# Orchestrator capabilities by class Date: 29 October 2026. Source revision: `b424656e1daff0afe8598162a4cce9966387c884`. [Italian version](2026-10-29-orchestrator-features_it.md). ## Scope and evidence The orchestrator is a workspace, operating rules, skills and scripts. TUICommander supplies sessions, mail, repository actions and scheduled jobs. Application development takes place in the target repository, not here. This is a source inventory, not a live health certificate. **In use** means a current rule, configured hook or source caller selects the capability. **Partial** means an optional integration, an external installation not checked here, or a stated limitation. **Unused** means not selected by the named current entry point, not proof that nobody ever calls it. **Broken** requires decisive failure evidence; this audit did not execute operational commands to find failures. A test file is evidence of intended protection, not evidence of a passing run today. MDKB was queried first: `mdkb search 'orchestrator tools capabilities operations hooks' --root orchestrator --limit 15`. It returned related reports and `orchestration.md`, not a complete tool inventory. The audit therefore inspected the requested rules, decisions, skill instructions, hook configuration, and script headers, commands, functions and callers. Tests and fixtures are validation support, not separate user capabilities. Generated `.wiz` state and Python bytecode are not entry points. Paths below are relative to the repository root, except named home-directory policies. The entry point is also the implementation evidence. The final column records the caller, rule or limitation. “Manual” means the coordinator or an authorized agent invokes it; it does not mean Boss must run it. ## 1. Agent lifecycle | ID | What it does | Entry point | When | Status and evidence | |---|---|---|---|---| | A01 | Start or resume an agent with a complete brief, parent identity and optional isolated worktree. | `tools/agents/tuic-spawn.sh`, `tools/agents/spawn-roles` | Manual, each assignment | In use: README; role admission, resume, brief validation and child registration are implemented; policy mismatch noted in I4. | | A02 | Deliver work to an idle agent, or send mail when it is busy; resolve its stable name. | `tools/agents/tuic-say.sh` | Each assignment or course change | In use: `operations/agent-work.md` requires this delivery path. | | A03 | Run a long command outside the turn and wake the caller with its exit result. | `tools/agents/tuic-bg.sh`, `tools/agents/bg_wakes.py` | Commands over one minute | In use: managed rules require it; wake receipts protect children from premature closure. | | A04 | Serialize ownership changes and recover interrupted child closure or checkout renames. | `tools/agents/child_state.py` | Spawn, associate, reconcile, close | In use: spawn, cleanup and child wrappers call the locked atomic writer. | | A05 | Distinguish pre-existing checkout changes from work written by the child or a named other owner. | `tools/agents/checkout_baseline.py` | Before spawn; handover and shared-checkout edits | In use: spawn snapshots/binds entries; child and landing writers use the baseline. | | A06 | Close an agent only after recording its successor; protect unsaved work and running test apps. | `tools/agents/child-close.sh` | After RESULT | In use: calls the durable close transaction and checks agent-test processes. | | A07 | Record a next action while reusing a child, without pretending the new brief was delivered. | `tools/agents/child-continue.sh` | After RESULT | In use: explicitly prints “NOT delivered”; A02 still sends the work. | | A08 | Close finished idle children but keep blocked children, protected sessions and live background work. | `tools/agents/close-idle.sh` | Pump housekeeping; default idle window 15 minutes | In use: `pump_housekeep.py` calls it. | | A09 | Show live, missing and stalled sessions, including terminal tails and old background jobs. | `tools/agents/tuic-status.sh` | Resume, pump, manual diagnosis | In use: `--json`, `--dead`, `--stalled`; unknown inventory is not treated as a dead roster. | | A10 | Find an ego child’s saved conversation and the recorded reason for its last stop. | `tools/agents/ego-children.sh` | Restart recovery and idle diagnosis | In use: resume procedure and pump cause reporting select it. | | A11 | Ask one read-only advisor, or ask three independent advisors the same precise question. | `tools/agents/tuic-consult.sh`, `tools/agents/tuic-triumvirate.sh` | Manual, justified unresolved decisions | Partial: consult logging exists; global escalation restrictions still apply, and advisor availability was not checked. | | A12 | Audit Claude agent token use and elapsed time by agent and story. | `tools/agents/agent-cost.py` | Manual audit | Partial: parser covers Claude JSONL, not ego or Codex usage. | ## 2. Event delivery and unfinished work | ID | What it does | Entry point | When | Status and evidence | |---|---|---|---|---| | E01 | Batch mail and result files, retain delivery progress and route actionable events. | `tools/agents/coord-pump.sh` | Background while children work | In use: `orchestration.md` requires pump plus refresher; one pump per coordinator. | | E02 | Offer an alternative watcher for mail, result files and child exits. | `tools/agents/mail-watch.sh` | Alternative background watcher | In use as a documented alternative, not an extra consumer of the same inbox; singleton and output guards are implemented. | | E03 | Refresh unfinished-work state without reading mail and wake on changed alarms. | `tools/agents/open-work-refresh.sh` | Beside the pump, refresh normally every ten minutes | In use: explicit two-loop rule in `orchestration.md`. | | E04 | Show owners, orphan work, pending landings, holds and publication debt in a bounded summary. | `tools/agents/pending.sh`, `tools/agents/open_work.py` | Refresh, terminal events, manual audit | In use: pending delegates to the computed snapshot; a full view is retained separately. | | E05 | Read story stores, match ownership and collect state with bounded external commands. | `tools/agents/open_work_inventory.py`, `tools/agents/open_work_state.py` | During E04 | In use: imported by the snapshot collector. | | E06 | Add coordinator queues, parked branches, dirty trunks and stashes that story status cannot show. | `tools/agents/open_work_sources.py`, `tools/agents/repo_key.py` | During refresh and parked-hold reads | In use: source collector and worktree tools share canonical repository keys. | | E07 | Keep RESULT/BLOCKED reports open until a recorded disposition takes responsibility for them. | `tools/agents/open_work_lifecycle.py` | Refresh and hook-baseline installation | In use: joins durable pump events to close/continue dispositions. | | E08 | Retain Boss prompts and agent reports, then link each to a story or an explicit dismissal. | `tools/agents/intake.sh`, `tools/agents/open_work_intake.py`, `tools/agents/open_work_spool.py` | Prompt hooks, pump intake, manual classification | In use: both CLI hook files select intake; spool survives delivery bookkeeping. | | E09 | Select unseen result files and list unread results after a restart. | `tools/agents/pump_results.py`, `tools/agents/unread_results.py` | Pump scan and resume | In use: callers preserve delivery watermarks and report unknown reads. | | E10 | Detect idle episodes, exclude live background runners and issue one scoped nudge per episode. | `tools/agents/idle_children.py` | Pump checks | In use: pump calls the candidate collector and tracks assignment generations. | | E11 | Find free staffing capacity and eligible release items without a live owner. | `tools/agents/pump_staffing.py`, `tools/agents/pump_unassigned.py` | Pump checks | In use: pump calls both; the second reads only fresh cached alarms. | | E12 | Run one locked maintenance pass for idle children, merged worktrees and gate dispatch. | `tools/agents/pump_housekeep.py` | Normally every ten minutes | In use: pump passes close-idle, sweep and gate tick commands. | | E13 | Apply only mechanical follow-up and testing-skill feedback edits with a restricted model worker. | `tools/agents/coord-pump-worker.sh` | Optional pump worker | Partial: narrow sandbox and escalation response exist; enabled use was not checked. | | E14 | Detect a ready batch that no integrator has taken. | `tools/agents/wave-idle.sh`, `tools/agents/wave_index.py` | Pump records events; Stop hook checks | In use: durable ready-wave index and Stop caller. | | E15 | Find ownerless release stories and reject malformed claims that only Boss can unblock them. | `tools/agents/unowned-work.sh`, `tools/agents/waiting-boss-lint.sh` | Stop hook and parking checks | In use: syntax checks are implemented; the coordinator must still verify that the reason is genuinely Boss-only. | ## 3. Stories, incoming changes and scope | ID | What it does | Entry point | When | Status and evidence | |---|---|---|---|---| | S01 | Avoid assigning a story already carried by an unmerged branch. | `tools/agents/story-claimed.sh` | Before spawn | In use: spawn invokes claim checks; gate-only and dead start branches have explicit rules. | | S02 | Find overdue story reviews and start an evidence-based classifier of real remaining work. | `tools/agents/triage-due.sh`, `tools/agents/story-triage.sh`, `tools/agents/briefs/story-triage.md` | Resume; manual dispatch when review is due | In use: dispatch ledger and three-day due check; implementation defaults to ego. | | S03 | Show Boss requests without a non-documentation commit, including progress through child stories. | `tools/agents/boss-unbuilt.sh` | Resume | In use: resume caller and history scan. | | S04 | Refuse story completion without acceptable criteria, named bug evidence and valid commit provenance. | `tools/agents/story-close.sh`, `tools/agents/story-close-preflight.cjs` | Read-only preflight before handoff; close after landing | In use: installed Wiz validators; Proof is not fabricated or filled with generic evidence. | | S05 | Execute a repository’s stories sequentially, with one combined gate and one owner of all failures. | `tools/agents/briefs/batch.md` | Coordinator prepares a batch | In use: `decisions.md`, 21 October; one critic covers the complete batch. | | S06 | Keep release membership separate from next work and valid waits on Boss. | `decisions.md`, `release-scope.tsv`, `tracker.md` | Manual decisions; derived status | In use: freeze rule updated 29 October; new Boss stories also go to next. Scope is not an automatic promise to publish. | | S07 | Review external pull requests against local trunk, verify candidates and maintain a living PR map. | `.agents/skills/pr-intake/SKILL.md`, `.agents/skills/pr-intake/brief-triage.md` | On request | In use: skill separates code verdict, test proof and authorization for GitHub actions. | ## 4. Quality gates and evidence | ID | What it does | Entry point | When | Status and evidence | |---|---|---|---|---| | Q01 | Generate a bounded independent critic assignment from the actual story and branch. | `tools/agents/critic-brief.sh`, `tools/agents/critic-brief.template.md` | After implementer RESULT, rounds one and two | In use: current script takes repository, story, worktree and round; narrow production fixes are allowed in round two. | | Q02 | Convert scored findings into FIX, STORY or ACCEPT with a fixed rule. | `tools/agents/critic-triage.sh`, `tools/agents/critic-triage.md` | Each scored review | In use: verdict considers value, effort, risk and added machinery. | | Q03 | Retain an older critic brief beside the active template. | `tools/agents/briefs/critic.md` | Legacy reference | Unused by the current generator: Q01 loads `critic-brief.template.md`; the two texts differ. See I1. | | Q04 | Run TUIC’s pinned Rust batch and record each completed suite. | `tools/gates/tuic-batch.sh` | Once per integration batch, in background | In use: default/headless nextest, strict all-targets clippy and headless binary build; local frontend suite is separate. | | Q05 | Record suite results by source tree and expose historical mutation pointers without claiming reuse. | `tools/gates/tested.sh`, `tested.tsv` | Before/after runs and status answers | In use: mutation check exits 3 and requires the registry; a historical row is not a pass for current inputs. | | Q06 | Capture landing obligations, dispatch bounded gate batches and retain receipts and explicit recovery. | `tools/gates/gate-queue.py`, `tools/gates/queue.json`, `gate-queue.md` | Landing/result intake, pump tick, upkeep sweep | In use: one locked backlog; at most twelve rows per dispatch; a dead peer becomes blocked, not passed. | | Q07 | Store mutation artifacts atomically, query exact mutants, export evidence and compare current source context. | `tools/gates/mutation-registry.py`, `tools/gates/mutation_context.py`, `tools/gates/mutation_identity.py` | Collection and before reuse | In use: canonical-root SQLite store; source, tests, tools, configuration and target must match. The local ignored database is not transferred by Git. | | Q08 | Decide whether the complete requested mutation scope can skip execution. | `tools/gates/mutation_gate.py`, `tools/gates/MUTATION-GATE.md` | Registry gate command | In use: complete universe and independent evidence are required; observed old outcomes alone are insufficient. | | Q09 | Measure current remote source and environment without building or syncing. | `tools/remote/rb-current-measurements.py` | Before mutation reuse | Partial: implementation and fixture integration exist; no live probe was run in this audit. See I6. | | Q10 | List the exact requested cargo-mutants scope and bind byte-exact mutant identities to current inputs. | `tools/remote/rb-requested-universe.py`, `tools/remote/rb_mutation_universe.py`, `tools/remote/REQUESTED-UNIVERSE.md` | Before whole-scope reuse | Partial: read-only listing supports restricted arguments and explicit `--no-config`; it does not execute mutants. | | Q11 | Reduce an assigned test batch without changing production, using retained behavior and AWS mutation proof. | `.agents/skills/test-janitor/SKILL.md`, `.agents/skills/test-janitor/references/index.md` | On request only | In use as a skill; missing compatible evidence prevents acceptance of deletions, not permission for local mutations. | | Q12 | Compare complete AWS mutation runs and propose same-behavior test dominance candidates. | `.agents/skills/test-janitor/scripts/evidence.py` | Test