}.plist` had a predictable name in a world-writable directory. A pre-created symlink at that path would cause writeFileSync to follow the link and clobber a file the attacker has write access to but not the current user. Replaced with `mkdtempSync(tmpdir() + "zeta-shadow- launchagent-")` which creates a 0700 directory with a random suffix — symlink-resistant, and we clean up both the candidate file and the directory in a `finally` block so we don't leak. Threads remaining after this push (auto-resolution candidates): - Test coverage P1 (5th time flagged): already filed as B-0528 with rationale. Will reply on the thread. - Cross-refs to .claude/rules/shadow-star-shorthand-* (5th time flagged): file IS on main via #3339. CI merge-ref check confirms. Will resolve as outdated. Verified: - bunx tsc --noEmit clean on install-launchagent.ts - --dry-run produces valid plist - --bun-path /tmp rejected as "not a regular file" - --bun-path <non-executable> rejected as "not executable" Co-Authored-By: Claude <noreply@anthropic.com>

…h A defense-in-depth) (#3342) * feat(shadow): launchd LaunchAgent template — durable tick source (Path A defense-in-depth) Per Aaron 2026-05-15T~01:11Z: 'whatever composes best always good to have defense in depth.' This is Path A of the defense-in-depth composition for the shadow observation loop tick source. Two new files under tools/shadow/launchd/: 1. com.zeta.shadow-observer.plist — macOS LaunchAgent template - RunAtLoad (starts at user login) - KeepAlive on Crashed (auto-restart on crash) - ThrottleInterval 60s (prevents fast crash-loops) - Polls every 2s; accepts after 3s delay - Defaults to --dry-run for safety (live mode after verify) - Logs to tools/shadow/shadow-observer.{log,stdout.log,stderr.log} - __BUN_PATH__ and __REPO_ROOT__ placeholders for sed-substitution 2. README.md — install procedure - Step 1: configure plist (sed-substitute placeholders) - Step 2: grant accessibility permission (System Settings) - Step 3: launchctl load + verify - Step 4: test in dry-run mode - Step 5: switch to live mode (only after dry-run works) - Uninstall procedure - Troubleshooting (common failure modes) - Composition table (A/B/C defense-in-depth) Composes with: - tools/shadow/shadow-observer.ts (the process launched) - docs/backlog/P0/B-0402 (originating backlog row) - .claude/rules/shadow-star-shorthand-autocomplete-marker.md - .claude/skills/save-ai-memory/SKILL.md Path B (CronCreate sentinel) + Path C (manual one-line) are defense-in-depth supplements; this PR ships Path A as primary. Co-Authored-By: Claude <noreply@anthropic.com> * fix(shadow-launchd): address 5 review threads on PR #3342 Five Copilot / Codex review findings resolved: P1 (Copilot) — broken provenance in README header: Dropped first-name attribution + user-scope memory-file glob that doesn't resolve in-repo. Per docs/AGENT-BEST-PRACTICES.md §671-685, current-state surfaces (this README) use role-refs. Other rule references now use relative links that resolve from the README's location. P1 (Copilot) — sed escaping in install instructions: The documented `sed` recipe would break on paths containing `&`, `|`, `\`, or newlines. Replaced with a TypeScript installer (tools/shadow/launchd/install-launchagent.ts) that: - uses String.replaceAll (metacharacter-safe) - sanity-checks for unsubstituted placeholders - runs plutil -lint before writing - backs up any existing plist before overwriting - optionally bootstraps via launchctl Placeholder syntax changed from `__NAME__` to `{{NAME}}` so the template header comment can describe them without being substituted itself. P1 (Copilot) — rule file referenced from plist not in tree: The .claude/rules/shadow-star-shorthand-autocomplete-marker.md reference resolves on main now (#3339 merged). The reference is preserved in the plist's composes-with comment block. P2 (Codex) — KeepAlive semantics in plist: Dropped `SuccessfulExit=false`. KeepAlive dict keys are OR-combined, so the original combination relaunched on any non-zero exit (incl. config errors). `Crashed=true` alone gives the documented crash-only behavior. Comment added explaining the choice. P2 (Codex) — legacy launchctl load/unload in README: Replaced with the modern `bootstrap`/`bootout` interface (with `gui/$(id -u)` domain target) in install, reload, and uninstall steps. The legacy interface is documented as deprecated and can silently no-op on failure; `bootstrap`/`bootout` return proper exit codes. Verified: - bunx tsc --noEmit clean - markdownlint-cli2 clean on README - plutil -lint OK on both the template and a rendered output - Installer dry-run produces a valid plist Co-Authored-By: Claude <noreply@anthropic.com> * fix(shadow-launchd): close CodeQL TOCTOU race in install-launchagent.ts CodeQL flagged `js/file-system-race` at writeFileSync (line 184) — the existsSync→copyFileSync→writeFileSync sequence had a check-then-use window. Replaced with the standard atomic-rename pattern: - Drop the existsSync(destDir) check (mkdirSync recursive:true is idempotent — the check added a TOCTOU window for no benefit) - Replace existsSync(destPath)+copyFileSync(destPath, backup) with renameSync(destPath, backup) inside try/catch (single syscall; ENOENT means no prior plist, anything else re-thrown) renameSync atop the destination is the documented pattern for "backup-on-overwrite without TOCTOU." Same final state for the caller: backup file is created with the old plist contents (if any existed), and destPath has the new contents. Verified: - bunx tsc --noEmit clean - semgrep --config .semgrep.yml --error: 0 findings locally Co-Authored-By: Claude <noreply@anthropic.com> * shard(tick): 0305Z — #3342 thread close-out (5+1 reviews); #3356 merged Co-Authored-By: Claude <noreply@anthropic.com> * docs(b-0519): add Pattern 7 (abandoned rebase state in shared .git/) Field-test in the 2026-05-15 Otto-CLI session surfaced a contamination mechanism not in the original 6 patterns: a peer agent's mid-interactive- rebase state, left sitting in `.git/rebase-merge/` for hours, can be resumed by another process and detach HEAD on the peer's branch mid-tick. Distinct