feat(routines): git-tracked Claude Desktop routines + autonomous-loop registration#3034
Merged
AceHack merged 12 commits intoMay 13, 2026
Merged
Conversation
…omous-loop registration Aaron asked for a power-user / scripting / version-control path for Claude Desktop scheduled tasks (the "Routines" sidebar panel), composed with the existing CLI <<autonomous-loop>> cron substrate. This introduces tools/routines/ as canonical-source-of-truth for Desktop routines and registers the first one (autonomous-loop) as a Desktop-side every-2-hour fallback for the CLI every-minute tick. Two-layer architecture: tools/routines/<id>/ — git-tracked canonical (SKILL.md + schedule.json) ~/.claude/scheduled-tasks/ — runtime location, generated from canonical tools/routines/install.ts — idempotent Bun TS installer (rule-0 compliant) The scheduled-tasks MCP server is cross-host — same server wired into both CLI Claude Code and Claude Desktop. Both surfaces read/write the same disk store. This PR makes the prompt-body authoring path git-canonical so routines are diffable, PR-reviewable, and shareable across maintainer machines. Three-layer catch-43 defence now in place: 1. CLI session cron — every minute, cheap re-prompt, dies on exit 2. Desktop routine — every 2 hours, persistent cold-boot, survives restart 3. tools/routines/ repo — git canonical, survives runtime corruption Composes with: - .claude/rules/tick-must-never-stop.md (catch-43 substrate) - .claude/rules/rule-0-no-sh-files.md (TS installer, not bash) - .claude/rules/holding-without-named-dependency-is-standing-by-failure.md (PR #3029) - .claude/rules/dv2-data-split-discipline-activated.md (canonical vs runtime change-rate split) - docs/AUTONOMOUS-LOOP.md (canonical tick procedure) Tick shard: docs/hygiene-history/ticks/2026/05/13/2125Z.md Co-Authored-By: Claude <noreply@anthropic.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 01fcf40a7e
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
![github-advanced-security[bot]](https://avatars.githubusercontent.com/in/57789?s=60&v=4){kind=small}
There was a problem hiding this comment.
Pull request overview
Adds a git-tracked “canonical routines” substrate under tools/routines/ for Claude Desktop scheduled tasks, plus an initial autonomous-loop routine and a Bun/TS installer to sync the canonical SKILL.md into the Desktop runtime directory (~/.claude/scheduled-tasks/). This complements the existing CLI session-scoped cron tick by providing a durable Desktop-side cadence.
Changes:
- Introduces
tools/routines/documentation and a first routine (autonomous-loop) withSKILL.md+schedule.json. - Adds
tools/routines/install.tsto sync canonical routine prompts into the Desktop runtime directory. - Records the work as a hygiene-history tick shard.
Reviewed changes
Copilot reviewed 5 out of 5 changed files in this pull request and generated 5 comments.
Show a summary per file
| File | Description |
|---|---|
| tools/routines/README.md | Documents the canonical-vs-runtime routines architecture and authoring workflow. |
| tools/routines/install.ts | Bun installer that syncs canonical SKILL.md files into ~/.claude/scheduled-tasks/<id>/SKILL.md. |
| tools/routines/autonomous-loop/SKILL.md | Adds the Desktop routine prompt body for the autonomous loop. |
| tools/routines/autonomous-loop/schedule.json | Adds a 2-hour cron schedule metadata file for the routine. |
| docs/hygiene-history/ticks/2026/05/13/2125Z.md | Tick shard capturing the work, verification trace, and rationale. |
Comments suppressed due to low confidence (2)
tools/routines/install.ts:61
readSchedulecatches JSON parse failures but then returns{ missing: false }, which hides an invalidschedule.jsonand produces ambiguous output (no cron printed and no “missing schedule” note). Consider surfacing parse errors explicitly (flag + warning) and/or failing the run so broken schedules can’t silently slip through.
