@ours.network/fleet 0.17.6 → 0.17.8

package/dist/build-info.json

@@ -1,9 +1,9 @@
 {
-  "version": "0.17.6",
-  "buildId": "a2f7650e3a39",
-  "commit": "96df9112202bb3957d18d95bc4210c1cf49d867f",
+  "version": "0.17.8",
+  "buildId": "39544c6acdb4",
+  "commit": "da17b486c7b7d486f1877852598a60d1fed48abe",
   "dirty": false,
-  "builtAt": "2026-08-17T11:51:49.146Z",
+  "builtAt": "2026-08-18T13:28:16.260Z",
   "capabilities": [
     "monitor.interrupt.after_tool"
   ]

package/dist/cli.js

@@ -648,7 +648,10 @@ function renderLoopRows(role, state, loop) {
     return Object.entries(state.loops).filter(([name]) => !loop || name === loop).map(([name, item]) => `${role}/${name} ${item.enabled && !item.operatorDisabled ? 'enabled' : 'disabled'} `
         + `${item.activeRunId ? 'running' : 'idle'} next=${item.nextDueAt} last=${item.lastOutcome ?? 'never'} `
         + `counts=${item.counts.started}/${item.counts.completed}/${item.counts.failed} `
-        + `skip=${item.counts.skipped}(busy=${item.counts.skippedBusy},missed=${item.counts.skippedMissed})`);
+        + `skip=${item.counts.skipped}(busy=${item.counts.skippedBusy},missed=${item.counts.skippedMissed})`
+        + (item.missedGap
+            ? ` gap=${item.missedGap.count}@${item.missedGap.fromAt}..${item.missedGap.throughAt}`
+            : ''));
 }
 cOpt(loopsCommand.command('status [role] [loop]').description('show live or stored loop state'))
     .option('--json', 'emit stable JSON')

package/dist/docs.d.ts

@@ -4,7 +4,7 @@
  * Keep this concise enough to place directly in an agent context. Unlike
  * Commander's per-command help, this describes how the pieces compose.
  */
-export declare const AI_DOCS = "# ours-fleet reference\n\nours-fleet runs persistent or temporary, identity-bound AI roles. A role selects\na harness independently from its session backend:\n\n- harness: `claude-code` or `codex`\n- session: `tmux` (default) or `acp`\n- lifetime: permanent (supervised, restartable) or `spawn --temp`\n\n## Discover and validate\n\n```sh\nours-fleet docs                         # this complete reference (`man` is an alias)\nours-fleet help <command>               # exact flags for one command\nours-fleet config [-c FILE]             # validate and print the merged plan; no changes\nours-fleet doctor [-c FILE] [--harness codex|claude-code]\nours-fleet version [--json]             # build identity, capabilities, every install on PATH\n```\n\nDefault configuration is `~/fleet.yaml` plus sorted `~/fleet.d/*.yaml` role\ndrop-ins. An explicit `-c FILE` replaces `~/fleet.yaml`; fleet.d still adds\nroles. Validate with `config` and `doctor` before starting or restarting.\n\nThe CLI never writes the base file: `spawn` writes `~/fleet.d/Name.yaml`. The\nweb console does write it, as a whole document \u2014 its setup wizard and\nconfiguration editor may create, change or remove any top-level block, including\n`vars:`, `defaults:`, `roles:`, `watchdogs:` and `loops:`. Only the base\nfile may hold `defaults:`, `watchdogs:` and `loops:`; a fleet.d drop-in may\ndeclare `roles:` and nothing else. Unrecognised top-level keys are round-tripped\nuntouched. Console edits are applied as surgical splices against the file's exact\nbytes, so an unchanged save is byte-identical and lines outside the edit keep their\ncomments and spacing. One exception: changing the length of a block sequence\n(`watch:`, `oversee:`, `roles:`, `wake_sources:`) may replace that collection\nwholesale and drop inline comments written on its items; lines outside that\ncollection remain byte-preserved. Each save is revision-guarded, reviewed as a diff\nof the real file before anything is written, validated by the real loader, and\nbacked up next to the file first.\n\n## Build identity and install provenance\n\n`--version` prints a semver and nothing else, and a semver does NOT identify an\nartifact. Version bumps land in a release commit of their own, so every build cut\nbetween two releases carries the PREVIOUS version while already containing new\nbehaviour. One host ran two installs that both reported 0.16.0 \u2014 same version,\ndifferent build. One accepted `monitor.interrupt: after_tool`, the other\nrejected it as invalid. Their\n`dist/cli.js` were byte-identical \u2014 the divergence was in other modules.\n\nEvery build therefore stamps `dist/build-info.json` with a build id (first 12 hex\nof a sha256 over the rest of `dist/`), the commit it was cut from, and the\ncapability tokens the shipped code declares \u2014 for example\n`monitor.interrupt.after_tool`. Ask any executable what it is:\n\n```sh\nours-fleet version                      # ours-fleet 0.17.0+9f1c2a3b4d5e, capabilities, PATH installs\nours-fleet version --json               # the same as machine-readable JSON, no environment values\n```\n\nRead a capability, never a version number, to decide whether a setting is\nsupported. When a build rejects a value it knows the name of, it says which\ncapability is missing and which build rejected it, because another install on the\nsame host may accept the identical file. `config` prints the build that resolved\nthe plan; `status <Name>` says so when the build reporting on a role is not the\none that created it (roles record their creating build in `creation.json`).\n\n`ours-fleet doctor` runs an `install` check that lists every `ours-fleet` on\nPATH plus the one executing, and FAILS when two installs share a semver but are\ndifferent builds, or when the running artifact is a DIFFERENT artifact from the\none PATH resolves to. A second prefix holding identical content is not a skew\nand is not reported. A PATH entry the shell would not execute \u2014 a directory, or\na file without its execute bit \u2014 is not counted as an install at all.\nInstalls built before this stamp existed report `+unknown`; they are compared by\nhashing their `dist/` instead, so two pre-provenance installs are still told\napart. To fix a flagged host, remove or update the stale install \u2014 do not rely on\nPATH order.\n\n## Lifecycle and console commands\n\n```sh\nours-fleet init\nours-fleet up|down [Name...]\nours-fleet restart [Name...]            # preserve/resume harness context\nours-fleet force-restart [Name...]      # fresh context; briefing is reloaded\nours-fleet ls\nours-fleet status|peek|attach|logs Name\nours-fleet logs -f Name\nours-fleet send Name \"prompt\"\nours-fleet send Name --key Enter        # tmux only\nours-fleet rm Name\nours-fleet watchdog-report <name> [run-id] [--list] [--json]\nours-fleet watchdog-run <name>\n```\n\n`peek`, `attach`, and text `send` work with tmux and ACP. ACP attachment\nalso accepts `/permit <permission-id> <option-id>`, `/interrupt`, and\n`/detach`. Raw `--key` input is tmux-only.\n\n## Local web console\n\nThe npm package includes the web console; installed users do not clone the repo\nor run `npm run build`:\n\n```sh\nnpm i -g @ours.network/fleet\nours-fleet init\nours-fleet doctor\nours-fleet web                         # install/update service, start, pair browser\n```\n\nThe normal command uses stable `http://127.0.0.1:49271/`, installs an\nowner-level systemd user service (Linux) or LaunchAgent (macOS), and opens a\nfive-minute one-use pairing link in the local browser. After pairing, bookmark\nthe plain URL or install the PWA. To pair a new, signed-out, or revoked browser,\nrun `ours-fleet web open`.\n\n```sh\nours-fleet web status\nours-fleet web start|stop|restart\nours-fleet web open\nours-fleet web revoke-all              # revoke every browser and active session\nours-fleet web uninstall\nours-fleet web serve --port 0 --no-open # isolated foreground/testing mode\n```\n\nThe console is IPv4-loopback-only by default. Both `localhost` and\n`127.0.0.1` are accepted locally. For an nginx/TLS reverse proxy, keep the\ndefault bind and declare the exact browser origin:\n\n`ours-fleet web install --public-origin https://fleet.example.com --password-file /secure/fleet-password`\n\nFleet reads the password file during setup and persists only a salted scrypt\nverifier. New browsers authenticate and retain rotating HttpOnly/SameSite\ntrusted-device credentials. If nginx already authenticates, the operator may\ndeliberately select `--no-password`; the CLI and browser warn that anyone\nreaching the origin can control the fleet. First setup requires an explicit\nchoice: `--password-file` or `--pairing` for protected access, or\n`--no-password` for intentional unprotected access.\n\nUse `--bind ADDRESS` only for an intentional direct listen. A non-loopback\nbind is rejected unless `--public-origin` is also present. Host/Origin checks\nuse the declaration and do not trust forwarded headers. Configure nginx to\nproxy HTTP and WebSocket upgrades to `127.0.0.1:49271` and terminate TLS;\nfleet accepts nginx's loopback upstream Host, so no Host rewrite is required.\nBrowser credentials add Secure for HTTPS, and `revoke-all` invalidates all\ntrusted devices. Role creation offers harness-scoped known-model choices\nwhile still accepting a typed model ID; blank explicitly uses the selected\nharness's own default.\n\n## Spawn\n\n```sh\nours-fleet spawn [--temp] [Name | --role Name] \\\n  --harness codex|claude-code --session tmux|acp \\\n  --mission \"one line\" --cwd /absolute/path --identity Identity \\\n  --coordinator Coordinator --model MODEL \\\n  --approval ask|auto|allow \\\n  --filesystem read-only|workspace|unrestricted \\\n  --unattended deny|wait \\\n  --bio-file /path/bio.md --persona-file /path/persona.md\n```\n\nPermanent spawn writes `~/fleet.d/Name.yaml` and starts a supervised role.\n`--temp` writes active state under `~/.ours-fleet/tmp` and starts an independent\ntransient supervisor (a collected systemd unit or submitted launchd job). It is\nnot enabled across reboot and does not die when the role that spawned it restarts.