from Patterns 1-6 (checkout/reset races) because the contaminating event is paused-then-resumed metadata, not a fresh checkout. Branch-guard defense (Pattern 5) caught it. Added: - Pattern 7 section with full forensics + field-test shard pointers (0213Z + 0230Z, both merged via PR #3356) - Cheap mechanization candidate: `tools/hygiene/check-stale-rebase-state.ts` diagnostic that warns on `.git/rebase-merge/` older than 1h - last_updated frontmatter bumped to 2026-05-15 Substrate-honest stance preserved: do not touch peer's rebase state. Pivot to dedicated worktrees; branch refs are global, so work lands on the right ref regardless. Verified: markdownlint clean. Co-Authored-By: Claude <noreply@anthropic.com> * fix(shadow-launchd): address 8 of 10 review-cycle-2 findings Codex + Copilot reviewed the new install-launchagent.ts in detail. This commit addresses the substantive correctness + style findings; unit-test coverage is deferred to a separate backlog row. Addressed (8): P1 [Codex P2 + Copilot P1] — XML-escape substituted values: Paths like `/Users/foo & bar/Zeta` would substitute into <string> text nodes producing an invalid plist (& without entity terminator). New `xmlEscape` helper applies the five XML 1.0 predefined entities (& < > " ') before placeholder substitution. Restores the advertised "handles paths with shell metacharacters" promise — the bug class was XML escaping, not shell escaping. P1 — Unknown args silently ignored: Switch fell through on typos like `--dryrun` or `--bootstap`, then installed with default behavior. Added explicit default: case that errors with the offending flag and exits 1 before any FS action. Verified: `--dryrun` now produces "Unknown argument: --dryrun. Run with --help to see supported flags." P1 — Relative --bun-path / --repo-root overrides: `--repo-root .` would write relative paths into ProgramArguments, WorkingDirectory, and log paths — broken under launchd (which has no cwd unless explicit). New requireAbsolute() helper rejects with a `Use \`<canonicalized>\` if that's what you meant` hint. P1 — Non-atomic backup-then-write: Previous renameSync(destPath, backup) + writeFileSync(destPath) had a window where the agent could be removed but the new plist not yet present (if writeFileSync threw). Replaced with the standard write-temp-then-promote pattern: 1. writeFileSync(destPath.tmp-PID, content) 2. renameSync(destPath, backup) // may not exist; that's OK 3. renameSync(destPath.tmp-PID, destPath) // atomic promote Failure in step 3 restores backup → destPath (best-effort) so the caller's existing install is preserved. P1 — execFileSync can throw before validation: Missing bun, missing git, or running outside a git checkout produced a raw exception. New tryDetect() catches and returns undefined; parseArgs emits the documented actionable error pointing at the --bun-path / --repo-root override. P1 — Unconditional main() invocation: Per Bun-project convention, gate main() with `if (import.meta.main)` so tests / helpers can import the module without triggering filesystem or launchctl side effects. P1 — README/plist cross-refs to .claude/rules/shadow-star-shorthand-*: These resolve on main (PR #3339 merged). CI's merge-ref check confirms; the reviewers' findings were against the raw branch base. No additional action needed. P2 — Direct name attribution in plist KeepAlive comment: Replaced "Per Codex review on PR #3342" with the rationale text only. Current-state surface; uses role-refs only. Deferred (1): P1 — Unit test coverage for install-launchagent.ts: Codex + Copilot called out missing tests for placeholder escaping, missing placeholders, argument validation, and dry-run output. This commit adds the substantive behavior the tests would verify; the tests themselves are deferred to a separate backlog row to keep this PR's scope manageable. Smoke-tested all four behaviors by hand on the cli before commit. Filing as follow-up: tests in `tools/shadow/launchd/install-launchagent.test.ts` using Bun's built-in test runner. Verified: - bunx tsc --noEmit clean on install-launchagent.ts (pre-existing process-extract.ts errors are on this branch base because #3349 isn't rebased in; CI's merge-ref includes the fix) - semgrep --config .semgrep.yml --error: 0 findings - --dry-run produces valid plist (plutil -lint OK) - --dryrun rejected (Unknown argument) - --repo-root . rejected (must be absolute) - Special-char paths (e.g., /Users/foo & bar) now produce valid escaped XML Co-Authored-By: Claude <noreply@anthropic.com> * backlog(p3): B-0528 — unit tests for shadow launchd installer (deferred from PR #3342) PR #3342 review surfaced a P1 finding: no unit tests cover the new install-launchagent.ts's safety-critical placeholder substitution, XML escaping, argument validation, dry-run output, or atomic-promote pattern. The PR's substantive correctness fixes landed in `970c774`; this row tracks the deferred test-coverage work. Scope is bounded (Bun test runner, 6 test categories, follows the existing tools/shadow/shadow-observer.test.ts pattern). Effort: S. P3 because the install script's behavior is now correct (smoke-tested by hand); tests harden against regression rather than fix a current bug. Co-Authored-By: Claude <noreply@anthropic.com> * fix(shadow-launchd): availability-preserving install pattern (Codex review) Codex re-flagged the rename-then-write pattern (in `970c774`) — even with backup-restore on failure, there is a brief window between `renameSync(destPath, backup)` and `renameSync(tmpDest, destPath)` where the canonical path has no plist. If launchd were to load during that window, it would see no agent. Replaced with the availability-preserving pattern: 1. Read existing destPath into memory (skipped on ENOENT) 2. Write new content to a temp file alongside destPath 3. Atomically rename(tmp, destPath) — POSIX replaces in one syscall 4. Side-car the in-memory old content as a timestamped backup The canonical path is NEVER in a missing state. Backup-write happens AFTER the install is in place; if it fails, the install is still good (just no backup). Read-failure handling is