try {
const parsed = JSON.parse(readFileSync(path, "utf8")) as { cronExpression?: string };
return { cronExpression: parsed.cronExpression, missing: false };
} catch {
return { missing: false };
}
tools/routines/install.ts:124
- Schedule changes won’t be detected:
needsRegistrationonly includes routines when SKILL.md was created/updated, so editingschedule.jsonalone won’t prompt re-registration and can lead to silent schedule drift. Consider tracking schedule state too (marker/hash) or always emitting registration instructions whencronExpressionis present.
const needsRegistration = results.filter(
(r) => r.cronExpression && (r.action === "created" || r.action === "updated"),
);
if (needsRegistration.length > 0) {
… + testability Bundles all 7 review findings from PR #3034 into one commit: install.ts - Fix tsc exactOptionalPropertyTypes violations (was the BLOCKING required check failure) - Eliminate TOCTOU races: replace existsSync+readFileSync pairs with try/catch around readFileSync via readFileOrUndefined helper (addresses CodeQL alert) - Surface invalid schedule.json instead of silently swallowing parse failures: new scheduleParseError field + console.error + parseErrors summary count (Codex P2) - Export listRoutines/readSchedule/syncRoutine/main with directory parameters so tests can drive deterministically without touching real homedir() or import.meta.dir (Copilot test-coverage suggestion; actual test suite is a follow-up) - Gate main() under if (import.meta.main) — aligns with other tools/** scripts that use the importable/testable pattern (Copilot) autonomous-loop/SKILL.md - Remove hardcoded /Users/acehack/Documents/src/repos/Zeta path from routine prompt (Codex P1, Copilot duplicate). Now portable across maintainer machines via "typically ~/Documents/src/repos/Zeta" guidance + project-metadata fallback. Installer will sync the new prompt to the live runtime SKILL.md on next run; the next routine fire (22:07Z) picks up the portable version. README.md - Fix Rule 0 wording: was "only `.sh` allowed there" (implied tools/setup/ is .sh-only); now ".sh files are restricted to under tools/setup/; other formats also live there" (matches .claude/rules/rule-0-no-sh-files.md) (Copilot) docs/hygiene-history/ticks/2026/05/13/2125Z.md - Update pr: TBD → pr: 3034 (Copilot — placeholder breaks consumers that parse frontmatter) Verify trace: 1. npx tsc --noEmit on routines files: clean ✓ 2. bun tools/routines/install.ts: [updated] autonomous-loop, parseErrors=0 ✓ 3. Runtime SKILL.md at ~/.claude/scheduled-tasks/autonomous-loop/SKILL.md now portable Follow-up (not in this commit): add bun:test test suite for install.ts (Copilot suggestion; refactor for testability landed here so the suite can land cleanly). Co-Authored-By: Claude <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 79c00b9df4
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
- SKILL.md:16 — wrap cron expression in backtick span to avoid MD037 (asterisks in "* * * * *" were parsed as emphasis markers) - 2125Z.md:43 — add blank line before list to satisfy MD032 (markdownlint requires blank lines surrounding lists) Co-Authored-By: Claude <noreply@anthropic.com>
Real-time observation: Otto-CLI hijacked the primary worktree branch context while Otto-Desktop was working there, in the SAME session that Otto-CLI authored PR #3032's claim-acquire-before-worktree rule. The rule was speculative when proposed; this observation is its first empirical validation. Files: - memory/feedback_split_brain_real_time_otto_cli_otto_desktop_primary_worktree_branch_hijack_pr_3032_claim_acquire_rule_validation_2026_05_13.md - docs/hygiene-history/ticks/2026/05/13/2140Z.md The memory extends PR #3032's rule operationally — adds 3 clauses beyond "claim acquire before worktree work": 1. Each Otto gets ONE dedicated worktree (never share primary) 2. Never git checkout on a worktree another Otto is using 3. Bus claim envelope should include 'worktree' field Composes with PR #3032 (validates), substrate-or-it-didn't-happen (rules in flight don't apply to behavior in flight), glass-halo-bidirectional (observation enabled diagnosis). Co-Authored-By: Claude <noreply@anthropic.com>
The non-required CI check "check memory file frontmatter completeness" flagged the file. Added YAML frontmatter (name/description/type) matching the convention used by other memory/feedback_*.md files. Co-Authored-By: Claude <noreply@anthropic.com>
…ion hint
P1: registration hint was gated on action=created|updated, so a
schedule-only change (SKILL.md unchanged) silently skipped the
create_scheduled_task reminder. Remove the action emitfilter
the hint whenever cronExpression is present, regardless of action.
P2: readSchedule returned { missing: false } with no parseError
when schedule.json parsed but lacked cronExpression, producing a
silent misconfiguration. Now surfaces a descriptive parseError so
the caller prints the malformed-schedule warning.
Addresses chatgpt-codex-connector review threads on PR #3034.
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…n optionality Two Copilot review threads on PR #3034: install.ts (Copilot — BP "No name attribution in code"): - Doc comment said "ask Otto (or call directly) to run create_scheduled_task" - Console output said "(in a Claude session, ask Otto to run create_scheduled_task for each)" - Both replaced with persona-agnostic phrasing: "invoke create_scheduled_task from an interactive Claude session (or via direct MCP API call)" README.md (Copilot — internal consistency): - Top-level routine description said each routine has SKILL.md + schedule.json without noting schedule.json is optional - But installer explicitly supports missing schedule.json (ad-hoc routines) - Now: SKILL.md marked **required**, schedule.json marked **optional** with the ad-hoc clarification both in the bullet list and in the Authoring section Verify trace: - bun tools/routines/install.ts → [skipped-unchanged] (idempotent confirmed) - npx tsc --noEmit → clean on routines files Co-Authored-By: Claude <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: fbdc1fabca
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…errors (Codex P2 x2) Two P2 findings from Codex on PR #3034: 1. Reject non-string cronExpression values (PRRT_kwDOSF9kNM6B5oEP): readSchedule now narrows parsed.cronExpression via typeof === "string" before accepting it. Non-string values (number, null, object) return a parseError with the offending type. Defends against schedule.json files that happen to round-trip through JSON but have semantically wrong types. 2. Fail installer when schedule parsing reports errors (PRRT_kwDOSF9kNM6B5oEU): main() now returns an exit code (0 on success, 1 if parseErrors > 0). The if (import.meta.main) entrypoint calls process.exit(main()) so CI catches malformed schedule.json files instead of silently counting them in the summary and exiting 0. Verify trace: - bun tools/routines/install.ts (clean run): exits 0, parseErrors=0 ✓ - npx tsc --noEmit: clean on routines files ✓ Co-Authored-By: Claude <noreply@anthropic.com>
…g correction Co-Authored-By: Claude <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: f0652b8e14
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…ntmatter The memory-index-integrity CI gate requires all four frontmatter fields: name, description, type, created. The prior fix commit (459a511) added the other three fields but omitted created. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
3 tasks
AceHack
added a commit
that referenced
this pull request
May 13, 2026
…fence layer (#3039) Aaron 2026-05-13 authorized this row ("yes that sounds good about the backlog too") after the same-session research surfaced Cloud Routines as a 4th catch-43 layer Anthropic ships: - Scheduled triggers (hourly/daily/nightly/weekdays/weekly) - API triggers (HTTP POST + bearer token per routine) - GitHub event triggers (pull_request.opened, push, issues, releases, check_runs) Usage caps: Pro 5/day, Max 15/day, Team/Enterprise 25/day. Builds on PR #3034's tools/routines/ substrate (the canonical-source-of-truth two-layer architecture established earlier today for Desktop routines). Cloud Routines extend the same canonical pattern with a sibling cloud-schedule.json file. Killer use case identified: GitHub event triggers unlock per-PR review ticks without any local cron infrastructure — cheap (uses GitHub event quota only, not per-day cron quota) and zero local-process risk. Decomposition hint included (B-0448.1 through B-0448.7) for the implementing slice. Co-authored-by: Claude <noreply@anthropic.com>
…ng, empty cron rejection Four threads addressed (3 P1 + 2 P2 from Codex/Copilot on PR #3034): install.ts: - readFileOrUndefined now distinguishes ENOENT (returns undefined → "missing file") from other errors (re-throws → installer fails loudly). Permission errors, IO failures no longer silently treated as missing. (Codex P2, Copilot P1) - readSchedule now rejects empty cronExpression strings with parseError "cronExpression must be a non-empty string". (Codex P2) autonomous-loop/SKILL.md + README.md: - Remove remaining "Otto" persona attribution per BP "No name attribution in code" on current-state surfaces under tools/. SKILL.md description + first paragraph now use "Autonomous-loop tick" without persona claim. README.md "Ask Otto" rephrased to direct MCP invocation; "Otto bootstream" rephrased to "canonical bootstream" with note that the literal filename is historical-surface substrate. (Copilot P1) Identity-stays-unified substrate (PR #3036) is preserved — the fresh-session that fires this routine identifies as Otto via the bootstream's Part 1, not via the routine prompt itself. Routine prompt is functional description; identity is established at cold-boot. Deferred to follow-up row: bun:test coverage for install.ts (Copilot P1 thread 5). Refactor for testability already shipped (exported pure functions accept directory params); actual test suite is its own task — will file as a separate backlog row. Verify trace: - bun tools/routines/install.ts: exits 0, parseErrors=0 ✓ - npx tsc --noEmit: clean on routines files ✓ Co-Authored-By: Claude <noreply@anthropic.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 6702abb4fd
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
3 tasks
…x P1 + P2) Same silent-failure-prevention pattern as readFileOrUndefined: 1. listRoutines now distinguishes ENOENT (returns empty array — the no
| { | ||
| "taskId": "autonomous-loop", | ||
| "cronExpression": "0 */2 * * *", | ||
| "description": "Otto autonomous-loop tick — fresh-session cold-boot, fires every 2 hours on Desktop", |
Comment on lines
+168
to
+170
| const routines = listRoutines(repoRoutinesDir).filter( | ||
| (id) => id !== "install.ts" && !id.endsWith(".md"), | ||
| ); |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: d79784ef53
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| import { homedir } from "node:os"; | ||
|
|
||
| export const DEFAULT_REPO_ROUTINES_DIR = resolve(import.meta.dir); | ||
| export const DEFAULT_RUNTIME_TASKS_DIR = join(homedir(), ".claude", "scheduled-tasks"); |
There was a problem hiding this comment.
The installer hardcodes ~/.claude/scheduled-tasks as the runtime path, but Desktop scheduled-task prompts can live under CLAUDE_CONFIG_DIR when that env var is set. In that environment this script writes to the wrong directory, reports success, and leaves the actual scheduled tasks unsynced, so routine updates/deletes in tools/routines will not take effect where Desktop reads them.
Useful? React with 👍 / 👎.
| return 0; | ||
| } | ||
|
|
||
| const results = routines.map((id) => syncRoutine(id, repoRoutinesDir, runtimeTasksDir)); |
There was a problem hiding this comment.
This flow only iterates routines present in tools/routines and never reconciles extra task directories already present in the runtime tree. If a routine directory is deleted from the canonical source, its old ~/.claude/scheduled-tasks/<id> prompt is left behind and can keep running, which violates the documented canonical→runtime source-of-truth model and creates stale autonomous behavior.
Useful? React with 👍 / 👎.