\nBoth lifetimes support `--session acp`. When a temporary role's bound identity\ncloses or its session ends, the supervisor, monitor and live roster entry retire\ntogether; state moves intact to `~/.ours-fleet/recovery/temporary` with a\ntermination record. Failed launches use the same archive rather than deleting\ntheir briefing, provenance, logs or partial supervisor metadata.\n\nNamed `down` and `rm` commands can target an exact state-backed temporary role\neven though it is absent from merged fleet YAML. The recorded transient unit/job\nis authoritative. Missing/incomplete ownership metadata is reconciled only from\nan exact `_run-temp <role>` process-table match: one match may be adopted, zero\nsettles as stopped, and ambiguity or an unreadable table fails closed. Launching\nrecords receive a bounded grace so a not-yet-registered transient unit cannot be\nmistaken for a stopped one. Stale recorded supervisors are reclaimed in bounded\nbatches by moving their state to the same recovery archive, never by blind deletion.\n\nTemporary-role identity bootstrap is capability-based. The generated briefing\nfirst tries to bind the exact assigned identity and preserves it when it already\nexists. If missing, it uses ours MCP `create_temporary_identity` when that tool\nis exposed, tying a newly-created identity to the connector session lifecycle;\nolder servers fall back to `create_identity`. Collisions and creation errors\nstop safely without force-adopting or deleting identity state. Permanent roles\nretain normal `create_identity` behavior.\n\nThe temporary supervisor treats its first positive identity observation as the\nlifecycle readiness gate: a cold harness may take as long as needed to read its\nbriefing and bind, without a fixed first-bind retirement timer. After readiness,\nonly sustained authoritative absence closes the role. Unreachable, malformed, or\nvalid-but-empty daemon indexes are ambiguous and reset closure debounce rather\nthan becoming cleanup authority.\n\nInside a managed ACP role, the same CLI automatically routes a real `spawn`\nthrough that role's authenticated supervisor control socket. `--role Name` is\naccepted as an alternative to the positional name, so a minimal delegated call\nis `ours-fleet spawn --role DeveloperX --temp`. The supervisor records the\ncalling role, performs creation, and only after success sends a structured\nspawn notice through the caller's owner channel when one is configured.\n\nOmitted harness, session, working directory, coordinator, neutral permissions,\nfleet monitor policy, and (when the harness is unchanged) model inherit from the\ncalling role. Explicit options always win. Selecting a different harness without\n`--model` leaves model selection to that harness/fleet defaults rather than\ncopying an incompatible caller model. This automatic proxy is a convenience and\nattribution mechanism, not an isolation boundary: an unrestricted role can still\ninvoke another binary path directly. Tmux roles and host/operator shells keep the\nordinary direct CLI behavior.\n\nCodex-specific spawn flags: `--sandbox`, `--permission-mode`, `--launcher`,\n`--profile`, `--search`, repeatable `--codex-config key=value`, repeatable\n`--add-dir`, and legacy `--monitor` (consent for the native Codex monitor,\nnot the `monitor.mode` wake-owner selector). Run `ours-fleet help spawn` for\nexact values.\n\n## fleet.yaml\n\n```yaml\nvars:\n  work_root: /home/me/work\nstart_stagger_ms: 0\ndefaults:\n  harness: codex\n  session: acp\n  model: gpt-model-id\n  permissions:\n    approval: ask\n    filesystem: workspace\n    unattended: deny\n  monitor:\n    mode: fleet                          # fleet (default) | native\nroles:\n  Coordinator:\n    harness: codex\n    session: acp\n    identity: Coordinator\n    cwd: ${work_root}/project\n    mission: Coordinate work and delegate implementation.\n    model: gpt-model-id\n    permissions:\n      approval: ask\n      filesystem: workspace\n      unattended: deny\n    session_options:                    # advanced overrides; normally omit\n      # acp:\n      #   command: [/custom/codex-acp, --flag]\n      tmux:\n        boot_grace_ms: 10000\n    monitor:\n      mode: fleet                        # fleet supervisor | native harness monitor\n      interrupt: false                    # false queues; true cancels; after_tool steers at an ACP tool boundary\n      wake_sources: [message_received, file_received, local_contact_request, pending_message]\n      batch_ms: 2000\n      inject: notification\n      turn_fail_threshold: 3\n    harness_options:\n      launcher: auto\n      sandbox: workspace-write\n      approval: on-request\n      search: false\n      profile: fleet\n      add_dirs: [/data/shared]\n      config:\n        model_reasoning_effort: high\n    bio: Public role card and when peers should engage it.\n    persona: Local operating contract, boundaries, and escalation policy.\n    briefing_file: /absolute/custom-briefing.md\n    coordinator: AnotherCoordinator\n    env:\n      KEY: value\n    oversee:\n      - { role: Worker, interval: 5m }\nwatchdogs:\n  nightwatch:                       # [A-Za-z0-9_-], must not collide with a role name\n    coordinator: FleetCoordinator   # required \u2014 where alerts go\n    # everything below is optional\n    enabled: true                   # default true; false = configured but never scheduled\n    interval: 10m                   # default 10m; 30s | 10m | 2h, minimum 1m\n    watch: [Alice, CodexReviewer]   # explicit lists are exact; omit for configured + live temp roles\n    harness: claude-code            # default: defaults.harness\n    model: claude-fable-5           # default: same resolution rule roles use (resolveRoleModel)\n    session: acp                    # default: defaults.session\n    identity: Watchdog-nightwatch   # default: Watchdog-<name>\n    timeout: 5m                     # default 5m; a run past this is killed and recorded as error\n    keep_reports: 50                # default 50 reports retained per watchdog\n    alert_cooldown: 60m             # default 60m before the same finding alerts again\n    prompt_file: /abs/extra.md      # optional extra focus, APPENDED to the fixed contract\n    isolation:                      # optional; omitted means no OS sandbox, like an ordinary role\n      backend: bubblewrap           # when present, the ordinary role isolation schema applies\n      network: broker\n      fs: { read: [/opt/watch-data] }\n```\n\nA watchdog observes and reports; it never restarts, stops, spawns, or removes a\nrole, answers a pending permission, edits a workspace, or approves anything on\nthe owner's behalf. `watchdogs:` may appear only in the base config\n(`~/fleet.yaml` or `-c FILE`), not in `~/fleet.d/*.yaml` drop-ins.\nWatchdogs are not isolated by default. An explicit watchdog `isolation:` block\nuses the same policy schema as a role and is applied unchanged; declare every\nextra filesystem access required by a custom prompt there.\nWhen `watch:` is omitted, each run watches the configured roles plus temporary\nfleet roles that are live when the run starts. An explicit `watch:` list is\nnever augmented.\n\nRole values override defaults. `${name}` substitutes entries from `vars`.\nOther role fields include `max_tokens`, `autocompact_pct`, and `isolation`.\nUse README.md for the complete isolation policy and resource-cap schema.\n\nSupervised roles connect to the operator-configured ours daemon; they do not own its\nlifecycle. Fleet forces `OURS_AUTOSTART=0` in tmux and ACP child processes after role\nenvironment overlays. Start the shared daemon only through an explicit operator or\ninstaller/setup flow.\n\n## Permissions\n\nPrefer the harness-neutral `permissions` block:\n\n- `approval: ask|auto|allow`: portable permission policy. `deny` remains a\n  deprecated, fail-closed compatibility alias for existing fleet files.\n- `filesystem: read-only|workspace|unrestricted`: filesystem intent\n- `unattended: deny|wait`: what ACP does when no console can answer a request\n\nThe backend translates this common intent. Harness-native settings in\n`harness_options` take precedence where supplied. Do not choose\n`allow`/`unrestricted`, Codex `never`/`danger-full-access`, or Claude\n`bypassPermissions` without explicit authorization.\n\n### Creation-time isolation\n\n`ours-fleet spawn --isolation-file <path>` supplies a role's sandbox policy at\ncreation, so the FIRST launch is already confined \u2014 a role that only gains\n`isolation:` on a later `up` ran unsandboxed until then.\n\nThe file holds exactly the `isolation:` mapping documented above and nothing\nelse \u2014 the same schema, validated by the same code, so a policy written here\ncannot mean something different from the identical block in fleet.yaml:\n\n```yaml\nnetwork: deny\nfs:\n  read: [/opt/reference]\nresources:\n  mem: 2G\n```\n\nInvalid files are rejected before anything is created: no config, no state\ndirectory, no identity reservation. Works for both permanent and `--temp` roles.\n\n### Never-prompt failure\n\nThe failure this section exists to prevent leaves no error message anywhere.\n\nAn unattended role has no console. When the harness needs a permission decision\nthere is nobody to ask, so the request is refused INSIDE the harness \u2014 no\nprompt, no error, no log line. The agent simply does less than its briefing told\nit to, reports success, and nothing distinguishes that from having done the\nwork. Two settings produce it:\n\n1. a permission mode that suppresses the prompt without granting the action\n   (Claude `dontAsk`, which is why neutral `allow` maps to\n   `bypassPermissions` instead); and\n2. `unattended: deny`, which refuses every request that reaches it.\n\n**Automatic decisions are now recorded.** Every permission request decided\nwithout a human emits a completed event into\n`~/.ours-fleet/agents/<Name>/.session-events.jsonl` carrying the decision,\nwhether policy or a person made it, the policy that produced it\n(`permissions.unattended=deny` vs `permissions.approval=deny`/`=allow`),\nthe reason, and the option selected. `ours-fleet peek` and `attach` render\nthem. Automatic denial asks for a one-shot rejection, never a standing one, so a\nsingle unattended refusal cannot disable a tool for the rest of the session.