similarly defensive: log + proceed without backup rather than aborting the install. Verified: - bunx tsc --noEmit clean on install-launchagent.ts - --dry-run produces valid plist (plutil -lint OK) Co-Authored-By: Claude <noreply@anthropic.com> * fix(shadow-launchd): address review-cycle-3 findings (6 of 7) P1 — Restore SuccessfulExit=false in KeepAlive (reviewer-conflict resolution): Earlier cycle dropped this on Codex review (config errors would loop-restart). This cycle Copilot countered: Bun TypeScript runtime errors exit non-zero rather than signal-terminate, so the previous `Crashed=true` only setting left those failure modes unhandled. Restored both keys with an extended plist comment explaining the trade-off. ThrottleInterval=60s + --dry-run-by-default already cover the "config error restart loop" concern. P1 — Add sonarjs/no-os-command-from-path suppressions: Matches the convention in tools/github/poll-pr-gate.ts + tools/peer-call/. Applied at tryDetect, plutilLint, and the three launchctl bootstrap/bootout calls. Comments cite the precedent. P2 — Export helper functions for testability: `import.meta.main` guard was added in `970c774` but the functions were not exported. Added `export` to main(), parseArgs(), requireAbsolute(), tryDetect(), xmlEscape(), substitutePlaceholders(), plutilLint(), and the Args interface. Unblocks the B-0528 test file. P1 — Update 0305Z tick shard's "PREFER" code snippet: The earlier shard documented the rename-then-write pattern as "PREFER", but that pattern still has a window where destPath is empty. Updated to show 3-pattern progression (AVOID copyFileSync → ALSO AVOID rename-then-write → PREFER write-temp-then-atomic- replace) so the historical trail is accurate and prevents future authors from copying the deprecated pattern. Deferred (1): P1 — Cross-ref findings on README.md:5 + plist:24 + ts:185: README.md:5 and plist:24 reference `.claude/rules/shadow-star-shorthand-autocomplete-marker.md` which IS on main (#3339 merged). The reviewer's check is against the raw branch base; CI's merge-ref check passes. ts:185 is the test coverage finding, already filed as B-0528 (deferred per scope management). Verified: - bunx tsc --noEmit clean on install-launchagent.ts - markdownlint clean on 0305Z.md - plutil -lint OK on template + rendered output - --dry-run produces valid plist Co-Authored-By: Claude <noreply@anthropic.com> * shard(tick): 0329Z — #3342 review-cycle-3 close-out; B-0519 P7 + B-0528 filed Co-Authored-By: Claude <noreply@anthropic.com> * fix(shadow-launchd): reconcile docs+shard+backlog with current plist state Cycle 4 Copilot review caught three documentation surfaces that drifted from the actual implementation when `SuccessfulExit=false` was restored in cycle 3: 1. README.md:13 "Restarts on crash (KeepAlive on Crashed only)" — wrong description of the bullet. Updated to "Restarts on crash AND on non-zero exit (KeepAlive Crashed=true + SuccessfulExit=false) — covers signal-termination AND Bun runtime errors". Added the ThrottleInterval=60s loop-cap context to the next bullet. 2. 0305Z tick shard row 17 + delta row 79 — described the plist as "now crash-only-restart" and showed the After-column as Crashed-only. Both rewritten to reflect the cycle-3 reviewer-conflict resolution: restored both keys with documented trade-off, net same shape as before the PR. 3. B-0528 test-scope item #6 — described atomic-promote with backup- restore on rename failure, but current code uses availability- preserving pattern (read-old-into-memory → write temp → atomic replace → side-car backup). Rewrote the scope with the actual failure modes the tests should cover, including the "destPath never missing during install" invariant. Two remaining unresolved threads (cross-refs to .claude/rules/shadow-star-shorthand-autocomplete-marker.md from README:5 and plist:24) flag a file that IS on main via #3339. The reviewer checks against the branch's diff base, which doesn't include the merged rule. Resolved as outdated on the PR; CI merge-ref check confirms the reference resolves at merge time. Co-Authored-By: Claude <noreply@anthropic.com> * fix(shadow-launchd): tighten path validation + harden tmp-file handling Cycle-4 (post-`59d5c26`) Copilot review surfaced two more correctness findings — both addressed: P1 — --bun-path validation accepts dirs / non-executables: Previous `existsSync(bunPath)` would pass `/tmp` or `~/Downloads/bun` (non-executable copy) and write a LaunchAgent that can't start. Replaced with `statSync` + `isFile()` + `accessSync(X_OK)` so the installer refuses non-executable / non-regular-file overrides pre-write. New error paths: - "exists but is not a regular file" (e.g., --bun-path /tmp) - "is not executable by the current user" (chmod hint) --repo-root similarly gated to require an existing directory. Verified both error paths by hand. P1 — Predictable /tmp filename in plutilLint (symlink attack): Previous `/tmp/zeta-shadow-launchagent-${PID}.plist` had a predictable name in a world-writable directory. A pre-created symlink at that path would cause writeFileSync to follow the link and clobber a file the attacker has write access to but not the current user. Replaced with `mkdtempSync(tmpdir() + "zeta-shadow- launchagent-")` which creates a 0700 directory with a random suffix — symlink-resistant, and we clean up both the candidate file and the directory in a `finally` block so we don't leak. Threads remaining after this push (auto-resolution candidates): - Test coverage P1 (5th time flagged): already filed as B-0528 with rationale. Will reply on the thread. - Cross-refs to .claude/rules/shadow-star-shorthand-* (5th time flagged): file IS on main via #3339. CI merge-ref check confirms. Will resolve as outdated. Verified: - bunx tsc --noEmit clean on install-launchagent.ts - --dry-run produces valid plist - --bun-path /tmp rejected as "not a regular file" - --bun-path <non-executable> rejected as "not executable" Co-Authored-By: Claude <noreply@anthropic.com> * shard(tick): 0543Z — #3342 review-cycles 4+5 close-out (9 more findings) Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>