Janitor evidence comparison | In use: skill reference selects it; it never deletes tests and rejects incompatible/incomplete runs. | ## 5. Landing, worktrees and publication | ID | What it does | Entry point | When | Status and evidence | |---|---|---|---|---| | L01 | Serialize landings across repositories, recover a dead lock holder and run an optional precheck before any merge. | `tools/land/land-serial.sh`, `tools/land/landing-lock.sh` | Background, each queued landing | In use: direct landings share the same lock; ordering of waiters is not guaranteed. | | L02 | Check the merge, formatting, types, dependency notices and Rust test compilation before landing; retain gate obligations and refresh changed JS dependencies. | `tools/land/tuic-land.sh` | Every landing | In use: implementation calls checks and cleanup; distinct exit codes distinguish a merge from incomplete cleanup, dependency refresh or gate intake. | | L03 | Run one full frontend suite on an exported trunk snapshot and clear pending batch rows only on success. | `tools/land/batch-vitest.sh` | After a batch of landings | In use: landings defer full vitest by default; no pending rows means no test run, not fresh proof. | | L04 | Prove integration before removing worktrees and branches; preserve permitted managed-file dirt and refuse live owners. | `tools/land/tuic-land-cleanup.sh` | After landing or cleanup retry | In use: TUIC repository actions, ownership recovery and integration checks precede deletion. | | L05 | Find the trunk, reconcile checkout renames and share worktree/integration/ownership checks. | `tools/land/trunk.sh` | Sourced by landing and worktree tools | In use: supports main or master and TUIC integration evidence. | | L06 | List merged checkouts left behind and safely sweep those eligible for cleanup. | `tools/land/landed-worktrees.sh`, `tools/land/wt-sweep.sh` | Resume; pump sweep; manual retry | In use: live child registrations protect fresh work; apply calls cleanup. | | L07 | Find committed or dirty worktrees with no live owner and show intentional holds. | `tools/land/wt-stranded.sh` | Resume, watcher and manual triage | In use: selects unmerged work or tracked dirt, with parked reasons. | | L08 | Rescue and triage stranded personal work, send useful work through review/landing and ask Boss only about discards. | `.agents/skills/wt-triage/SKILL.md`, `.agents/skills/wt-triage/brief.md` | Request or hourly health procedure | In use as a skill; some instructions are stale, see I2 and I3. WORK repositories remain report-only without authority. | | L09 | Archive a branch tip and prepare deletion commands for Boss without bypassing a deletion guard. | `tools/land/branch-retire.sh` | Manual, guarded branches | In use: verifies archive equals current tip; does not itself delete. | | L10 | Resolve append-only documentation conflicts by retaining both blocks and staging only resolved files. | `tools/land/merge-union.py` | Manual conflict repair | In use: README selects CHANGELOG/to-test use; malformed conflicts refuse. Not a general code merge strategy. | | L11 | Advance a clean initialized submodule and attribute only its pointer change to the coordinator. | `tools/land/advance-submodule.sh` | Manual shared-root update | In use: explicit procedure in `orchestration.md`; records exact-entry attribution. | | L12 | Wait for every GitHub Actions run of a full commit SHA and return success, failure or timeout. | `tools/land/ci-wait.sh` | Background after an authorized push | In use as the prescribed post-push check; no push performed here. | | L13 | Read authoritative successful CI and required job results for release checks and validated tag-only publication. | `tools/land/ci_evidence.py`, `tools/hooks/validated_tags.py` | Release alarm and pre-push exception | In use: callers require exact commit and successful completed CI; this is not permission to publish. | | L14 | Enforce release decisions, including green CI on every platform and separate authorization for publication. | `decisions.md`, `decisions/tuicommander.md`, `decisions/other-projects.md` | Release preparation | In use as policy: plugins-first order, ego private/public history separation and version rules remain repository-specific manual steps, not one release script. | ## 6. Build admission and isolated live checks | ID | What it does | Entry point | When | Status and evidence | |---|---|---|---|---| | B01 | Admit builds by memory, load, disk and configured concurrency; protect the running TUIC checkout and select remote execution. | `tools/build/build-slot.sh`, `tools/build/build-slot.conf` | Every managed build/test | In use: managed rules and wrappers select it; gates have a lower-priority lane. Automatic local fallback has restrictions. | | B02 | Freeze compilation machine-wide with a reason and an