\n\nA role that can auto-deny logs one line at startup saying so.\n\nTo detect an under-permissioned role BEFORE it runs, use the capability floor\nbelow: `ours-fleet doctor` fails such a role rather than letting it discover\nthe problem silently at work.\n\n### The unattended capability floor\n\nAn unattended role has no console, so a permission request cannot be answered \u2014\nit is refused, silently, inside the harness. The agent then does less than it\nwas told to and reports no error. To make that visible before launch,\n`ours-fleet config` and `ours-fleet doctor` resolve each role's neutral\npermissions through its harness and check the result against a fixed floor:\n\n- `read-state` \u2014 read its briefing, ROUTINES.md, and WORKLOG.md\n- `write-state` \u2014 append its WORKLOG and its own state files\n- `messaging` \u2014 bind its identity, send and receive ours mail\n- `monitor` \u2014 arm and observe its mail monitor\n- `workspace-edit` \u2014 edit and test files in its working directory\n- `status-commands` \u2014 run the inspection commands its briefing prescribes\n\n`doctor` reports this per role as `unattended floor: <Role>`. A role with\n`unattended: deny` that cannot meet the floor FAILS doctor, because it will\ndeny those requests with nobody to see it; with `unattended: wait` it warns,\nbecause a human can still attach and answer.\n\nSecurity meaning: `ask` maps to Codex `untrusted` and Claude `default`.\n`auto` selects Codex ACP `agent` (`on-request` + `workspace-write`) and\nClaude `acceptEdits`. `approval: allow` selects Codex ACP's fully\nnon-interactive yolo mode, reported as `agent-full-access` (`never` +\n`danger-full-access`), and Claude `bypassPermissions`. Codex tmux retains\nindependent approval and sandbox flags: `auto` is `on-request`, `allow`\nis `never`, and `filesystem` still selects the sandbox. These modes genuinely\npermit the actions the role was authorized to take \u2014\n`dontAsk` only suppresses the prompt while still refusing the action. Nothing\nother than an explicit `allow` becomes non-interactive. Legacy `deny` keeps\nits conservative Codex `on-request` / Claude `plan` translation. `allow` is therefore a real grant and\nrequires explicit authorization; per-role `isolation:` remains the outer\nboundary that a permission mode cannot cross.\n\nACP carries agent-advertised session mode IDs and `session/set_mode`, but those\nIDs are agent-specific and ACP defines no portable permission-policy capability.\nFleet therefore uses the ACP primitive where an adapter exposes a matching mode\nand otherwise performs the harness translation above. The bundled Codex ACP\nadapter couples approval and sandboxing in its advertised mode IDs. Neutral\n`allow` therefore selects `agent-full-access` and widens `filesystem:\nworkspace` or `read-only` to `danger-full-access`; neutral `auto` selects\n`agent` and `workspace-write` even when the neutral filesystem value differs.\nAn explicit `harness_options.sandbox` selects its corresponding ACP preset and\nstill wins, as does an explicit native approval override. `config` and\n`doctor` report a coupled-mode mismatch as approximate. Use per-role\n`isolation:` as the outer boundary for an `allow` ACP role. The live session\nreports both its effective normalized mode and the exact native mode selected.\n\nSee also: `spawn --approval/--filesystem/--unattended` set this intent at\ncreation, and `ours-fleet config` prints each role's neutral settings, their\nnative translation, and any warning \u2014 the same text `doctor` reports.\n\nClaude `harness_options`: `permission_mode` (default, acceptEdits, plan,\ndontAsk, bypassPermissions), `plugins`, `mem_palace`, and\n`mem_palace_midsession_autosave`.\n\nCodex `harness_options`: `launcher` (auto, ours-codex, codex), `sandbox`\n(read-only, workspace-write, danger-full-access), `approval` or\n`permission_mode` (untrusted, on-request, never), `profile`, `search`,\n`config`, `add_dirs`, and `monitor`.\n\n## ACP adapters\n\nThe maintained `@agentclientprotocol/codex-acp` and\n`@agentclientprotocol/claude-agent-acp` runtimes are bundled automatically as\noptional ours-fleet dependencies. The supervisor resolves their executable\nentrypoints internally, so default ACP roles do not depend on global PATH.\nThe maintained Claude adapter requires Node 22; tmux and Codex ACP continue to\nwork on the ours-fleet core minimum of Node 20.\n\nOverride an adapter only when necessary with `session_options.acp.command`\n(string or argv list). If optional dependencies were deliberately omitted,\nours-fleet falls back to a compatible globally installed `codex-acp` or\n`claude-agent-acp`. `ours-fleet doctor -c FILE` verifies the resolved adapter.\n\n## Reliable mail wake\n\n`monitor.mode` selects exactly one wake owner:\n\n- `fleet` (default): the ours-fleet supervisor consumes body-free daemon\n  events and advances its durable cursor only after delivery is accepted. ACP\n  uses live steering when supported and falls back to structured\n  `session/prompt`; tmux uses verified console injection.\n- `native`: ours-fleet starts no supervisor monitor; the generated briefing\n  instructs Claude Code or Codex to arm its harness-native wake mechanism.\n\nSet `monitor.interrupt: true` in fleet mode to cancel active work before every\nconfigured wake. Set it to `after_tool` to preserve an active ACP tool (and any\npending permission), then steer the wake at the first tool-terminal boundary\nwithout cancellation. A hung boundary is bounded at 120 seconds and falls back\nto non-cancelling steering/queueing; adapters without authenticated tool events\nuse the same conservative fallback. Explicit human/control interrupts remain\nimmediate. The policy is content-blind because the supervisor cannot inspect\nencrypted message bodies. Message bodies are released only when the role calls\nthe ours `get_messages` tool.\n\nThe default is `false`. For a temporary role whose mission intentionally arrives\nafter its readiness announcement, set `mode: fleet` and `interrupt: true`\nexplicitly. The readiness announcement does not change the transport: the\nmission remains ordinary ours mail, fleet injects only the body-free wake, and\nthe role calls `get_messages` before acting. Every later configured wake uses\nthe same interruption policy.\n\nLegacy `monitor.enabled: true|false` remains accepted as an alias for\n`mode: fleet|native`; use `mode` in new configuration. Codex's separate\n`harness_options.monitor: true` is native-monitor consent, not monitor-owner\nselection.\nInspect `ours-fleet status Name`, `peek Name`, role logs, and\n`~/.ours-fleet/agents/Name/.monitor-status` when diagnosing delivery.\n\n## Trusted owner channel\n\nAn ACP role may declare a separate, existing ours identity which fleet \u2014 never\nthe agent \u2014 binds:\n\n```yaml\nowner_channel:\n  identity: Coordinator Owner Channel\n  owners: [authenticated-owner-contact-cid]\n  agent: authenticated-managed-agent-cid\n  interrupt: false\n  progress_interval_ms: 30000\n  comments: true\n  attachments:\n    enabled: true\n    max_files_per_request: 4\n    max_file_bytes: 10485760\n    max_request_bytes: 20971520\n    retention_ms: 86400000\n    allowed_mime: [application/pdf, text/plain, image/png, audio/ogg]\n```\n\nThis does not replace the role identity. Normal identity mail remains untrusted\npeer input: the agent reads it through `get_messages` and replies through\n`send_message`. Mail arriving on the dedicated channel from a CID in `owners`\nis injected as a direct `[fleet-owner]` prompt. Mail from the exact `agent`\nCID is forwarded as a new message to the latest authenticated owner conversation;\nits files may also be relayed through this channel. A reply reference selects the\nowner of that authenticated source wire instead of the latest conversation.\nEvery other CID is rejected and warned about without reflecting its body. Fleet sends\naccepted/queued/progress/interrupted/failure notices and routes the ACP turn's\nfinal assistant text back to the authenticated sender with its source wire ID.\nFor file replies, fleet injects a request-specific outbox path into the owner\nprompt. The agent copies completed artifacts there; fleet sends every regular\nfile from the channel identity with the same source wire ID and removes the\ntemporary outbox only after successful delivery. For proactive or in-turn agent\nattachments, the agent calls ours `send_file` to the channel identity and may\npair it with a reply-linked caption; fleet, not the agent, chooses the owner.\nOwner messages whose trimmed text starts with `/` are deterministic\nsupervisor commands and never enter the model: `/help` (alias `/commands`),\n`/status`, `/comments [status|on|off]`, `/interrupt`, `/clear`,\n`/compact`, `/model <model-id>`, `/restart`, `/force-restart`, `/ls`,\n`/peek`, `/worklog`, and\n`/version`. Unknown or malformed commands answer with the help text instead of\nbeing forwarded; plain messages reach the agent unchanged. `/clear`,\n`/compact`, and `/model` are forwarded only when the role's bundled ACP\nadapter executes them locally (claude-code: all three; codex: `/compact`\nonly) and are otherwise refused with a notice, so slash text never reaches the\nmodel as a prompt.\n\nWhile a request runs, the agent's live ACP commentary is relayed as messages\nprefixed with the single stable label `\uD83D\uDFE1 Live update:`, so an owner can see\nexactly which messages the setting controls. `owner_channel.comments`\n(default `true`, so existing channels keep their current behavior) is the\nRESTART BASELINE; `/comments on|off` changes only the running session and is\ndeliberately not persisted, so a restart always returns to the checked-in\nconfiguration. `/comments status` reports the live value, the baseline, and\nwhether the backend emits live comments at all. Suppressing live comments never\nsuppresses receipts, progress notices, or the final answer.