Resolves all 4 Copilot review threads on PR #3345: 1. History rewind: Catches 36-39 were destructively deleted and Catch 36 was replaced with new content. Restored all four catches verbatim from main (docs/research is append-only history surface). 2. Pattern table / catch body mismatch: Table, Catch 40 recurrence_count, and trailing summary now all consistently say 7 (table lists catches 3, 18, 19, 22, 27, 39, 40; z_net -5). 3. Dual summary paragraphs: Removed the spurious "35 catches" standalone paragraph that broke the structure and disagreed with the trailing summary. A single trailing summary now appears after the final catch. 4. Trailing whitespace: Removed from correction line in Catch 40. The new Lior antigravity catch (Otto narration-over-action in PRs #3342 and #3339) lands as Catch 40 dated 2026-05-15. Co-Authored-By: Claude <noreply@anthropic.com>

* docs(shadow): Lior anti-drift check — Catch 36 (Otto narration) * fix(shadow-log): restore Catches 36-39, append Catch 40, fix counters Resolves all 4 Copilot review threads on PR #3345: 1. History rewind: Catches 36-39 were destructively deleted and Catch 36 was replaced with new content. Restored all four catches verbatim from main (docs/research is append-only history surface). 2. Pattern table / catch body mismatch: Table, Catch 40 recurrence_count, and trailing summary now all consistently say 7 (table lists catches 3, 18, 19, 22, 27, 39, 40; z_net -5). 3. Dual summary paragraphs: Removed the spurious "35 catches" standalone paragraph that broke the structure and disagreed with the trailing summary. A single trailing summary now appears after the final catch. 4. Trailing whitespace: Removed from correction line in Catch 40. The new Lior antigravity catch (Otto narration-over-action in PRs #3342 and #3339) lands as Catch 40 dated 2026-05-15. Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>