AceHack
added a commit
that referenced
this pull request
May 13, 2026
…rd + memory (#3041) * docs(rules): Otto inter-surface communication channels — reference card + memory Aaron 2026-05-13 asked Otto on both surfaces independently "do yall have a good way of communicating you should make sure and save it for future versions to remember." Two complementary observers landed independent partial lists; this PR synthesizes them as 8 channels in 2 classes (ambient vs explicit) and lands the substrate as both auto-loaded rule + detailed memory. Files: - .claude/rules/otto-channels-reference-card.md (auto-loaded; reference card) - memory/feedback_otto_inter_surface_communication_channels_8_channels_ambient_vs_explicit_aaron_2026_05_13.md (substantive empirical evidence) Ambient channels (state-of-the-world; both Ottos read continuously): Git, .claude/rules/ auto-load, Bootstream, Tick shards, Memory files, PR review threads Explicit channels (active signaling; meant to be observed by peer): Bus envelopes, Claim coordinator, Routines schedule, Aaron as ferry Per Otto on CLI 2026-05-13: "the bus is the explicit channel; git is the ambient one." Empirically validated: today's session exercised all 8 channels — 6 commits on PR #3034 across both processes (zero conflicts via rebase-on-pull), 3 memory files landed, PR #3032 rule auto-loaded for future, 9 bus envelopes scanned, multiple Aaron-as-ferry paste-relays, cross-lane PR thread resolution. Composes with PR #3032 (claim-acquire), PR #3036 (identity-stays-unified), PR #3037 (SENDER_IDS schema), B-0444 (bus envelope worktree field), and the existing wake-time-substrate / glass-halo-bidirectional / substrate-or-it-didn't-happen rules. Co-Authored-By: Claude <noreply@anthropic.com> * fix(rules): correct channel count (8→10) + commit count (6→9) consistency Codex P2 + Copilot P1 findings on PR #3041: - Carved sentence + 'all 8 channels' references inconsistent with enumerated list (actually 10: 6 ambient + 4 explicit) - 'Empirical evidence' section claimed 6 commits but commit table lists 9 Both factual inaccuracies. Aligned references throughout: - '8 channels' → '10 channels' in carved sentence + empirical-evidence header + memory description - '6 commits' → '9 commits (6 by Otto on Desktop + 3 by Otto on CLI surface)' in empirical evidence - 'all 8 channels exercised' → 'all 10 channels exercised' The 'Otto on CLI surfaced 6' / 'Otto on Desktop surfaced 8' counts in the inter-observer comparison are correct as-is — those refer to standalone observation counts per surface, not the combined synthesis (= 10). Co-Authored-By: Claude <noreply@anthropic.com> --------- Co-authored-by: Claude <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
tools/routines/as canonical-source-of-truth for Claude Desktop scheduled tasks (the "Routines" sidebar panel — same substrate as thescheduled-tasksMCP server)autonomous-loop— Desktop-side every-2-hour cold-boot tick, complementary to the CLI every-minute<<autonomous-loop>>cron sentineltools/routines/install.ts, Bun, rule-0 compliant) syncs canonical → runtime locationWhy
Aaron 2026-05-13 (this conversation arc, preserved in tick shard):
The
scheduled-tasksMCP server is cross-host — same server wired into both CLI Claude Code and Claude Desktop, both reading/writing~/.claude/scheduled-tasks/. UI Routines panel, MCP tools, raw-file edits, and now git-tracked canonical sources all route into the same disk substrate. Aaron's framing: "looks like there are multiple ways in here."Architecture
tools/routines/<id>/~/.claude/scheduled-tasks/<id>/tools/routines/install.tsEach routine:
SKILL.md(prompt body + YAML frontmatter) +schedule.json(cron + metadata).Three-layer catch-43 defence
* * * * *0 */2 * * *tools/routines/repoEach layer is a different distance from the operative tick — closer = faster recovery, farther = more durable across substrate failure modes.
Composes with
Test plan
bun tools/routines/install.tsfirst run:[updated](whitespace normalization vs MCP-generated file)bun tools/routines/install.tssecond run:[skipped-unchanged](idempotent verify)mcp__scheduled-tasks__list_scheduled_tasks→ autonomous-loop present,enabled: true,nextRunAt 2026-05-13T22:07:13Zbun test tools/routines/install.test.tsonce test suite landsTick shard
docs/hygiene-history/ticks/2026/05/13/2125Z.md
🤖 Generated with Claude Code