\n\nOwner documents, images, and voice messages use the same authenticated sender\nand source-wire boundary. Fleet inspects body-free metadata first and rejects\ndisabled, over-count, over-size, or disallowed-MIME requests before selective\nretrieval. Unauthorized CIDs are never retrieved or answered. Reply-linked text\nand files from the same sender become one ordered request; a file-only wake also\nstarts a turn. Retrieved bytes must match their structured size and SHA-256,\ntheir content signature must match the declared MIME, and symlinks or non-regular\npaths fail closed. Sanitized copies live only in a mode-0700 request directory as\nmode-0600 files and are removed after completion or bounded stale retention.\n\nVoice prompts include a bounded transcript only when ours-mcp reports success.\nFailure or unavailability is explicit and preserves the private audio path as the\nfallback. Run `ours-mcp voice-status --json` to inspect the host configuration.\nA mode-0600 crash journal contains only authenticated CID and wire routing data;\nit never stores captions, filenames, paths, transcript text, or bytes. Journaled\npost-retrieval files resume selectively through `save_file`. A deferred agent\ncaption is replayed with its processed files before the group is admitted. Fleet\nresolves one authenticated owner route before retrieving bytes, admits every file\nbefore emitting the caption or any file, and sends every part to that same route.\nUnknown correlated routes remain queued without retrieval and receive one bounded\ncorrelated notice. Admission rejection consumes the whole group with one NACK;\nonce emission starts, a transport error becomes terminal uncertain delivery and\nthe group is never blind-retried. Bounded v2 source-wire routing state is migrated\nfrom v1 on read. Corrupt state disables attachment admission rather than weakening\nprovenance checks.\n\nThe channel identity must be unique and must not be a role identity. The bridge\npersists bounded wire IDs only, never message/reply plaintext, and requeues input\nbefore starting its turn for at-least-once crash recovery. It currently requires\n`session: acp`: tmux has no structured, turn-correlated final answer, and pane\nscraping cannot provide the same reliable reply guarantee.\n\n### Live contact and owner administration\n\nThe supervisor which is already running the ACP role remains the sole binder of\n`owner_channel.identity`. The CLI reaches that exact live `OwnerChannel`\nthrough the role's token-authenticated, mode-0600 Unix control socket for contact\ninspection and setup; it never starts another ours client and never force-binds:\n\nRapid supervised restart is serialized by a role-scoped single-binder lease.\nThe predecessor closes its authenticated control socket and MCP proxy before\nreleasing ownership. The replacement waits at most five seconds and retries the\ndaemon bind only when PID/start-marker metadata proves the holder was the same\nrole and owner-channel identity. Foreign, live, corrupt, or otherwise\nunverifiable ownership remains fail-closed; fleet never uses `force=true`.\n\nIf that matching predecessor misses the bound, its still-authenticated control\nroute may send one fixed, digest-deduplicated recovery notice through the latest\nauthenticated owner conversation (or the sole configured owner). Notice\nplaintext is never persisted. With no safe deterministic route fleet guesses no\nrecipient and leaves the actionable failure in the web console and role logs.\nThe remote recovery action is `/restart`; inspect repeated failures with\n`ours-fleet logs <Role>` or the web console.\n\n```sh\nours-fleet owner-channel contact list <Role>\nours-fleet owner-channel contact invite <Role> [--name <label>]\nours-fleet owner-channel contact add <Role> (--invite-file <path> | --invite-stdin) [--name <label>]\nours-fleet owner-channel owner list <Role>\nours-fleet owner-channel owner authorize <Role> <exact-64-hex-contact-cid>\nours-fleet owner-channel owner revoke <Role> <exact-64-hex-contact-cid>\n```\n\nContact establishment and owner authorization are separate security steps.\n`contact add` never authorizes: invite redemption is pending until the peer\nverifies it. Once `contact list` reports the established contact, authorize\nits exact immutable CID explicitly. Invite creation emits invite material only\non stdout; acceptance reads it from a file or stdin, not argv.\n\nConfigured `owners` remain the baseline. On legacy channels without `agent`,\nlive authorizations/revocations are an immediately effective, restart-persistent\noverlay. Managed-agent CID gating makes fleet configuration authoritative and\ndisables live owner mutation and direct control-socket sends. `owner list` labels\nbaseline versus dynamic entries and effective status. The atomic mode-0600 file\ncontains bounded CIDs and audit actions only. Corruption disables all effective\nowners and refuses mutation rather than resurrecting authority; revoking the\nlast effective owner is always refused.\n\nA missing/stopped role, tmux session, role without `owner_channel`, unavailable\nMCP client, or a role entering shutdown returns an actionable error with no\nside effects. Management uses no network listener and never logs or persists\ninvite material.\n\nFor any non-final message\u2014progress, blocker, suggestion, or later proactive note\u2014\nthe managed agent calls ordinary ours `send_message` to the channel identity.\nFleet checks only that the authenticated sender CID exactly equals `agent`, then\nforwards the text as a new message. There is no task/request/update type, phase,\nreply correlation, or owner recipient argument. A sole owner is the safe fallback;\nwith multiple owners and no inbound route history the relay fails closed. Devices\nsharing one identity share its CID; separate owner identities hand off the route\nwhen either sends channel mail. The ACP final is separate: fleet extracts it from\nthe completed turn and deterministically replies to the initiating owner wire.\n\nThe bounded mode-0600 route state stores CIDs, wire IDs, timestamps, delivery state,\nand hashes but never message plaintext. Unauthorized attempts produce a bounded\nCID-only owner warning; attempted bodies are neither reflected nor persisted.\n\nFor a mobile owner, establish the contact first, wait for peer verification,\nauthorize its exact CID, and revoke that same CID when access ends. The bounded\nmode-0600 CID overlay survives supervisor restart and remains fail-closed on\ncorruption. Update bodies remain memory-only. After a crash/restart, unfinished\ndeferred owner input follows the existing at-least-once replay path; the restarted\nsupervisor remains the sole binder.\n\n## Stable config and YAML migration\n\n`ours-fleet config --json` emits schemaVersion 1 resolved plans. Environment\nvalues and mission/persona/bio bodies are withheld; environment keys are sorted\nand values are marked redacted. Additive fields may appear in schema 1, while a\nremoval or semantic reuse requires a new schema version.\n\nYAML parsing always rejects duplicate keys. The current default\n`--yaml-mode compat` warns with file/line/column for anchors, aliases, explicit\ntags, non-scalar keys, and multiple documents. Use `--yaml-mode strict` in CI\nnow; strict becomes the next-major default and compat is the temporary migration\nescape hatch.\n\n## Bounded worklogs, auth proxy, and model recovery\n\nAn optional `worklog: { max_kb, keep_tail_kb, max_archives }` policy rotates a\nstable snapshot at fleet-owned lifecycle points. Concurrent changes defer\nrotation. Archives remain beside WORKLOG.md with the same sensitive-state\nboundary; retention deletes only recognized fleet archive names.\n\n`auth_proxy: { kind: anthropic, base_url, required, health_url }` is Claude-only\nand loopback-only. Fleet injects only ANTHROPIC_BASE_URL and doctor rejects\ncredential env keys. The privileged reference companion is\n`contrib/anthropic-auth-proxy.mjs`; deploy it separately as a dedicated account\nwith a 0600 token file and per-role listener access. Fleet never installs it or\nreads its credential.\n\n`model_chain` is an ordered authorization list and its first entry must equal\n`model`. Only sustained high-confidence entitlement/quota 429 evidence advances\none entry. Transient 429, overload, auth, policy, and unknown errors never\ndown-shift. Runtime state is atomic in .model-recovery.json; exhaustion is\nfail-closed and held down. Change the declared chain/model and restart to\nreconcile explicitly; no chain preserves detection-only behavior.\n";