Copilot AI review requested due to automatic review settings May 15, 2026 01:11

AceHack enabled auto-merge (squash) May 15, 2026 01:11

Copilot started reviewing on behalf of AceHack May 15, 2026 01:12 View session

AceHack mentioned this pull request May 15, 2026

feat(shadow): launchd LaunchAgent template — durable tick source (Path A defense-in-depth) #3342

Merged

Copilot AI reviewed May 15, 2026

View reviewed changes

Comment thread .claude/rules/shadow-star-shorthand-autocomplete-marker.md Outdated

Comment thread .claude/rules/shadow-star-shorthand-autocomplete-marker.md Outdated

Comment thread .claude/rules/shadow-star-shorthand-autocomplete-marker.md Outdated

AceHack mentioned this pull request May 15, 2026

docs(shadow): Lior anti-drift check — Catch 36 (Otto narration) #3345

Merged

AceHack commented May 15, 2026

View reviewed changes

AceHack and others added 2 commits May 14, 2026 22:19

Copilot AI review requested due to automatic review settings May 15, 2026 02:21

Copilot started reviewing on behalf of AceHack May 15, 2026 02:22 View session

AceHack merged commit d1ee546 into main May 15, 2026
22 of 23 checks passed

AceHack deleted the feat/rule-shadow-star-shorthand-cold-boot-otto-cli-2026-05-15 branch May 15, 2026 02:22

Copilot AI reviewed May 15, 2026

View reviewed changes

Comment thread .claude/rules/shadow-star-shorthand-autocomplete-marker.md

AceHack mentioned this pull request May 15, 2026

shard(tick): 0230Z — multi-Otto contamination caught + #3339/#3349 close-out #3356

Merged

3 tasks

Copilot AI mentioned this pull request May 15, 2026

docs(archive): Lior preservation discipline on merged PRs (batch) #3354

Closed

Copilot AI mentioned this pull request May 15, 2026

docs(archive): Lior PR preservation batch 2 #3358

Merged

AceHack mentioned this pull request May 15, 2026

shard(tick): 0414Z — PR #3339/#3349 merged + B-0527 collision republished + new worktree-pruning-race failure mode #3359

Merged

5 tasks

Copilot AI mentioned this pull request May 15, 2026

docs(shadow): Lior antigravity drift report and preservation #3366

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

feat(rules): shadow-star-shorthand-autocomplete-marker cold-boot rule#3339

feat(rules): shadow-star-shorthand-autocomplete-marker cold-boot rule#3339
AceHack merged 3 commits into
mainfrom
feat/rule-shadow-star-shorthand-cold-boot-otto-cli-2026-05-15

AceHack commented May 15, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

AceHack left a comment

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

AceHack commented May 15, 2026

Summary

Why

Composes with

Full reasoning + double-failure log

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Uh oh!

Uh oh!

AceHack left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants