| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Three changes to the same layer.
I split testing.md the way I split commits.md, by what has to be resident rather than by size. What stays is the standing directive: TDD is the default, write the failing test first, and every unit needs Normal, Boundary, and Error cases. That has to fire before any code is written, which is exactly when no skill has been summoned, so it can't ride a trigger. Everything else moved to the testing-standards skill: characterization recipes, the per-category detail, property-based and mutation testing, the pyramid, integration rules, naming, the test-quality and mocking rules, coverage targets, the spike exception, and the anti-patterns. 2,824 words down to 347.
I fixed the approval gate in the publish flow. It decided whether to ask for approval by checking whether .ai/ is tracked, using that as a proxy for "team repo." The proxy was wrong in the direction that matters: rulesets, home, and work all track .ai/ while all three are private single-user repos, so the rule skipped the gate on the three projects I use most. It now checks whether any remote is on a host other than cjennings.net, which is the thing that actually decides whether someone else reads the log. Every current project resolves to gate-applies, which matches how the flow has actually been run.
I also added a first-person directive to the always-loaded core. One already existed for commit bodies and PR prose, but it moved into the publish skill with everything else, and it never covered code comments at all. Now everything I author in or about the repo is first person, with one carve-out: a comment describing what the code does stays third person, because there the code is the actor and not me.
Separately I split the publish skill internally. PR descriptions and the three review shapes moved to references/pull-requests.md, since a plain commit never needs them.
Always-loaded rules are now about 28,900 tokens, down from 57,800 this morning.
One risk on the record. testing.md's margin is thinner than commits.md's was. If testing-standards fails to trigger while I'm writing tests I lose the mocking-boundary rules, which is a quality regression rather than a permanent one, but it is a real bet where commits.md's was not. I also moved the TDD rationalization table rather than cutting it. The posts argue that kind of over-argument is counterproductive now, but removing your defense against me skipping TDD is your call, not mine.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
commits.md was 12,800 tokens, larger than the next four rules combined, and it loaded in every session in every project whether or not anything was being published. Most of it is procedure that only applies when a commit or PR is in play.
I split it by blast radius rather than by size. What stays always-loaded is what must hold whether or not I'm publishing, and where a violation is permanent and reaches other people: author identity, the no-AI-attribution ban, the generated-document byline rule, and the public-artifact content-scope rules. That core is 1,027 words, about 2,342 tokens.
What moved into the publish skill is everything about how a message gets written, reviewed, approved, and published. The pre-flight reconcile, the code-review gate, the draft and voice and approval gate, conventional-commit format, Voice and Focus, PR description structure, the three review shapes, merge strategy, and the pre-commit checklist.
The failure modes are why the line falls there. If the skill doesn't load I don't know the publish flow and have to be told, which is visible and recoverable. I don't silently commit with AI attribution, because that guard never moved. Only the recoverable half is on the skill-triggering bet.
I verified the split by using it. The skill registered mid-session and I invoked it to publish this commit, and it loaded with the flow intact.
Content is conserved: 5,561 words in, 5,898 across both files, the difference being the frontmatter and the pointer I added to the core. I repointed five cross-references in voice, review-code, inbox.org, and no-approvals.org that named sections which moved.
One thing I left alone. The skill is a single 4,871-word file, and both posts argue a long skill should split internally rather than sit as one blob. It loads on demand now, which is the win worth taking today. Splitting it further is its own change.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
wrap-it-up drops /tmp/ai-wrap-teardown-<project> to ask the Stop hook to kill the tmux session once the wrap certifies clean. When certification fails the Stop hook blocks and leaves the sentinel armed, which I did on purpose so a wrap blocked by a dirty tree can retry on a later stop without re-running the workflow.
I never bounded that retry to the session. An uncertified sentinel sat armed indefinitely and fired in whatever session next reached a clean tree. work's 11:37 wrap today left one armed. The 13:20 session ran startup, committed the task filing and the template sync, went clean, and the next stop consumed the two-hour-old sentinel and killed the terminal mid-work. Every stop before those commits had been blocked by the same sentinel failing certification, so the session spent its whole life either blocked or dead. archsetup's had been armed since Saturday on a live attached terminal, and home's was armed and waiting.
session-start-disarm.sh clears the project's sentinels at SessionStart. A new session means the wrap that armed one is gone, so its pending teardown is meaningless. Within-session retry is untouched, since the hook only runs at session start, and a test pins that. If teardown is still wanted, wrap-it-up re-arms it.
I disarmed the three live ones by hand before writing this, backed up under /tmp/disarmed-sentinels.
Four tests cover the disarm, one pins the retry behavior I did not want to lose. The scoping test matters most: a concurrent session in another project keeps its own sentinel.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
I added a Collaborative Peer Reasoning section to the interaction rules. It governs how an interpretation forms before any rule about presenting choices: infer first and clarify only at material forks, test a conclusion against its strongest alternative, let a correction update the downstream model instead of just the wording. The file's framing line widened to match.
Two probes were failing silently. The startup KB nudge looked up the best-practices node by grepping file content for its slug. A roam node's slug lives in its filename, so the lookup always returned empty. The nudge pointed at nothing in every project and every session, for as long as it shipped. It matches the filename now, through find rather than a glob so zsh doesn't abort on no match.
The browser rule told agents to open URLs with a form ending in &>/dev/null &. That discards the "Opening in existing browser session." line confirming the tab opened. Warm and cold start now split: foreground and read the confirmation when Chrome is already up, detach only when it isn't. The confirmation is on stdout, verified rather than assumed.
I filed two tasks from handoffs. The sentry triage split needs a work-vs-personal classification mechanism before its wording can move, because the current rule excludes by category and category can't express that split. The publish-lock design is approved but carries three open gaps. The load-bearing one is a lock held across an unbounded human approval pause.
I also swept the old processed handoffs out of inbox/. History keeps them.
|
| |
|
|
| |
Centralize repository-state checks, bind teardown to a certified clean HEAD, and allow inbox-only refreshes. Guard installed symlinks from cross-project writes and add regression coverage.
|
| |
|
|
| |
Clarify that speedrunnable means the solo autonomy tag; quick remains only an effort hint.
|
| |
|
|
|
|
|
|
| |
org-lint's invalid-block checker flags both delimiters of a paired example or src block whose body holds a heading-shaped line. Its parser reads the `** ` as a structural break and loses the open block. Every org file that documents org syntax inside a block collects two false findings.
The checker is org-lint's own, not one lint-org implements, so I filtered its output instead of patching a local checker. `lo--matched-block-regions` pairs blocks by scanning lines under org's real rule (once open, only the block's own `#+end_TYPE` closes it). A finding inside a matched region gets dropped, delimiters included, since that's where org-lint reports. I scanned lines rather than asking org because org's parser is what's confused.
An unterminated block never lands in a matched region, so it still reports. Three tests cover the false positives: a heading in an example block, a literal `#+end_example` inside a src block, and uppercase delimiters. Two more guard the real finding, including a file with one paired and one unterminated block that proves the suppression is per-block.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
An overnight hygiene run over this repo, reviewed adversarially afterward. Every problem below was reproduced before it was fixed, and every fix was re-checked by a skeptic who was told to refute it.
cj-remove-block could silently destroy content. Its range check validated only the first and last lines, so a span from one cj block's opener to a later block's closer passed and the removal deleted everything between — prose, headings, whole tasks, with a zero exit. That is the failure the check exists to prevent, and drift is its normal case, since the calling skill edits the file while processing. It now refuses a range holding more than one block, backs up first, and writes atomically while preserving the file's mode and following symlinks to the real target.
todo-cleanup rewrote todo.org with no backup, alone among the three tools that mutate these files. It now copies to the temp dir like lint-org does. Both backup helpers suffix rather than overwrite, because a second-resolution stamp collapsed two back-to-back invocations into one backup holding already-mutated content — and running todo-cleanup twice in a row is the shipped path.
audit.bats failed about one run in eight, in teardown rather than in its assertions, so a passing test reported as a false red. git commit spawns a detached maintenance process that is still writing packs when the commit returns, racing the cleanup. The fixture disables it.
route_recommend downgraded a correct strong match to weak when two projects shared a basename. lint.sh now sweeps the scripts that get symlinked onto PATH, which were the only shell in the repo with no gate over them.
Two test defects found in review and fixed here: a suite that deleted real backups from the shared temp dir, and one that depended on that directory's contents. Isolation now lives in an autouse fixture rather than in each test's memory.
Known and filed, not closed: the range check still cannot prove the range is the block that was scanned, so drift by exactly one whole block deletes the wrong annotation. That needs a content assertion and a CLI contract change.
|
| |
|
|
|
|
|
|
|
|
| |
Two additions from Craig, both landing on the throwaway branch that already makes a bad night a deleted branch.
Pass 11 now finds refactoring opportunities alongside bugs, filing each as a graded task. Still find-never-fix: a fresh finding files a task and stops, because a just-found bug can be a misread (one was filed and retracted two fires apart on 2026-07-23). The task is the checkpoint.
New pass 12 implements the backlog's solo, decision-free tasks on the branch, off by default behind :SENTRY_MAY_IMPLEMENT:. It reuses work-the-backlog's unattended-loop contract: only the ready bucket runs, anything needing a decision defers to a VERIFY, and nothing pushes. The morning review is the gate. Its /review-code runs the premise check first (reproduce the bug before judging the diff), because a filed claim never gets a review, and that skeptical review is what makes fixing on a branch safe. A finding that fails its premise check isn't implemented.
The marker is separate from :COMMIT_AUTONOMY: on purpose: hygiene is a two-minute merge, implemented code is a review session, so the owner opts in per project.
|
| |
|
|
|
|
|
|
| |
The five todo-format-family checkers encode todo.org completion conventions, but they ran on every org file, so they misfired across docs/specs/. That was 100 findings over nine specs in this repo, and the same noise in every project carrying specs. A spec's Decisions section legitimately uses a level-2 DONE with no CLOSED cookie. Its review-history uses level-2 dated headings. Neither is a completion defect there.
A path guard now skips the family on any file under a docs/specs/ segment, the canonical spec home per the docs-lifecycle rule. Link, table, and structural checks still run on specs, so a broken link in one still flags. todo.org and every non-spec org file keep the full checker set.
Folds in task-missing-last-reviewed alongside the four named checkers, since it's the same family and misfires the same way on a spec's phase tasks. Correcting an earlier note in the task: these are lint-org.el's own checkers, not org-lint's, so the fix is a local guard rather than filtering upstream output.
|
| |
|
|
|
|
| |
main caught ValueError and FileNotFoundError around the send, so an unreadable source raised a PermissionError as a raw traceback instead of the clean "inbox-send: <message>" every other failure produces. Widening the catch to OSError covers the unreadable source, the missing source, and any atomic-write failure alike.
discover_projects appended without deduping, so a roots config naming both a parent directory and one of its children listed the same project at two indices. It now dedupes on the resolved path, first-seen order preserved.
|
| |
|
|
|
|
|
|
| |
inbox-send wrote straight to the destination path, and write_text truncates on open, so any mid-write failure left a zero-byte .org in another project's inbox. inbox-status counted that phantom as a pending handoff and blocked a turn in the receiving project over a file with no content and no sender.
Both send paths now write to a temp sibling and os.replace it into place, so the inbox only ever sees a complete file. A caught failure removes the temp and re-raises, leaving no debris. encoding is pinned to utf-8 on the write and on the roots-config read, which closes the locale-dependent failure that first surfaced this.
inbox-status also skips the .inbox-send-* temp, so the brief window before the rename can't read as pending either. The atomicity is only complete once the consumer ignores the in-flight file.
|
| |
|
|
|
|
|
|
|
|
| |
The python and typescript bundles carried rules and a coverage script but no pre-commit hook, so any project installing one got no credential scan on commit. Both now ship all four components the README documents: the shared secret scan, a validate-on-edit hook, settings wiring, and a seed CLAUDE.md. Verified against a real repo: a commit carrying an AWS key is refused.
install-lang now warns when a bundle is missing a documented component. That's the half that keeps this from recurring. Whoever adds the sixth bundle will forget something too, and today the installer prints success either way.
Two things fell out of the build. node --check is unusable on TypeScript: it ignores --experimental-strip-types, so it rejects valid TS and accepts broken TS. The hook uses tsc filtered to syntactic diagnostics instead. And completing the bundles means every pair now collides on settings.json and pre-commit, so no two compose without FORCE=1. The earlier "bundles already compose" reading rested on these two being incomplete. Filed for a real decision; clock-panel is the project that wants it.
Also stamps :LAST_REVIEWED: at task creation, with a lint checker to catch misses. Writing a task is reviewing it, so leaving the stamp off pushed every fresh task to the top of the next review batch to be re-derived by someone with less context than its author had. Tonight's sweep sent the staleness count from 13 to 22 while the list got more accurate.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
Four proposals from other projects, reviewed and applied together.
wrap-it-up gains a Clear temp/ step in Step 3. temp/ is gitignored everywhere, so nothing there survives deletion. The step confirms before removing anything that reads as in-progress rather than throwaway, and skips where temp/ isn't gitignored. This closes the last open clause of the working/temp work from 2026-07-20.
interaction.md drops the fenced-code-block carve-out. Chat output carries no markup at all now, fences included. Fences don't invert the way inline spans do, but the carve-out kept pulling them back in.
sentry.org folds in four notes from its first two live runs elsewhere. Two are traps for anyone reviewing a night by git log: archive-done touches .gitignore on its first run, and a mirror-only project's quiet fires leave no commits at all. The others make property sweeps sanctioned quiet-fire work and split the task audit into a mechanical hourly subset plus a nightly judgment half.
The notes.org template now lints clean, down from four flags in every project on every sweep. Two of those were mechanical, so lint-org --fix anywhere would have rewritten the template and drifted it from canonical.
|
| |
|
|
|
|
|
|
|
|
| |
Sentry no longer checks email or messengers. Pass 3 loads only the calendar and PR/ticketing triage plugins. The mail and messenger ones stay out: cmail, Gmail, Telegram, Signal, chat DMs. A manual triage intake still scans everything. The probe follows, so a project whose only declared sources are mail or messengers now skips the pass instead of running it empty.
New pass 11 hunts bugs. It runs linters and static analysis, checks config sanity, and reads one rotating code area per fire so coverage builds over a night. Verified findings get filed as graded bug tasks. It never fixes unattended.
I dropped the suite run the order called for. Running the suite every fire is the per-pass anti-pattern this same file forbids, and the entry baseline already has the result.
The KB personal-project roots now cover ~/.dotfiles, which was classifying Unknown and blocking writes from there.
|
| |
|
|
|
|
|
|
| |
Auto mode is the away-from-desk mode. A sweep worth seeing should reach Craig's phone, not just print into a session he isn't watching. I folded a Phone delivery subsection into canonical triage-intake.org: a full-three-section sweep now pushes to the phone over Signal via agent-text. A pointer from End-of-sweep output and a Living Document note tie it in.
Signal-only, by Craig's ruling: a quiet sweep's "nothing" heartbeat never reaches the phone. Silent-until-signal governs the phone channel too, so it only fires on real signal. The in-session heartbeat stays as proof the loop ran.
Only the send half ships. Reply polling (the retired ntfy design's phone-recv) waits on the reply-correlation follow-up: a Signal reply fans out to every linked device and neither knows which page it answers. That spec owns the recv wiring. I re-pointed the transport off dead ntfy onto agent-text (the renamed pager). The mirror is synced and the DONE task records the deferral.
|
| |
|
|
|
|
| |
"page me" and "text me" now name distinct channels: page is the desktop notification, text is the Signal phone push, and "text and page me" fires both. Before, "page me" defaulted to desktop and the phone was an unnamed "on my phone" variant, so you couldn't name the phone without naming the machine.
I renamed the Signal tool agent-page to agent-text to match, with a deprecated agent-page shim delegating to it so callers and other machines don't break before their next install. The dispatch and tests are otherwise unchanged. I rewrote protocols.org "Reaching Craig", page-me.org, and work-the-backlog's away-run logic around the three phrasings, and updated the runbook.
|
| |
|
|
|
|
|
|
| |
The pager worked but was undocumented and drifting stale: signal-cli warned the account had gone 47 days without a receive, and no runbook existed to hand a future session or a non-Claude runtime.
I added the runbook under docs/design/ and a roam-sync-shaped receive timer that drains the queue every 15 minutes, keeping the account warm and surfacing reply messages. I also generalized agent-page: it now sends directly from any machine holding the pager account and relays to velox only when it doesn't. ratio is linked as a device of the account, so a page lands even when velox is down.
The receive script no-ops cleanly where the account isn't registered, since the units stow onto every machine via the shared common package.
|
| |
|
|
|
|
| |
The general triage plugins (personal Gmail, cmail, calendar, Telegram, GitHub PRs) are template-synced into every project, so the old "plugins present" probe self-activated triage everywhere. The sentry live trial caught it reaching personal inboxes from a project that isn't a triage target. Now a general plugin activates only when the project names its basename in a :TRIAGE_SOURCES: line in notes.org Workflow State. A project-specific plugin stays active by presence. The gate lives in triage-intake Phase 0, so it fixes the interactive over-pull too, and sentry's pass-3 probe reads the same signal.
Presence is capability, the declaration is activation. A project that declares nothing and owns no project plugin pulls nothing. Migration handoffs went to home and work. rulesets correctly declares nothing.
|
| |
|
|
|
|
| |
Auto triage-intake and auto inbox-zero now collapse an empty check to a single "<workflow> at HH:MM: nothing" line, matching the sentry quiet-fire heartbeat from Phase 1. Detection stays in-session, so the MCP-auth sources are scanned in full and only the output collapses. Inbox already had an acknowledge-only-on-empty rule, now unified to the standard format; triage's three-section sweep output gains the empty-sweep collapse.
Completes the silent-until-signal spec (Phases 2-5): the spec is the shared-policy home the three workflows reference, and a manual-testing entry covers the live-loop verification. Spec flipped to IMPLEMENTED.
|
| |
|
|
|
|
| |
A suspend now ends by detaching the tmux client viewing the aiv-<project> session, so the parked session drops out of the alt-space rotation instead of cluttering it. Detach is not teardown: the session and agent process stay alive in the background and only the view disconnects, so it runs inline as the last action (the handoff text is preserved in the pane and shown on re-attach) rather than deferring to a Stop hook the way the session-killing teardown must. Degrades gracefully when not in tmux.
Originated in archsetup as a local stopgap and reconciled to canonical here.
|
| |
|
|
|
|
| |
A sentry fire whose passes all probe-skip or no-op now writes a single "sentry at HH:MM: nothing" line instead of a full per-pass digest block. A fire that runs or queues anything writes the full digest as before. This is Phase 1 of the silent-until-signal policy. The live trial's fires 3-8 were walls of no-op digests, the noise this removes.
The "no silent skip" discipline is reconciled, not broken: inside a working fire every pass still writes its ran/skipped line, and a quiet fire's heartbeat is the explicit "nothing to do" record rather than a hidden skip.
|
| |
|
|
|
|
|
|
|
|
| |
working/ is the tracked home of in-progress work, version-controlled from creation rather than staged locally until it graduates. Filing on completion reorganizes durable artifacts into permanent homes. It doesn't mark when they became durable. Ephemeral, disposable artifacts go in a gitignored temp/ (or /tmp), never working/.
install-ai.sh emits a temp/ ignore in both track and gitignore modes. Ephemerality is independent of whether a project tracks its .ai/ tooling, so temp/ rides neither the track .gitkeep step nor the gitignore-only tooling block. working/ is never ignored, on purpose. sweep-gitignore-tooling.sh backfills temp/ as a separate mode-independent pass, not an IGNORE_SET member: folding it into that set would skip every track-mode project, the set that most needs it.
While implementing I confirmed the canonical machinery never ignored working/ (the tooling set is only .ai/ .claude/ CLAUDE.md AGENTS.md), so a project that ignored working/ did so as a local deviation.
Documents the convention in working-files.md and mirrors it in protocols.org. 7 new bats cover temp/ in both modes, never working/, and idempotency.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
convention
Home proposed two linked cross-project norms and Craig approved rolling them out. Both now ship in the template so every project inherits them.
A new "Colloquialisms and Expansions" section in protocols.org maps Craig's shorthand to a defined action the agent applies without asking. Two seed entries: "put X on the list" appends to a session-scoped Before-Close Queue, and "tell <project> <message>" drops the message in that project's inbox via inbox-send. The set is extensible.
The queue lives under a "* Before-Close Queue" heading in the session anchor, so it resets when the anchor is archived at wrap. Anything that must outlive the session stays a todo.org task. wrap-it-up Step 1 gets a sub-step that works the queue oldest-first before the Summary, so any resulting edits ride the wrap commit, with unfinished items surfaced in the valediction rather than dropped.
I settled three design calls. The reference lives in protocols.org rather than a per-project notes.org (home's local reference), since a synced section is what makes it a shared norm. The queue lives in the session anchor, as home did. The wrap step sits at the front of Step 1 rather than a new half-step, which keeps the "Steps 1 through 5" framing intact. A documentation-integrity bats guards the wiring and the before-the-Summary ordering against a future edit or sync dropping it.
|
| |
|
|
|
|
|
|
| |
The "check inbox/ at every task boundary" rule was prose-only in protocols.org, so it held only as well as the agent remembered it and a handoff could pass a turn unseen. inbox-boundary-check.sh is a Stop hook that backs the rule: when inbox-status -q reports pending items it blocks the yield once and injects the count, so the agent processes them before returning.
It soft-nudges rather than hard-blocks. On the harness re-entry (stop_hook_active: true) the hook steps aside and lets the turn end, so a mid-task pause to ask a question, or an item the agent genuinely can't process, never wedges the session. A hard block would instead push inbox processing ahead of that clarifying question.
The hook self-skips where it doesn't apply: no inbox/ dir, no inbox-status, or a clean inbox each exit 0 silently, so one global hook covers every project with no config. It prefers the project-local .ai/scripts/inbox-status and falls back to one on PATH. It's wired into the Stop array ahead of ai-wrap-teardown in both the tracked settings.json (which the live ~/.claude/settings.json symlinks to) and the documented snippet. bats covers pending, clean, the re-entry step-aside, no-inbox, and the absent-status degrade.
|
| |
|
|
|
|
|
|
| |
I moved the decision logic out of the launcher's I/O-coupled functions so it's unit-testable on plain inputs, the hardening the audit called for. Three pure cores now carry logic that was buried in git and tmux calls: _git_prep_action classifies a repo's state into none/pull/report, _order_windows sorts the window listing, and _match_window_id resolves a name to a window id. _git_is_dirty replaces the staged/unstaged/untracked check that was copy-pasted across three functions.
prep_git_single and auto_pull_if_clean both route their pull-or-report choice through _git_prep_action, so the "never auto-pull a dirty repo" rule lives in one place with its own test instead of two parallel branches. sort_windows keeps its two-pass move (park high, then reassign) but hands the ordering to _order_windows, and find_window_id is now a thin tmux wrapper over _match_window_id.
Each core gets a Normal/Boundary/Error set: the git-prep truth table including diverged and behind-while-dirty, the window ordering including whole-line-not-substring matching. The 42 launcher tests stay green and a live black-box run still prints the right launch command for claude and codex, so the extraction is behavior-neutral. I also ran shfmt (-i 2 -ci) to match the house style the other bin/ scripts follow.
|
| |
|
|
|
|
|
|
| |
Brings the session launcher under characterization tests ahead of the hardening refactor, so the pure-core extractions and footgun fixes land against a green net. The tmux/git-coupled functions can't be reached by running the launcher black-box, so I wrapped dispatch in main() behind a run-vs-sourced guard: run directly it dispatches as before, sourced it exposes the functions to the tests.
The net pins current behavior (record-not-spec) with a Normal/Boundary/Error set per unit: git_status_indicator across no-upstream/ahead/behind/dirty/clean, candidate discovery and filtering, selection parsing, and functional tests over create_window, find_window_id, and sort_windows ordering against a throwaway tmux server. Every tmux call runs on a private socket so nothing touches the live ai session, and the throwaway repos disable git auto-maintenance to dodge the teardown race that flaked rename-ai.
Also cleared the two pre-existing shellcheck warnings on the touched paths: quoted the "@{u}" revspecs (SC1083) and marked the intentionally-literal display tilde (SC2088). Both are behavior-neutral, and the existing runtime tests confirm the launch path is unchanged.
|
| |
|
|
|
|
|
|
|
|
| |
A task phrased as an absence ("find bugs until none remain," "refactor until nothing worthwhile is left," "clean it up") can't be :solo:, because it fails the verifiable-by-the-agent gate. Absence isn't falsifiable, so "done" is a judgment call, which is what :solo: forbids. This kept surfacing (the ai-launcher-hardening task is the live case) with no written recipe for fixing it.
todo-format.md gains a subsection, right after the :solo:/:quick: definitions, on converting a fuzzy goal into a measurable one: bound the surface (enumerate the units, that's the done-set), net the behavior with characterization tests, disposition every finding (fixed / filed / declined-with-reason, never "looked and it's fine"), and gate on an objective floor (static-analysis clean, suite green, coverage of the enumerated surface). The qualifying answer becomes a dispositioned report, not a claim of perfection, and the honest limit stays honest: "no visible bugs" means "every enumerated path passes its characterization set and clears the audit." Once a task carries those criteria it can carry :solo:.
work-the-backlog's keystone defer item now names open-ended goals as a specific, recognizable failure of test-writability, so a mistagged :solo: task with no writable acceptance test is deferred and routed to get criteria rather than guessed at.
start-work has the same Justify/Approach gate and wants the same pointer, but its skill lives outside rulesets, so that edit is a follow-up in its home.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
quality sweep
The rule already said "write a characterization test before refactoring" but not what one is or how many to write per unit. Two additions close that.
testing.md's "Adding Tests to Existing Untested Code" now defines a characterization test as recording what the code actually does, not what it should (Feathers' assert-wrong-read-the-real-value recipe), and requires the same Normal/Boundary/Error set as any unit rather than one happy-path capture. The reason is the useful part: on a characterization test the negative and boundary cases are the bug-finders, because untested legacy code is weakest at the empty input and the missing upstream, and pinning what it currently does there writes the wrong behavior down where it becomes a visible bug. A pinned case that turns out to be a bug graduates from recording current behavior to asserting correct behavior. In-unit bugs fall to this set. Composition bugs need a functional test.
The refactor-for-testability section gains the framing that when the untestable code is legacy you're hardening, extracting the pure core is the hardening, not a detour around it. A function you can't characterize without mocking tmux or git can't be refactored safely, so "needs too much mocking" is the signal to reshape it, never a license to skip the boundary and error cases.
code-quality.org's passes claim to preserve behavior, so it now states the precondition that makes that true: on untested scope, a characterization net comes first, or "behavior-preserving" is an assertion the green suite can't actually verify.
|
| |
|
|
|
|
|
|
|
|
| |
Craig wanted a curated "DeepSat Fundamentals" deck: the ~100 most fundamental cards out of the 465-card deepsat.org deck, marked with a second org tag (:fundamental:) so they stay grep-able in the source. Both flashcard-to-anki.py and flashcard-stats.py keyed a card on a heading ending in exactly :drill:, so adding any second tag turned the heading into :fundamental:drill:, which that anchor didn't match. The 100 tagged cards would have silently dropped from the full-deck apkg and been undercounted by stats. That's the passing-gate-skips-your-file failure.
CARD_RE now matches a trailing org tag block, and a heading is a card when drill is among its tags. The other tags ride along as Anki tags next to the section tag. Card bodies are bounded by any L1/L2 heading (HEADING_RE) rather than only a section or a drill heading. flashcard-to-anki.py gains --tag-filter <tag> (emit only cards carrying that org tag, the subset deck) and --guid-salt <s> (a distinct GUID space so the subset imports non-empty instead of colliding with the full deck on GUID and importing empty). No salt is unchanged, so the existing deck's GUIDs and SRS state are untouched. flashcard-stats.py gets the same CARD_RE/HEADING_RE broadening plus a drill-membership guard.
This is the rulesets-canonical reconcile of a stopgap the work project ran locally. I re-derived it against the current canonical rather than copying the preserved files, which predate the #+TITLE deck-name fix (060a938) and would have reverted it.
parse()'s third element is now the Anki-tag list rather than a single tag string. Only build() and the tests consumed it. The existing parse tests move to the list contract, and new tests cover multi-tag parsing, --tag-filter, the guid salt, and the stats count. A bats case guards that a :fundamental:drill: card still counts. The full deck's real 465/100 check needs the work deck and isn't runnable here. Verified end to end instead on a synthetic multi-tag deck (2 cards, --tag-filter fundamental returns 1) through the real genanki path.
|
| |
|
|
|
|
|
|
|
|
| |
The flashcard pipeline was one-directional (org-drill to apkg). A deck curated on the phone, or an orphaned apkg whose org source was never saved, had no way back into the org source of truth. apkg-to-orgdrill.py closes the loop.
An apkg is a zip holding an Anki sqlite collection, so reading it needs no third-party library: stdlib zipfile + sqlite3. genanki is only needed to write apkgs, not read them, so the converter has no runtime deps. It mirrors flashcard-to-anki.py inverted: deck name to #+TITLE, note Front to the ** heading, note Back HTML to the body (<br> to newlines, entities unescaped, the answer <hr> stripped), the note tag to a * section, and a fresh :ID: UUID per card so the output is org-drill-valid.
The entity unescape runs in the order that inverts escape_html, which escaped & first, so & is unescaped last here. Otherwise a literally-escaped < in the source would wrongly collapse to <. Only Front/Back note types convert; a cloze or other model is skipped with a warning rather than silently dropped, and a note referencing media is flagged since the org side has no media path.
17 tests cover the HTML-to-org conversion, the apkg read (single and multiple decks, Default-deck exclusion, non-basic skip, .anki2 vs .anki21, media flag), the org emission, and errors. The round-trip test closes the loop through flashcard-to-anki.py's own parse(): a known org source becomes an apkg fixture, converts back, and the recovered (front, back, tag) tuples match the original. Verified end to end against a real genanki-produced apkg too.
|
| |
|
|
|
|
|
|
| |
The sentry live trial surfaced this. A sentry inbox-zero pass running from the work project parked the whole 19-item roam inbox as a cross-project boundary crossing and refused to tidy it unattended, reading knowledge-base.md's work-denylist as "don't touch roam from work."
That over-reads the rule. The denylist is a confidentiality guard on one thing: promoting a durable fact into a new agents/ node, so work-confidential material stays out of the personal cross-machine store. Roam is a shared resource, not another project's product scope. Reading it and tidying the shared roam inbox are housekeeping any project may do, work included.
knowledge-base.md gains a "Scope of the denylist" paragraph stating that and naming the mis-park as the error it closes. sentry.org's inbox-zero pass and inbox.org's roam mode each get a one-line note pointing at the rule, so a future agent doesn't re-derive the refusal. Only durable-node promotion stays work-denylisted.
|
| |
|
|
|
|
|
|
|
|
| |
Phase 3 of the sentry supervisor: reconcile the existing roam writers and wrap-up so sentry's locks actually guard something, and its shutdown has one enforced entry point.
The roam-write lock only helps if every roam writer takes it, so both writers now do. knowledge-base.md's write recipe and inbox.org core §5 acquire the roam-write lock around their edit and trigger roam-sync instead of committing themselves. That closes a gap the one-git-owner rule already implied but the KB recipe still violated: the recipe told agents to run git add -A && commit && push against a tree that's chronically dirty from live captures, which could sweep an in-flight capture into a stray commit. roam-sync stays the roam repo's only committer. Agents edit-plus-trigger under the lock, and roam-sync.sh's header now states that contract instead of the old "agents commit inline" note.
Both writers degrade as the spec settled: an absent agent-lock proceeds unlocked (today's behavior), and only a present helper reporting the lock busy after its bounded wait defers or surfaces.
wrap-it-up.org gains a Step 0 that refuses while sentry is live. It checks the single-runner lock and points at "stop sentry" rather than archiving the anchor and tearing down the buffer under a still-firing loop. Both files derive the lock name the same way (sentry-<repo-basename>), so the guard and the engine agree. triage-intake.org notes that it also runs as sentry's triage pass under the no-approvals contract, with its trigger phrases unchanged.
|
| |
|
|
|
|
|
|
|
|
| |
Phase 2 of the sentry supervisor. sentry.org is the engine: an interval loop (default hourly) that walks a fixed pass list (roam pull, inbox zero, triage, todo cleanup, task audit, working-files hygiene, spec board, link integrity, git health, prep freshness) and commits each writing pass to a throwaway sentry/<date>-<host> branch. Nothing pushes. The morning teardown (review, squash-merge, delete) stays Craig's.