+export declare const AI_DOCS = "# ours-fleet reference\n\nours-fleet runs persistent or temporary, identity-bound AI roles. A role selects\na harness independently from its session backend:\n\n- harness: `claude-code` or `codex`\n- session: `tmux` (default) or `acp`\n- lifetime: permanent (supervised, restartable) or `spawn --temp`\n\n## Discover and validate\n\n```sh\nours-fleet docs                         # this complete reference (`man` is an alias)\nours-fleet help <command>               # exact flags for one command\nours-fleet config [-c FILE]             # validate and print the merged plan; no changes\nours-fleet doctor [-c FILE] [--harness codex|claude-code]\nours-fleet version [--json]             # build identity, capabilities, every install on PATH\n```\n\nDefault configuration is `~/fleet.yaml` plus sorted `~/fleet.d/*.yaml` role\ndrop-ins. An explicit `-c FILE` replaces `~/fleet.yaml`; fleet.d still adds\nroles. Validate with `config` and `doctor` before starting or restarting.\n\nThe CLI never writes the base file: `spawn` writes `~/fleet.d/Name.yaml`. The\nweb console does write it, as a whole document \u2014 its setup wizard and\nconfiguration editor may create, change or remove any top-level block, including\n`vars:`, `defaults:`, `roles:`, `watchdogs:` and `loops:`. Only the base\nfile may hold `defaults:`, `watchdogs:` and `loops:`; a fleet.d drop-in may\ndeclare `roles:` and nothing else. Unrecognised top-level keys are round-tripped\nuntouched. Console edits are applied as surgical splices against the file's exact\nbytes, so an unchanged save is byte-identical and lines outside the edit keep their\ncomments and spacing. One exception: changing the length of a block sequence\n(`watch:`, `oversee:`, `roles:`, `wake_sources:`) may replace that collection\nwholesale and drop inline comments written on its items; lines outside that\ncollection remain byte-preserved. Each save is revision-guarded, reviewed as a diff\nof the real file before anything is written, validated by the real loader, and\nbacked up next to the file first.\n\n## Build identity and install provenance\n\n`--version` prints a semver and nothing else, and a semver does NOT identify an\nartifact. Version bumps land in a release commit of their own, so every build cut\nbetween two releases carries the PREVIOUS version while already containing new\nbehaviour. One host ran two installs that both reported 0.16.0 \u2014 same version,\ndifferent build. One accepted `monitor.interrupt: after_tool`, the other\nrejected it as invalid. Their\n`dist/cli.js` were byte-identical \u2014 the divergence was in other modules.\n\nEvery build therefore stamps `dist/build-info.json` with a build id (first 12 hex\nof a sha256 over the rest of `dist/`), the commit it was cut from, and the\ncapability tokens the shipped code declares \u2014 for example\n`monitor.interrupt.after_tool`. Ask any executable what it is:\n\n```sh\nours-fleet version                      # ours-fleet 0.17.0+9f1c2a3b4d5e, capabilities, PATH installs\nours-fleet version --json               # the same as machine-readable JSON, no environment values\n```\n\nRead a capability, never a version number, to decide whether a setting is\nsupported. When a build rejects a value it knows the name of, it says which\ncapability is missing and which build rejected it, because another install on the\nsame host may accept the identical file. `config` prints the build that resolved\nthe plan; `status <Name>` says so when the build reporting on a role is not the\none that created it (roles record their creating build in `creation.json`).\n\n`ours-fleet doctor` runs an `install` check that lists every `ours-fleet` on\nPATH plus the one executing, and FAILS when two installs share a semver but are\ndifferent builds, or when the running artifact is a DIFFERENT artifact from the\none PATH resolves to. A second prefix holding identical content is not a skew\nand is not reported. A PATH entry the shell would not execute \u2014 a directory, or\na file without its execute bit \u2014 is not counted as an install at all.\nInstalls built before this stamp existed report `+unknown`; they are compared by\nhashing their `dist/` instead, so two pre-provenance installs are still told\napart. To fix a flagged host, remove or update the stale install \u2014 do not rely on\nPATH order.\n\n## Lifecycle and console commands\n\n```sh\nours-fleet init\nours-fleet up|down [Name...]\nours-fleet restart [Name...]            # preserve/resume harness context\nours-fleet force-restart [Name...]      # fresh context; briefing is reloaded\nours-fleet ls\nours-fleet status|peek|attach|logs Name\nours-fleet logs -f Name\nours-fleet send Name \"prompt\"\nours-fleet send Name --key Enter        # tmux only\nours-fleet rm Name\nours-fleet watchdog-report <name> [run-id] [--list] [--json]\nours-fleet watchdog-run <name>\n```\n\n`peek`, `attach`, and text `send` work with tmux and ACP. ACP attachment\nalso accepts `/permit <permission-id> <option-id>`, `/interrupt`, and\n`/detach`. Raw `--key` input is tmux-only.\n\n## Local web console\n\nThe npm package includes the web console; installed users do not clone the repo\nor run `npm run build`:\n\n```sh\nnpm i -g @ours.network/fleet\nours-fleet init\nours-fleet doctor\nours-fleet web                         # install/update service, start, pair browser\n```\n\nThe normal command uses stable `http://127.0.0.1:49271/`, installs an\nowner-level systemd user service (Linux) or LaunchAgent (macOS), and opens a\nfive-minute one-use pairing link in the local browser. After pairing, bookmark\nthe plain URL or install the PWA. To pair a new, signed-out, or revoked browser,\nrun `ours-fleet web open`.\n\n```sh\nours-fleet web status\nours-fleet web start|stop|restart\nours-fleet web open\nours-fleet web revoke-all              # revoke every browser and active session\nours-fleet web uninstall\nours-fleet web serve --port 0 --no-open # isolated foreground/testing mode\n```\n\nThe console is IPv4-loopback-only by default. Both `localhost` and\n`127.0.0.1` are accepted locally. For an nginx/TLS reverse proxy, keep the\ndefault bind and declare the exact browser origin:\n\n`ours-fleet web install --public-origin https://fleet.example.com --password-file /secure/fleet-password`\n\nFleet reads the password file during setup and persists only a salted scrypt\nverifier. New browsers authenticate and retain rotating HttpOnly/SameSite\ntrusted-device credentials. If nginx already authenticates, the operator may\ndeliberately select `--no-password`; the CLI and browser warn that anyone\nreaching the origin can control the fleet. First setup requires an explicit\nchoice: `--password-file` or `--pairing` for protected access, or\n`--no-password` for intentional unprotected access.\n\nUse `--bind ADDRESS` only for an intentional direct listen. A non-loopback\nbind is rejected unless `--public-origin` is also present. Host/Origin checks\nuse the declaration and do not trust forwarded headers. Configure nginx to\nproxy HTTP and WebSocket upgrades to `127.0.0.1:49271` and terminate TLS;\nfleet accepts nginx's loopback upstream Host, so no Host rewrite is required.\nBrowser credentials add Secure for HTTPS, and `revoke-all` invalidates all\ntrusted devices. Role creation offers harness-scoped known-model choices\nwhile still accepting a typed model ID; blank explicitly uses the selected\nharness's own default.\n\n## Spawn\n\n```sh\nours-fleet spawn [--temp] [Name | --role Name] \\\n  --harness codex|claude-code --session tmux|acp \\\n  --mission \"one line\" --cwd /absolute/path --identity Identity \\\n  --coordinator Coordinator --model MODEL \\\n  --approval ask|auto|allow \\\n  --filesystem read-only|workspace|unrestricted \\\n  --unattended deny|wait \\\n  --bio-file /path/bio.md --persona-file /path/persona.md\n```\n\nPermanent spawn writes `~/fleet.d/Name.yaml` and starts a supervised role.\n`--temp` writes active state under `~/.ours-fleet/tmp` and starts an independent\ntransient supervisor (a collected systemd unit or submitted launchd job). It is\nnot enabled across reboot and does not die when the role that spawned it restarts.\nBoth lifetimes support `--session acp`. When a temporary role's bound identity\ncloses or its session ends, the supervisor, monitor and live roster entry retire\ntogether; state moves intact to `~/.ours-fleet/recovery/temporary` with a\ntermination record. Failed launches use the same archive rather than deleting\ntheir briefing, provenance, logs or partial supervisor metadata.\n\nNamed `down` and `rm` commands can target an exact state-backed temporary role\neven though it is absent from merged fleet YAML. The recorded transient unit/job\nis authoritative. Missing/incomplete ownership metadata is reconciled only from\nan exact `_run-temp <role>` process-table match: one match may be adopted, zero\nsettles as stopped, and ambiguity or an unreadable table fails closed. Launching\nrecords receive a bounded grace so a not-yet-registered transient unit cannot be\nmistaken for a stopped one. Stale recorded supervisors are reclaimed in bounded\nbatches by moving their state to the same recovery archive, never by blind deletion.\n\nTemporary-role identity bootstrap is capability-based. The generated briefing\nfirst tries to bind the exact assigned identity and preserves it when it already\nexists. If missing, it uses ours MCP `create_temporary_identity` when that tool\nis exposed, tying a newly-created identity to the connector session lifecycle;\nolder servers fall back to `create_identity`. Collisions and creation errors\nstop safely without force-adopting or deleting identity state. Permanent roles\nretain normal `create_identity` behavior.\n\nThe temporary supervisor treats its first positive identity observation as the\nlifecycle readiness gate: a cold harness may take as long as needed to read its\nbriefing and bind, without a fixed first-bind retirement timer. After readiness,\nonly sustained authoritative absence closes the role. Unreachable, malformed, or\nvalid-but-empty daemon indexes are ambiguous and reset closure debounce rather\nthan becoming cleanup authority.\n\nInside a managed ACP role, the same CLI automatically routes a real `spawn`\nthrough that role's authenticated supervisor control socket. `--role Name` is\naccepted as an alternative to the positional name, so a minimal delegated call\nis `ours-fleet spawn --role DeveloperX --temp`. The supervisor records the\ncalling role, performs creation, and only after success sends a structured\nspawn notice through the caller's owner channel when one is configured.\n\nOmitted harness, session, working directory, coordinator, neutral permissions,\nfleet monitor policy, and (when the harness is unchanged) model inherit from the\ncalling role. Explicit options always win. Selecting a different harness without\n`--model` leaves model selection to that harness/fleet defaults rather than\ncopying an incompatible caller model. This automatic proxy is a convenience and\nattribution mechanism, not an isolation boundary: an unrestricted role can still\ninvoke another binary path directly. Tmux roles and host/operator shells keep the\nordinary direct CLI behavior.\n\nCodex-specific spawn flags: `--sandbox`, `--permission-mode`, `--launcher`,\n`--profile`, `--search`, repeatable `--codex-config key=value`, repeatable\n`--add-dir`, and legacy `--monitor` (consent for the native Codex monitor,\nnot the `monitor.mode` wake-owner selector). Run `ours-fleet help spawn` for\nexact values.\n\n## fleet.yaml\n\n```yaml\nvars:\n  work_root: /home/me/work\nstart_stagger_ms: 0\ndefaults:\n  harness: codex\n  session: acp\n  model: gpt-model-id\n  permissions:\n    approval: ask\n    filesystem: workspace\n    unattended: deny\n  monitor:\n    mode: fleet                          # fleet (default) | native\nroles:\n  Coordinator:\n    harness: codex\n    session: acp\n    identity: Coordinator\n    cwd: ${work_root}/project\n    mission: Coordinate work and delegate implementation.\n    model: gpt-model-id\n    permissions:\n      approval: ask\n      filesystem: workspace\n      unattended: deny\n    session_options:                    # advanced overrides; normally omit\n      # acp:\n      #   command: [/custom/codex-acp, --flag]\n      tmux:\n        boot_grace_ms: 10000\n    monitor:\n      mode: fleet                        # fleet supervisor | native harness monitor\n      interrupt: false                    # false queues; true cancels; after_tool steers at an ACP tool boundary\n      wake_sources: [message_received, file_received, local_contact_request, pending_message]\n      batch_ms: 2000\n      inject: notification\n      turn_fail_threshold: 3\n    harness_options:\n      launcher: auto\n      sandbox: workspace-write\n      approval: on-request\n      search: false\n      profile: fleet\n      add_dirs: [/data/shared]\n      config:\n        model_reasoning_effort: high\n    bio: Public role card and when peers should engage it.\n    persona: Local operating contract, boundaries, and escalation policy.\n    briefing_file: /absolute/custom-briefing.md\n    coordinator: AnotherCoordinator\n    env:\n      KEY: value\n    oversee:\n      - { role: Worker, interval: 5m }\nwatchdogs:\n  nightwatch:                       # [A-Za-z0-9_-], must not collide with a role name\n    coordinator: FleetCoordinator   # required \u2014 where alerts go\n    # everything below is optional\n    enabled: true                   # default true; false = configured but never scheduled\n    interval: 10m                   # default 10m; 30s | 10m | 2h, minimum 1m\n    watch: [Alice, CodexReviewer]   # explicit lists are exact; omit for configured + live temp roles\n    harness: claude-code            # default: defaults.harness\n    model: claude-fable-5           # default: same resolution rule roles use (resolveRoleModel)\n    session: acp                    # default: defaults.session\n    identity: Watchdog-nightwatch   # default: Watchdog-<name>\n    timeout: 5m                     # default 5m; a run past this is killed and recorded as error\n    keep_reports: 50                # default 50 reports retained per watchdog\n    alert_cooldown: 60m             # default 60m before the same finding alerts again\n    prompt_file: /abs/extra.md      # optional extra focus, APPENDED to the fixed contract\n    isolation:                      # optional; omitted means no OS sandbox, like an ordinary role\n      backend: bubblewrap           # when present, the ordinary role isolation schema applies\n      network: broker\n      fs: { read: [/opt/watch-data] }\n```\n\nA watchdog observes and reports; it never restarts, stops, spawns, or removes a\nrole, answers a pending permission, edits a workspace, or approves anything on\nthe owner's behalf. `watchdogs:` may appear only in the base config\n(`~/fleet.yaml` or `-c FILE`), not in `~/fleet.d/*.yaml` drop-ins.\nWatchdogs are not isolated by default. An explicit watchdog `isolation:` block\nuses the same policy schema as a role and is applied unchanged; declare every\nextra filesystem access required by a custom prompt there.\nWhen `watch:` is omitted, each run watches the configured roles plus temporary\nfleet roles that are live when the run starts. An explicit `watch:` list is\nnever augmented.\n\nRole values override defaults. `${name}` substitutes entries from `vars`.\nOther role fields include `max_tokens`, `autocompact_pct`, and `isolation`.\nUse README.md for the complete isolation policy and resource-cap schema.\n\nSupervised roles connect to the operator-configured ours daemon; they do not own its\nlifecycle. Fleet forces `OURS_AUTOSTART=0` in tmux and ACP child processes after role\nenvironment overlays. Start the shared daemon only through an explicit operator or\ninstaller/setup flow.\n\n## Permissions\n\nPrefer the harness-neutral `permissions` block:\n\n- `approval: ask|auto|allow`: portable permission policy. `deny` remains a\n  deprecated, fail-closed compatibility alias for existing fleet files.\n- `filesystem: read-only|workspace|unrestricted`: filesystem intent\n- `unattended: deny|wait`: what ACP does when no console can answer a request\n\nThe backend translates this common intent. Harness-native settings in\n`harness_options` take precedence where supplied. Do not choose\n`allow`/`unrestricted`, Codex `never`/`danger-full-access`, or Claude\n`bypassPermissions` without explicit authorization.\n\n### Creation-time isolation\n\n`ours-fleet spawn --isolation-file <path>` supplies a role's sandbox policy at\ncreation, so the FIRST launch is already confined \u2014 a role that only gains\n`isolation:` on a later `up` ran unsandboxed until then.\n\nThe file holds exactly the `isolation:` mapping documented above and nothing\nelse \u2014 the same schema, validated by the same code, so a policy written here\ncannot mean something different from the identical block in fleet.yaml:\n\n```yaml\nnetwork: deny\nfs:\n  read: [/opt/reference]\nresources:\n  mem: 2G\n```\n\nInvalid files are rejected before anything is created: no config, no state\ndirectory, no identity reservation. Works for both permanent and `--temp` roles.\n\n### Never-prompt failure\n\nThe failure this section exists to prevent leaves no error message anywhere.\n\nAn unattended role has no console. When the harness needs a permission decision\nthere is nobody to ask, so the request is refused INSIDE the harness \u2014 no\nprompt, no error, no log line. The agent simply does less than its briefing told\nit to, reports success, and nothing distinguishes that from having done the\nwork. Two settings produce it:\n\n1. a permission mode that suppresses the prompt without granting the action\n   (Claude `dontAsk`, which is why neutral `allow` maps to\n   `bypassPermissions` instead); and\n2. `unattended: deny`, which refuses every request that reaches it.\n\n**Automatic decisions are now recorded.** Every permission request decided\nwithout a human emits a completed event into\n`~/.ours-fleet/agents/<Name>/.session-events.jsonl` carrying the decision,\nwhether policy or a person made it, the policy that produced it\n(`permissions.unattended=deny` vs `permissions.approval=deny`/`=allow`),\nthe reason, and the option selected. `ours-fleet peek` and `attach` render\nthem. Automatic denial asks for a one-shot rejection, never a standing one, so a\nsingle unattended refusal cannot disable a tool for the rest of the session.\n\nA role that can auto-deny logs one line at startup saying so.\n\nTo detect an under-permissioned role BEFORE it runs, use the capability floor\nbelow: `ours-fleet doctor` fails such a role rather than letting it discover\nthe problem silently at work.\n\n### The unattended capability floor\n\nAn unattended role has no console, so a permission request cannot be answered \u2014\nit is refused, silently, inside the harness. The agent then does less than it\nwas told to and reports no error. To make that visible before launch,\n`ours-fleet config` and `ours-fleet doctor` resolve each role's neutral\npermissions through its harness and check the result against a fixed floor:\n\n- `read-state` \u2014 read its briefing, ROUTINES.md, and WORKLOG.md\n- `write-state` \u2014 append its WORKLOG and its own state files\n- `messaging` \u2014 bind its identity, send and receive ours mail\n- `monitor` \u2014 arm and observe its mail monitor\n- `workspace-edit` \u2014 edit and test files in its working directory\n- `status-commands` \u2014 run the inspection commands its briefing prescribes\n\n`doctor` reports this per role as `unattended floor: <Role>`. A role with\n`unattended: deny` that cannot meet the floor FAILS doctor, because it will\ndeny those requests with nobody to see it; with `unattended: wait` it warns,\nbecause a human can still attach and answer.\n\nSecurity meaning: `ask` maps to Codex `untrusted` and Claude `default`.\n`auto` selects Codex ACP `agent` (`on-request` + `workspace-write`) and\nClaude `acceptEdits`. `approval: allow` selects Codex ACP's fully\nnon-interactive yolo mode, reported as `agent-full-access` (`never` +\n`danger-full-access`), and Claude `bypassPermissions`. Codex tmux retains\nindependent approval and sandbox flags: `auto` is `on-request`, `allow`\nis `never`, and `filesystem` still selects the sandbox. These modes genuinely\npermit the actions the role was authorized to take \u2014\n`dontAsk` only suppresses the prompt while still refusing the action. Nothing\nother than an explicit `allow` becomes non-interactive. Legacy `deny` keeps\nits conservative Codex `on-request` / Claude `plan` translation. `allow` is therefore a real grant and\nrequires explicit authorization; per-role `isolation:` remains the outer\nboundary that a permission mode cannot cross.\n\nACP carries agent-advertised session mode IDs and `session/set_mode`, but those\nIDs are agent-specific and ACP defines no portable permission-policy capability.\nFleet therefore uses the ACP primitive where an adapter exposes a matching mode\nand otherwise performs the harness translation above. The bundled Codex ACP\nadapter couples approval and sandboxing in its advertised mode IDs. Neutral\n`allow` therefore selects `agent-full-access` and widens `filesystem:\nworkspace` or `read-only` to `danger-full-access`; neutral `auto` selects\n`agent` and `workspace-write` even when the neutral filesystem value differs.\nAn explicit `harness_options.sandbox` selects its corresponding ACP preset and\nstill wins, as does an explicit native approval override. `config` and\n`doctor` report a coupled-mode mismatch as approximate. Use per-role\n`isolation:` as the outer boundary for an `allow` ACP role. The live session\nreports both its effective normalized mode and the exact native mode selected.\n\nSee also: `spawn --approval/--filesystem/--unattended` set this intent at\ncreation, and `ours-fleet config` prints each role's neutral settings, their\nnative translation, and any warning \u2014 the same text `doctor` reports.\n\nClaude `harness_options`: `permission_mode` (default, acceptEdits, plan,\ndontAsk, bypassPermissions), `plugins`, `mem_palace`, and\n`mem_palace_midsession_autosave`.\n\nCodex `harness_options`: `launcher` (auto, ours-codex, codex), `sandbox`\n(read-only, workspace-write, danger-full-access), `approval` or\n`permission_mode` (untrusted, on-request, never), `profile`, `search`,\n`config`, `add_dirs`, and `monitor`.\n\n## ACP adapters\n\nThe maintained `@agentclientprotocol/codex-acp` and\n`@agentclientprotocol/claude-agent-acp` runtimes are bundled automatically as\noptional ours-fleet dependencies. The supervisor resolves their executable\nentrypoints internally, so default ACP roles do not depend on global PATH.\nThe maintained Claude adapter requires Node 22; tmux and Codex ACP continue to\nwork on the ours-fleet core minimum of Node 20.\n\nOverride an adapter only when necessary with `session_options.acp.command`\n(string or argv list). If optional dependencies were deliberately omitted,\nours-fleet falls back to a compatible globally installed `codex-acp` or\n`claude-agent-acp`. `ours-fleet doctor -c FILE` verifies the resolved adapter.\n\n## Reliable mail wake\n\n`monitor.mode` selects exactly one wake owner:\n\n- `fleet` (default): the ours-fleet supervisor consumes body-free daemon\n  events and advances its durable cursor only after delivery is accepted. ACP\n  uses live steering when supported and falls back to structured\n  `session/prompt`; tmux uses verified console injection.\n- `native`: ours-fleet starts no supervisor monitor; the generated briefing\n  instructs Claude Code or Codex to arm its harness-native wake mechanism.\n\nSet `monitor.interrupt: true` in fleet mode to cancel active work before every\nconfigured wake. Set it to `after_tool` to preserve an active ACP tool (and any\npending permission), then steer the wake at the first tool-terminal boundary\nwithout cancellation. A hung boundary is bounded at 120 seconds and falls back\nto non-cancelling steering/queueing; adapters without authenticated tool events\nuse the same conservative fallback. Explicit human/control interrupts remain\nimmediate. The policy is content-blind because the supervisor cannot inspect\nencrypted message bodies. Message bodies are released only when the role calls\nthe ours `get_messages` tool.\n\nThe default is `false`. For a temporary role whose mission intentionally arrives\nafter its readiness announcement, set `mode: fleet` and `interrupt: true`\nexplicitly. The readiness announcement does not change the transport: the\nmission remains ordinary ours mail, fleet injects only the body-free wake, and\nthe role calls `get_messages` before acting. Every later configured wake uses\nthe same interruption policy.\n\nLegacy `monitor.enabled: true|false` remains accepted as an alias for\n`mode: fleet|native`; use `mode` in new configuration. Codex's separate\n`harness_options.monitor: true` is native-monitor consent, not monitor-owner\nselection.\nInspect `ours-fleet status Name`, `peek Name`, role logs, and\n`~/.ours-fleet/agents/Name/.monitor-status` when diagnosing delivery.\n\n## Trusted owner channel\n\nAn ACP role may declare a separate, existing ours identity which fleet \u2014 never\nthe agent \u2014 binds:\n\n```yaml\nowner_channel:\n  identity: Coordinator Owner Channel\n  owners: [authenticated-owner-contact-cid]\n  agent: authenticated-managed-agent-cid\n  interrupt: false\n  progress_interval_ms: 30000\n  comments: true\n  attachments:\n    enabled: true\n    max_files_per_request: 4\n    max_file_bytes: 10485760\n    max_request_bytes: 20971520\n    retention_ms: 86400000\n    allowed_mime: [application/pdf, text/plain, image/png, audio/ogg]\n```\n\nThis does not replace the role identity. Normal identity mail remains untrusted\npeer input: the agent reads it through `get_messages` and replies through\n`send_message`. Mail arriving on the dedicated channel from a CID in `owners`\nis injected as a direct `[fleet-owner]` prompt. Mail from the exact `agent`\nCID is forwarded as a new message to the latest authenticated owner conversation;\nits files may also be relayed through this channel. A reply reference selects the\nowner of that authenticated source wire instead of the latest conversation.\nEvery other CID is rejected and warned about without reflecting its body. Fleet sends\naccepted/queued/progress/interrupted/failure notices and routes the ACP turn's\nfinal assistant text back to the authenticated sender with its source wire ID.\nFor file replies of every kind \u2014 a response artifact, a proactive note, or an\nin-turn attachment \u2014 the agent calls ours `send_file` to the channel identity\nand may pair it with a reply-linked caption; fleet, not the agent, chooses the\nowner. That is the only delivery route an agent is given: a tool call either\ndelivers or reports an error, where a file written to disk does neither.\nOwner messages whose trimmed text starts with `/` are deterministic\nsupervisor commands and never enter the model: `/help` (alias `/commands`),\n`/status`, `/comments [status|on|off]`, `/interrupt`, `/clear`,\n`/compact`, `/model <model-id>`, `/restart`, `/force-restart`, `/ls`,\n`/peek`, `/worklog`, and\n`/version`. Unknown or malformed commands answer with the help text instead of\nbeing forwarded; plain messages reach the agent unchanged. `/clear`,\n`/compact`, and `/model` are forwarded only when the role's bundled ACP\nadapter executes them locally (claude-code: all three; codex: `/compact`\nonly) and are otherwise refused with a notice, so slash text never reaches the\nmodel as a prompt.\n\nWhile a request runs, the agent's live ACP commentary is relayed as messages\nprefixed with the single stable label `\uD83D\uDFE1 Live update:`, so an owner can see\nexactly which messages the setting controls. `owner_channel.comments`\n(default `true`, so existing channels keep their current behavior) is the\nRESTART BASELINE; `/comments on|off` changes only the running session and is\ndeliberately not persisted, so a restart always returns to the checked-in\nconfiguration. `/comments status` reports the live value, the baseline, and\nwhether the backend emits live comments at all. Suppressing live comments never\nsuppresses receipts, progress notices, or the final answer.\n\nOwner documents, images, and voice messages use the same authenticated sender\nand source-wire boundary. Fleet inspects body-free metadata first and rejects\ndisabled, over-count, over-size, or disallowed-MIME requests before selective\nretrieval. Unauthorized CIDs are never retrieved or answered. Reply-linked text\nand files from the same sender become one ordered request; a file-only wake also\nstarts a turn. Retrieved bytes must match their structured size and SHA-256,\ntheir content signature must match the declared MIME, and symlinks or non-regular\npaths fail closed. Sanitized copies live only in a mode-0700 request directory as\nmode-0600 files and are removed after completion or bounded stale retention.\n\nVoice prompts include a bounded transcript only when ours-mcp reports success.\nFailure or unavailability is explicit and preserves the private audio path as the\nfallback. Run `ours-mcp voice-status --json` to inspect the host configuration.\nA mode-0600 crash journal contains only authenticated CID and wire routing data;\nit never stores captions, filenames, paths, transcript text, or bytes. Journaled\npost-retrieval files resume selectively through `save_file`. A deferred agent\ncaption is replayed with its processed files before the group is admitted. Fleet\nresolves one authenticated owner route before retrieving bytes, admits every file\nbefore emitting the caption or any file, and sends every part to that same route.\nUnknown correlated routes remain queued without retrieval and receive one bounded\ncorrelated notice. Admission rejection consumes the whole group with one NACK;\nonce emission starts, a transport error becomes terminal uncertain delivery and\nthe group is never blind-retried. Bounded v2 source-wire routing state is migrated\nfrom v1 on read. Corrupt state disables attachment admission rather than weakening\nprovenance checks.\n\nThe channel identity must be unique and must not be a role identity. The bridge\npersists bounded wire IDs only, never message/reply plaintext, and requeues input\nbefore starting its turn for at-least-once crash recovery. It currently requires\n`session: acp`: tmux has no structured, turn-correlated final answer, and pane\nscraping cannot provide the same reliable reply guarantee.\n\n### Live contact and owner administration\n\nThe supervisor which is already running the ACP role remains the sole binder of\n`owner_channel.identity`. The CLI reaches that exact live `OwnerChannel`\nthrough the role's token-authenticated, mode-0600 Unix control socket for contact\ninspection and setup; it never starts another ours client and never force-binds:\n\nRapid supervised restart is serialized by a role-scoped single-binder lease.