allow-list. | `tools/build/build-freeze.sh` | Explicit operational stop | In use: build-slot and rb honor BUILD_FREEZE; not a routine waiting mechanism. | | B03 | Check changed Rust and JS/TS files with their formatter, without formatting unrelated modules or vendor code. | `tools/build/fmt-changed.sh` | Before handoff and landing | In use: landing caller and managed rules. Markdown-only changes do not need a Rust or frontend suite. | | B04 | Audit stale Tauri build metadata in the owning mbx view and optionally quarantine it by rename. | `tools/build/mbx-tauri-quarantine.sh` | Diagnosed missing-permission build failure | In use: managed recovery procedure; requires builds for that worktree to stop first; no deletion. | | B05 | Signal when an isolated desktop test instance answers, rather than waiting for its long-running launcher to exit. | `tools/build/agent-test-up.sh` | Background beside make agent-test | In use: managed rules specify unique instances; the application launcher belongs to TUIC, not this repository. | | B06 | Verify an ego child can drive its required browser wrapper and retain a page capture and transcript. | `tools/agents/browser-wrapper-smoke.py` | Manual access check | In use: managed procedure cites the helper and retained live report; this audit did not launch a browser. | | B07 | Keep native GUI, browser and device checks away from Boss’s daily app and protect his active screen. | `operations/agent-work.md`, `orchestration.md`, `tools/agents/managed-rules.md` | Live validation | In use as policy: isolated app/Vite, MacControl, scoped screenshots, shared simulator limits and process cleanup. External tools supply the actual interaction. | ## 7. Remote execution and disposable environments | ID | What it does | Entry point | When | Status and evidence | |---|---|---|---|---| | R01 | Snapshot a worktree or pinned revision, run a command on rb and collect results and mutation evidence. | `tools/remote/rb` | Through build-slot --remote | In use: shared remote runner; validates transfer prerequisites, macOS-only changes and evidence preservation. mbx is not the normal remote command path. | | R02 | Admit normal, short and gate jobs fairly, bound their runtime, watch client loss and report changed source. | `tools/remote/rb-run` | Installed runner on rb | In use by R01/queue: source defines class slots, age rules and client-loss cleanup; live installed version was not checked. | | R03 | Queue long gates durably so they survive the Mac session, then list, cancel or collect them. | `tools/remote/rb-queue`, `tools/remote/rb-queue.service`, `tools/remote/rb-queue-install.py` | rb queue commands; remote service worker | In use: rb provisions missing queue infrastructure without overwriting existing policy; service health was not probed. | | R04 | Preserve exact transport environment and standard input, bound short SSH calls and protect active replica generations. | `tools/remote/rb-transport`, `tools/remote/rb-ssh.py`, `tools/remote/rb-reserve.py`, `tools/remote/rb-fingerprint.py` | Internal rb calls | In use: rb invokes these helpers for transport, leases and manifest identity. | | R05 | Reclaim remote caches and old replicas safely, and prune old local pinned snapshots only when unlocked. | `tools/remote/rb-gc-box.sh`, `tools/remote/rb-refs-prune.sh` | rb execution and snapshot maintenance | In use: caller integrates GC; running/waiting jobs and reservations protect source. | | R06 | Provision the build user, runner, queue and analysis tools; retain an opt-in Linux cache experiment. | `tools/remote/rb-box-setup.sh`, `tools/remote/rb-crap-install.sh` | Box setup/deployment; explicit experiment | Partial: setup and CRAP installer exist; experimental mbx is separate from the default, and installation was not run. | | R07 | Run an unmodified headless Claude CLI with a token scoped only to that wrapper. | `tools/remote/rb-claude` | Authorized box-side agent run | Partial: wrapper exits 78 without its token; token presence and installed CLI were not checked or read. | | R08 | Create or destroy an isolated Linux VM, install agent CLIs, transfer permitted keys and run agent marker scenarios. | `tools/vm/agent-vm.sh`, `tools/vm/agent-vm.yaml`, `tools/vm/guest/install-agents.sh` | Manual compatibility matrix | Partial: Lima configuration and per-agent installation report exist; VM availability and actual installs were not checked. Key values must not be printed. | ## 8. Boss communication and reports | ID | What it does | Entry point | When | Status and evidence | |---|---|---|---|---| | C01 | Collect unresolved decisions, apply previous ticks/comments first and present plain-language choices with consequences and a recommendation. | `.agents/skills/open-asks/SKILL.md`, `boss-operating-rules.md` | Request; when a real Boss-only decision remains | In use as a skill; current skill output path is stale, see I2. Reversible choices are decided, not turned into blocking questions. | | C02 | Keep investigation answers visible until Boss acknowledges them. | `followups.md`, `tools/reports/boss-unread.sh` | Every Boss prompt | In use: Claude and Codex UserPromptSubmit hooks inject the Da leggere section. | | C03 | Give every release a separate remaining-work count, split into release work, Boss waits and next items. | `tools/reports/release-countdown.sh`, `tools/reports/release_story_states.py` | End of each Boss reply | In use: reads scoped story status; emits one-time NOTIFY at zero. Extra aggregate output conflicts with presentation policy, see I5. | | C04 | Alarm when a recorded release CI condition is satisfied and all scoped stories are complete. | `tools/reports/release_gate_alarm.py` | Countdown | In use: calls authoritative CI reader; does not publish or close the release gate. | | C05 | Find project reports made stale by source changes and generate a new evidence-based progress report. | `tools/reports/report-stale.sh`, `tools/agents/briefs/report.md` | Resume scan; manual report assignment | In use: Source revision, age and relevant commits drive the stale check. | | C06 | Find expired topic reports or reports without expiry dates. | `tools/reports/report-expired.sh` | Resume and tidyup | In use: lists candidates; tidyup handles deletion and preserves unread answers. | | C07 | Embed local images in a digest so the document does not depend on inaccessible local assets. | `tools/reports/digest-inline-images.py` | Manual compatibility workaround | Partial: source says “until TUIC 1335 ships”; need for the workaround today was not checked. | | C08 | Deliver short chat answers with required markers and long bilingual material in persistent report tabs. | `boss-operating-rules.md`, `AGENTS.md`, `.claude/wiz-claude.md` | Each Boss interaction | In use as policy: 💬 answers, intent markers, clickable links, progress record, no toasts or Mission Control iframe. | | C09 | Explain a reported bug with evidence for cause, regression history and prevention; report what is tested and missing. | `boss-operating-rules.md` | Bug reply and status answer | In use as policy; not an automatic root-cause detector. | ## 9. Health and housekeeping | ID | What it does | Entry point | When | Status and evidence | |---|---|---|---|---| | H01 | Recover coordinator identity and a bounded briefing of children, builds, results, reports, due triage, follow-ups and budgets. | `tools/coord-resume.sh`, `tools/coord-state.sh` | Restart, clear or compaction | In use: AGENTS requires it for coordinators; per-coordinator state separates children and delivery while selected locks remain shared. | | H02 | Capture TUIC responsiveness, freeze signals, footprint and slow operations through its local socket. | `tools/health/tuic-health.sh` | Prescribed hourly check | Partial: source and hourly procedure exist; scheduler installation was not checked. | | H03 | Recover a sidecar whose executable bit or stale macOS signature prevents execution, with a bounded retry. | `tools/health/tuic-heal.sh` | Specific CLI failures | In use: spawn/bg/child/Stop callers; it does not rebuild or replace the app. | | H04 | Restart only an explicitly authorized dedicated make-dev process group and supervise its replacement. | `tools/health/tuic-dev-restart.py` | Boss-authorized restart | Partial: manual recovery helper with PID validation; not automatic permission to restart the daily app. | | H05 | Reap orphan agent process trees while protecting live owners, recent work, main checkouts and shared daemons. | `tools/health/reap.sh`, `tools/health/reap.py` | Guard; manual dry-run/apply | In use: guard calls apply on Mac/rb; an unreadable TUIC roster refuses local reaping. | | H06 | Alarm on reaped processes, low free memory or newly failed enabled automations. | `tools/health/guard.sh`, `tools/health/it.straus.orch-guard.plist` | Configured launchd interval 15 minutes | Partial: tracked schedule calls the guard; loaded launchd state and current automation health were not queried. | | H07 | Show model, context, effort, directory and cost in the private Claude status line. | `tools/health/c2-statusline.sh` | Claude status rendering | Partial: README names external settings; those settings were not audited here. | | H08 | Find old stashes using metadata only, without exposing their messages or contents. | `tools/agents/stash-stale.py` | Resume | In use: resume caller; selects stashes over seven days in recently active repositories. | | H09 | Diagnose machine/process load, thermal state and network activity through ProcessGuard. | `processguard.md` | Manual, before blaming a slow build | Partial: documented external socket API; daemon availability was not checked. | ## 10. Memory, policy and upkeep | ID | What it