The file owns the whole engine: the :COMMIT_AUTONOMY: entry ticket, the interactive entry gates (clean tree, green suite, prior-branch check, ff-only reconcile) that run with Craig present, the single-runner and roam-write locks via agent-lock, the probe-work-session-context-commit pass contract, the digest and morning-approval queue, the skip-not-degrade safety model, the conditional fire-end suite, the multi-day stall notification, and the stop-sentry shutdown.
Every one of the spec's ten decisions and twelve review findings is reflected: host-suffixed branch so the two daily drivers don't collide, tmpfs lock homes so a lock never rides roam-sync to the other machine, roam-sync kept as the roam repo's only committer, spine-excluded dirty checks so sentry's own bookkeeping can't trip its safety skip, and the queue-don't-execute rule for every judgment or destructive action.
Phase 3 wires the roam writers (inbox.org, knowledge-base.md) to acquire the roam-write lock and adds the wrap-it-up guard. Until then the engine is documented but the companions don't yet cooperate. INDEX.org lists it under Tasks and planning.
|
| |
|
|
|
|
|
|
|
|
| |
Phase 1 of the sentry supervisor workflow. sentry needs two locks: a single-runner lock so overlapping /loop fires don't stack, and a roam-write lock so agent edits to the shared roam tree serialize. flock can't serve either, because every Bash call an agent makes is its own short-lived shell, so an flock dies with the call that took it.
agent-lock persists the lock on disk between calls. An atomic mkdir is the acquire. A metadata file inside records pid, host, ISO timestamp, and TTL. Staleness is age-based on the meta mtime, so a crashed holder's lock expires instead of wedging every later acquire, and every reclaim surfaces a note rather than clearing silently. refresh re-touches the mtime as a heartbeat, release is idempotent, status reports free/held/stale, path prints the resolved dir. Locks live under XDG_RUNTIME_DIR/agent-locks (tmpfs: host-local, out of every repo, cleared on reboot), with a ~/.cache fallback where no runtime dir exists.
Reclaim claims the stale dir with an atomic rename before removing it, so two acquirers that both see a lock stale can't both win. A plain remove-then-mkdir would let the loser delete the winner's fresh lock and double-acquire, defeating the roam-write lock's purpose.
Nothing calls it yet. 18 bats tests cover fresh win, contention, release, stale reclaim, heartbeat, path resolution, and usage errors.
|
| |
|
|
|
|
|
|
| |
Two batched changes to the todo hygiene tooling.
Dated-seal archiving: --archive-done aging now keeps one month in-file (retain default 7 → 31, named as tc-archive-retain-days-default) and a new --seal renames the working task-archive.org to resolved-YYYY-MM-DD.org beside it, letting the next aging start a fresh working file. Dating the sealed file by the seal run rather than the calendar quarter is what avoids mislabeling a task closed late in a quarter and archived after the boundary. The unparseable-CLOSED case is now explicit in the contract: a keyword-complete task with no readable close date ages out too. The sealed file inherits the todo file's gitignore status.
Planning-line strip: --convert-subtasks now drops the whole planning line when it rewrites a done sub-task to a dated event-log entry, not just the CLOSED cookie. A dated-log heading carries its date in the heading, so an active SCHEDULED/DEADLINE left on it pins the finished entry to the agenda as weeks-overdue forever (org renders any active planning timestamp, keyword or not). New lint-org checker dated-log-heading-active-timestamp backstops any that predate this. todo-format.md's sub-task and VERIFY completion rules gain the strip step.
|
| |
|
|
| |
Added claude-templates/bin/install-ai, a thin launcher that resolves its own path through the symlink chain and execs scripts/install-ai.sh. make install's bin loop symlinks it into ~/.local/bin/install-ai, the same mechanism that links ai and agent-page, so the fresh-project bootstrapper now runs as install-ai from anywhere. The symlink always points at the canonical, so no separate dotfiles copy is needed. Three launcher bats cover the wrapper, including invocation through a symlink from another directory (the case a naive dirname "$0"/.. would break).
|
| |
|
|
|
|
|
|
| |
New stdlib-only upcoming_birthdays.py reads ~/sync/org/contacts.org for :BIRTHDAY: properties and prints a daily-prep block: name, date, days-away, and the age turned when the birth year is known (1900 is the org-contacts unknown-year placeholder, rendered date-only). Birthdays inside 7 days come back flagged so a gift or plan gets prompted. 19 pytest cases cover the window boundaries, leap-day fallback to Feb 28, the placeholder year, and the CLI.
daily-prep gains a Phase A source that runs the script and a Heads-Up line that folds in the block, so contact birthdays the calendar doesn't carry still surface.
Built and tested in the home project. Promoted here because the daily-prep hook lives in this synced workflow and the contacts file is present on every machine.
|
| |
|
|
|
|
| |
Phase C now renders a TASKS / FYI / MISC digest with a two-option close offer. The per-source breakdown and the itemized action menu move to on-request. Phase D becomes the close and runs by default after the digest: file the TASKS items, run mail hygiene on every scanned account without itemized confirmation, advance the sentinel, tear down anything the sweep started. Prose sends under my name and destructive non-mail actions stay gated. The reroute modifier delivers outside-project items to their owners via inbox-send. The stay-open-until-confirmed exit loop is retired. I updated daily-prep's Sub-step 3b to match.
I ratified this in the work project on 2026-07-18 and applied it here as sent from that handoff (inbox 2026-07-18-1730-from-work).
|
| |
|
|
|
|
|
|
| |
The residue probe cut off at before:<anchor-YYYY/MM/DD> while the scan started at after:<anchor-epoch>. Gmail's before:<date> excludes the named day, so the anchor day fell between the two queries. Each sweep then advanced the anchor past that window and never looked back, so an evening anchor hid most of a day for good. That's the exact failure the probe was added on 2026-07-08 to prevent.