does | Entry point | When | Status and evidence | |---|---|---|---|---| | M01 | Search repository knowledge first, verify source and retain durable decisions and verified lessons. | `AGENTS.md`, `boss-operating-rules.md`, `orchestration.md`, `operations/agent-work.md` | Research, decisions and lessons during work | In use as policy: MDKB and topic files are separate from transient delivery state. | | M02 | Audit root-file ownership and budgets, compact generated triage history and verify the generated gate view. | `tools/upkeep/root-upkeep.py` | Tidyup start/apply and final/check | In use: detects unresolved dirt without deleting Boss choices, evidence or mbx-owned targets. | | M03 | Consolidate live decisions/lessons, remove eligible expired reports, reconcile memory and save the day’s durable corrections. | `.agents/skills/tidyup/SKILL.md` | Nightly scheduler request, manual /tidyup or budget warning | In use as a skill: final check is mandatory; current nightly schedule health is not certified by this report. | | M04 | Keep model routing, delegation limits, escalation and managed handoff under one global authority. | `~/.codex/ORCHESTRATION.md`, `~/.codex/SUBAGENT.md`, `tools/agents/managed-rules.md` | Assignment, execution and final handoff | In use: current files read; repository routing duplicates and test-policy drift remain in I4. | | M05 | Repair a failed owned tool in its repository rather than silently working around it; retain one active home per lesson. | `orchestration.md`, `AGENTS.md` | Diagnosed tool failure | In use as policy: ownership table identifies TUIC, MDKB, Wiz, MacControl, ego and this repository. mbx is external. | ## 11. Hooks and guards | ID | What it does | Entry point | When | Status and evidence | |---|---|---|---|---| | G01 | Refuse a coordinator stop when release work lacks owners, ready waves lack integration or staffing is below the justified floor. | `tools/hooks/stop-unowned.sh` | Claude Stop hook | In use: `.claude/settings.json`; repeated active Stop passes to avoid a hook loop. Not configured as a Codex Stop hook here. | | G02 | Require research briefs to say MDKB first, and require a recent repository-scoped knowledge lookup before recursive searches. | `tools/hooks/agent-mdkb-guard.sh`, `tools/hooks/mdkb-first.sh` | PreToolUse | In use: Claude wires both; Codex wires the research-agent guard but not mdkb-first in its local hook file. | | G03 | Reject an invisible shell-backgrounded pump and interactive questions that would block Boss. | `tools/hooks/no-bg-pump.sh`, `tools/hooks/no-ask-user.sh` | PreToolUse; Claude permission deny | In use: Claude wires no-bg-pump and denies AskUserQuestion; Codex wires no-ask-user. | | G04 | Refuse pattern-wide process kills and zsh assignments that accidentally replace PATH. | `tools/hooks/no-pkill.sh`, `tools/hooks/no-zsh-path.sh` | When installed in tool hooks | Partial: implementations and fixtures exist; neither local CLI hook file selects these scripts. Global installation was not checked. | | G05 | Reject newly added unowned deferrals while retaining legacy text and valid Boss decision options. | `tools/hooks/deferral-check.py`, `tools/hooks/install.sh`, `.githooks/pre-commit`, `tools/hooks/deferral-baseline.json` | Hook installation on resume; pre-commit | In use: tracked pre-commit refuses only proven invalid deferrals; checker infrastructure failures print UNKNOWN and pass. | | G06 | Refuse publication while the repository still owes a full frontend batch, except a validated tag-only transaction. | `tools/hooks/batch-vitest-push.sh` | Global pre-push chain | In use by the tracked publication wrapper; checks pending rows without clearing them. Actual global Git configuration was not checked. | | G07 | Scan outgoing messages, paths, additions and tags for denied publication material, while preserving repository hooks. | the orchestrator pre-push publication hook | Global pre-push, when installed | Partial: source and README installation procedure exist; external denylist and active global hook config were not read. Missing denylist refuses. | ## Coverage and verification boundaries All tracked operational scripts under `tools/agents`, `land`, `gates`, `health`, `hooks`, `reports`, `upkeep`, `remote`, `build`, `vm`, and both top-level shell files are named above, including imported helper modules. All five local skills and their supporting instructions are included. Script tests under each `tools/*/tests/`, `tools/tests/`, and the Test Janitor evidence test are validation support; their names and fixtures were inventoried, but their suites were not executed. Remote live checks, VM installs, builds, mutation runs, scheduler changes and GitHub publication were deliberately not performed for this documentation task. The requested policy sources were read: `AGENTS.md`, `boss-operating-rules.md`, `orchestration.md`, `operations/agent-work.md`, `decisions.md`, `decisions/tuicommander.md`, `decisions/other-projects.md`, `tools/README.md`, all local skill instructions, `.claude/settings.json`, and `~/.codex/ORCHESTRATION.md`. The audit also inspected `.codex/hooks.json`, tracked Git hooks, schedule/service files and referenced brief templates. Existing product decisions are policy inputs, not claims that this repository implements those application features. ## IMPROVEMENT notes - **I1 — Duplicate and stale critic documentation.** `tools/README.md` has two incompatible `critic-brief.sh` signatures. The implementation accepts ` [focus]` and loads `critic-brief.template.md`, not `briefs/critic.md`. The latter still requests package-wide frontend vitest while the active template requests named files plus typecheck. README also describes FIX/SKIP although the script emits FIX/STORY/ACCEPT, and describes generic Proof filling that story-close explicitly forbids. Reconcile the index and retire the unused template only after confirming external users. - **I2 — Digest paths contradict persistent delivery.** Open-asks and worktree-triage skills still write under `~/Gits/.tmp/boss/`. Boss operating rules require persistent bilingual reports under this repository because mobile cannot open untrusted temp paths. Update both skills and preserve/apply existing ticks before moving any live digest. - **I3 — Worktree skill cleanup instructions are stale.** The worktree-triage skill still ends with raw Git worktree/branch deletion and allows direct low-risk landing. Current managed rules require TUIC tracking, child-close ownership and the selected review flow; operations also says every branch needs a critic. Reconcile before the next destructive triage, without discarding existing work. - **I4 — Routing and validation rules have several competing copies.** Global policy is authoritative, but spawn roles, legacy generated guides and local templates contain other routing defaults or admission limits (for example Luna is implemented but not admitted by `spawn-roles`). Per-story RED/strict-clippy rules, deferred GREEN rules and batch retrospective RED instructions also differ across managed rules and templates. Keep one authoritative contract and make wrappers/templates derive it; do not resolve this by silently choosing a cheaper gate. - **I5 — Countdown output still contains a cross-release total.** `release-countdown.sh` ends with an `all:` line, but Boss rules require separate release lines and never one combined total. Per-release calculation exists; remove or explicitly separate the unwanted presentation only in a scoped tested change. - **I6 — Mutation documentation understates current implementation.** README still says the current-measurement protocol fixture/end-to-end work is pending as of 3 October. The repository now contains `rb-current-measurements.py`, the producer fixture and `rb-current-measurements-e2e.test.py`. Update the historical note after checking their recorded result; file presence alone does not prove a current live pass. A deletion-only reuse bridge remains explicitly absent. - **I7 — Optional compatibility helpers need retirement checks.** `close-idle.sh` names a future native replacement, and digest-inline-images names a temporary TUIC fix. This audit did not verify those target story states. Check the target product before retaining or removing either workaround. - **I8 — Hook coverage differs by CLI.** The local Claude and Codex hook files select different guards; no-pkill and no-zsh-path are present without local wiring. Verify the effective global/runtime installation and document coverage, rather than assuming one configuration protects every harness. - **I9 — Installed health is not source health.** Tracked launchd/systemd definitions, optional worker paths and external status-line settings do not prove those jobs run. A separate read-only installation/last-run audit would establish actual in-use, broken or unused runtime status. This report intentionally makes no such runtime claim. - **I10 — Mutation evidence needs an operational backup owner.** The registry is local and Git-ignored, and the README says a safe backup is needed. No backup entry point was found among the audited tool capabilities. Name the owner and safe SQLite backup procedure before relying on the store as the only surviving evidence copy. ## Documentation checks Static checks passed: 105 capability rows in 11 classes; identical row IDs and entry points in both languages; all 122 tracked operational scripts covered; valid expiry, table structure and cited local paths. The instruction-link checker passed for the current repository. No application suite was run because only these two Markdown reports changed. No operational code or live state was changed by this inventory.