The probe now cuts off at the epoch, matching the scan. I hoisted the epoch rule to cover both anchored queries once rather than twice. I scoped it to the anchor windows so the date-slice walk's deliberate day resolution doesn't read as a bug.
home caught it: after acting on 38 messages the probe reported zero while a plain is:unread in:inbox still returned two messages from the anchor day. personal-gmail is the only source pairing an anchored scan with a residue probe, so nothing else needed the fix.
|
| | |
|
| |
|
|
|
|
| |
The meeting recorder switched to lossless FLAC on 2026-07-01, and the transcript workflow's scan globbed only .mkv and .m4a, so a .flac recording was silently never found. FLAC now threads through the audio-only path: found by the scan, transcribed natively by AssemblyAI with no ffmpeg extraction, kept on cleanup, and archived with its own gitignore pattern.
Adopted verbatim from work's handoff (2026-07-14-0105), diff-verified against the canonical.
|
| |
|
|
|
|
| |
Both helpers treated any pipe-led line as a table row, so ASCII art in an example block got mangled into a bordered table (a work file took a 1949-line reformat on 2026-07-09). The scanners now track the open block's type and skip its content. Only the matching end marker closes a block, so a literal end_src quoted inside an example block can't re-expose it.
The deeper cause was wrap-org-table's load-time dispatch: it fired when lint-org merely required the library, running the reformatter over files lint-org was only asked to report on. It now dispatches only when named as the entry script. lint-org's CLI is also report-only by default now. Writes require --fix, and the wrap-up workflow and lint command pass it.
|
| |
|
|
| |
The interactive flow now opens with a runtime picker: claude first (Enter-Enter preserves the old two-keystroke habit), codex labeled as ChatGPT, and one local:<model> line per ollama model, queried live with a short timeout so a dead server just drops those lines. The pick feeds the existing runtime map, then the annotated project multi-select runs as before. --runtime and AI_RUNTIME skip the question, single-directory mode is unchanged, and --print-runtimes exposes the choice list for the two new bats tests.
|
| |
|
|
| |
ai --runtime local now launches codex against the machine's ollama (explicit --local-provider=ollama, model from AI_LOCAL_MODEL, default gpt-oss:120b). Verified end to end with a codex exec completion through the local model. The dependency check probes AGENT_BIN now that AGENT_CMD carries flags. This makes codex-over-ollama the de-facto answer to the runtime spec's first-supported-local-CLI blocker.
|
| |
|
|
| |
The Signal pager reconcile found one pager identity, registered in velox's signal-cli, reachable from other machines only by hand-rolled ssh. agent-page wraps that: direct signal-cli on velox, ssh relay over the tailnet from everywhere else, a desktop-fallback hint when the relay fails, and the UUID target baked in so nobody pages the unregistered phone number again. Four bats tests stub ssh/uname/signal-cli to verify command construction; the real path was verified with a live phone push today. protocols.org's Paging Craig section now teaches both channels (notify for the desk, agent-page for the phone) and demotes signal-mcp to a velox-local nicety; page-me.org, work-the-backlog, and the INDEX carry the same two-channel story. Every project inherits on its next startup sync and make install.
|
| |
|
|
| |
The ai launcher can now start codex sessions: --runtime claude|codex (or AI_RUNTIME), a runtime-to-CLI map with local reserved behind a clear error until the local-model evaluation picks a CLI, and a runtime-aware dependency check. A new --print-launch mode prints the pane launch command without touching tmux or fzf, and is the seam the six new bats tests drive. Both current CLIs take the opening line as a positional prompt, so only the command name varies.
|
| |
|
|
| |
claude-templates/AGENTS.md points any harness at protocols.org, the rules locations, and the /name-to-file resolution rule, and tells it to degrade per each rule's documented fallback instead of skipping gates. make install links it to ~/.codex/AGENTS.md (new CODEX_DIR stanza) and install-ai.sh seeds a project-owned copy at bootstrap, never overwriting an existing one. The rulesets root carries a tracked symlink as dogfood. Five new bats tests cover the link idiom and the seeding. This resolves the generic-instruction-file blocker from the runtime spec, and velox picks up the global link on its next session's make install.
|
| |
|
|
| |
interaction.md gains a Showing Craig Visuals section: EAT renders SendUserFile and inline images as a text line, so visuals open in the browser or imv (per desktop-capture.md), with inline sixel as the quick-glance lane behind a capability gate (patched EAT answering CSI 14 t plus tmux terminal-features sixel, diagnosed 2026-07-13). protocols.org gains the zsh no-word-split gotcha next to the exa alias note. takuzu sent the proposal as inbox handoff 2026-07-11-2137; the sixel findings correct its capability claim.
|
| |
|
|
|
|
| |
task-review-staleness.sh expected a bare 2026-07-09 and treated an org inactive timestamp ([2026-07-09 Thu], the form matching the CREATED:/CLOSED: cookies in the same drawer) as unparseable. The count branch folded that into the stale count, so a freshly-reviewed task reported as never-reviewed, and a full review pass never dropped the startup nudge. Both the bare and bracketed forms now normalize to the ISO date. A value that is neither warns loudly to stderr (file:line:value) and stays out of the count, since a data error shouldn't hide as "never reviewed." task-review.org documents the accepted format.
Tested red/green: 5 new bats cases (bracketed fresh and stale, list-mode sort by real date, malformed warns and is excluded). Full suite 273 ok.
|