\nThe predecessor closes its authenticated control socket and MCP proxy before\nreleasing ownership. The replacement waits at most five seconds and retries the\ndaemon bind only when PID/start-marker metadata proves the holder was the same\nrole and owner-channel identity. Foreign, live, corrupt, or otherwise\nunverifiable ownership remains fail-closed; fleet never uses `force=true`.\n\nIf that matching predecessor misses the bound, its still-authenticated control\nroute may send one fixed, digest-deduplicated recovery notice through the latest\nauthenticated owner conversation (or the sole configured owner). Notice\nplaintext is never persisted. With no safe deterministic route fleet guesses no\nrecipient and leaves the actionable failure in the web console and role logs.\nThe remote recovery action is `/restart`; inspect repeated failures with\n`ours-fleet logs <Role>` or the web console.\n\n```sh\nours-fleet owner-channel contact list <Role>\nours-fleet owner-channel contact invite <Role> [--name <label>]\nours-fleet owner-channel contact add <Role> (--invite-file <path> | --invite-stdin) [--name <label>]\nours-fleet owner-channel owner list <Role>\nours-fleet owner-channel owner authorize <Role> <exact-64-hex-contact-cid>\nours-fleet owner-channel owner revoke <Role> <exact-64-hex-contact-cid>\n```\n\nContact establishment and owner authorization are separate security steps.\n`contact add` never authorizes: invite redemption is pending until the peer\nverifies it. Once `contact list` reports the established contact, authorize\nits exact immutable CID explicitly. Invite creation emits invite material only\non stdout; acceptance reads it from a file or stdin, not argv.\n\nConfigured `owners` remain the baseline. On legacy channels without `agent`,\nlive authorizations/revocations are an immediately effective, restart-persistent\noverlay. Managed-agent CID gating makes fleet configuration authoritative and\ndisables live owner mutation and direct control-socket sends. `owner list` labels\nbaseline versus dynamic entries and effective status. The atomic mode-0600 file\ncontains bounded CIDs and audit actions only. Corruption disables all effective\nowners and refuses mutation rather than resurrecting authority; revoking the\nlast effective owner is always refused.\n\nA missing/stopped role, tmux session, role without `owner_channel`, unavailable\nMCP client, or a role entering shutdown returns an actionable error with no\nside effects. Management uses no network listener and never logs or persists\ninvite material.\n\nFor any non-final message\u2014progress, blocker, suggestion, or later proactive note\u2014\nthe managed agent calls ordinary ours `send_message` to the channel identity.\nFleet checks only that the authenticated sender CID exactly equals `agent`, then\nforwards the text as a new message. There is no task/request/update type, phase,\nreply correlation, or owner recipient argument. A sole owner is the safe fallback;\nwith multiple owners and no inbound route history the relay fails closed. Devices\nsharing one identity share its CID; separate owner identities hand off the route\nwhen either sends channel mail. The ACP final is separate: fleet extracts it from\nthe completed turn and deterministically replies to the initiating owner wire.\n\nThe bounded mode-0600 route state stores CIDs, wire IDs, timestamps, delivery state,\nand hashes but never message plaintext. Unauthorized attempts produce a bounded\nCID-only owner warning; attempted bodies are neither reflected nor persisted.\n\nFor a mobile owner, establish the contact first, wait for peer verification,\nauthorize its exact CID, and revoke that same CID when access ends. The bounded\nmode-0600 CID overlay survives supervisor restart and remains fail-closed on\ncorruption. Update bodies remain memory-only. After a crash/restart, unfinished\ndeferred owner input follows the existing at-least-once replay path; the restarted\nsupervisor remains the sole binder.\n\n## Stable config and YAML migration\n\n`ours-fleet config --json` emits schemaVersion 1 resolved plans. Environment\nvalues and mission/persona/bio bodies are withheld; environment keys are sorted\nand values are marked redacted. Additive fields may appear in schema 1, while a\nremoval or semantic reuse requires a new schema version.\n\nYAML parsing always rejects duplicate keys. The current default\n`--yaml-mode compat` warns with file/line/column for anchors, aliases, explicit\ntags, non-scalar keys, and multiple documents. Use `--yaml-mode strict` in CI\nnow; strict becomes the next-major default and compat is the temporary migration\nescape hatch.\n\n## Bounded worklogs, auth proxy, and model recovery\n\nAn optional `worklog: { max_kb, keep_tail_kb, max_archives }` policy rotates a\nstable snapshot at fleet-owned lifecycle points. Concurrent changes defer\nrotation. Archives remain beside WORKLOG.md with the same sensitive-state\nboundary; retention deletes only recognized fleet archive names.\n\n`auth_proxy: { kind: anthropic, base_url, required, health_url }` is Claude-only\nand loopback-only. Fleet injects only ANTHROPIC_BASE_URL and doctor rejects\ncredential env keys. The privileged reference companion is\n`contrib/anthropic-auth-proxy.mjs`; deploy it separately as a dedicated account\nwith a 0600 token file and per-role listener access. Fleet never installs it or\nreads its credential.\n\n`model_chain` is an ordered authorization list and its first entry must equal\n`model`. Only sustained high-confidence entitlement/quota 429 evidence advances\none entry. Transient 429, overload, auth, policy, and unknown errors never\ndown-shift. Runtime state is atomic in .model-recovery.json; exhaustion is\nfail-closed and held down. Change the declared chain/model and restart to\nreconcile explicitly; no chain preserves detection-only behavior.\n";
 /**
  * What every shipped spawn-skill variant must say, and must not say (7.1).
  *

package/dist/docs.js

@@ -526,12 +526,11 @@ owner of that authenticated source wire instead of the latest conversation.
 Every other CID is rejected and warned about without reflecting its body. Fleet sends
 accepted/queued/progress/interrupted/failure notices and routes the ACP turn's
 final assistant text back to the authenticated sender with its source wire ID.
-For file replies, fleet injects a request-specific outbox path into the owner
-prompt. The agent copies completed artifacts there; fleet sends every regular
-file from the channel identity with the same source wire ID and removes the
-temporary outbox only after successful delivery. For proactive or in-turn agent
-attachments, the agent calls ours \`send_file\` to the channel identity and may
-pair it with a reply-linked caption; fleet, not the agent, chooses the owner.
+For file replies of every kind — a response artifact, a proactive note, or an
+in-turn attachment — the agent calls ours \`send_file\` to the channel identity
+and may pair it with a reply-linked caption; fleet, not the agent, chooses the
+owner. That is the only delivery route an agent is given: a tool call either
+delivers or reports an error, where a file written to disk does neither.
 Owner messages whose trimmed text starts with \`/\` are deterministic
 supervisor commands and never enter the model: \`/help\` (alias \`/commands\`),
 \`/status\`, \`/comments [status|on|off]\`, \`/interrupt\`, \`/clear\`,

package/dist/loops/manager.d.ts

@@ -68,8 +68,32 @@ export declare class ScheduledLoopManager implements ScheduledLoopManagerHandle
     private armAbandon;
     private finish;
     private advance;
+    /**
+     * Coalesce a backlog into one skip. The counters alone say how many
+     * occurrences were lost but never when or for how long, so the window is
+     * recorded too and carried on the state until a run is actually told about it
+     * — a dropped pass has to stay visible to the next one, not just to whoever
+     * was reading the log at the time.
+     */
     private skipMissed;
-    private skipRestartMisses;
+    /**
+     * Restart is not, by itself, a reason to lose an occurrence a running manager
+     * would still have run. `poll` tolerates lateness up to one full interval and
+     * runs the tick late; this path used to drop anything already due however
+     * recently, so a role restarted seconds after its own tick came due lost it
+     * outright. For an oversight role that is precisely the pass which would have
+     * recorded why it restarted, so the failure erased its own witness.
+     *
+     * The tolerance is the only thing shared with `poll`. A backlog at least one
+     * interval deep is still coalesced into a single skip and never replayed —
+     * after a long outage exactly one occurrence survives, and `schedule` then
+     * arms it through the ordinary path rather than firing a burst here.
+     *
+     * Running the survivor late cannot outpace the configured cadence: `advance`
+     * moves the cursor by exactly one `intervalMs` per occurrence from the nominal
+     * time, so a loop that keeps restarting still runs at most once per interval.
+     */
+    private skipRestartBacklog;
     /**
      * A run the store could not record is dropped, not retried: the cursor has
      * already moved, so this can never become a busy loop, and the outage is
@@ -86,5 +110,10 @@ export declare class ScheduledLoopManager implements ScheduledLoopManagerHandle
      * until the process was restarted.
      */
     private recover;
+    /**
+     * The envelope is the only channel a scheduled pass has for learning about
+     * the passes that did not happen. A gap stated here is what lets an oversight
+     * role report its own outage instead of resuming as if nothing was missed.
+     */
     private envelope;
 }

package/dist/loops/manager.js

@@ -39,7 +39,7 @@ export class ScheduledLoopManager {
     }
     start() {
         if (!this.store.fresh)
-            this.skipRestartMisses();
+            this.skipRestartBacklog();
         this.schedule();
     }
     async stop() {
@@ -143,10 +143,15 @@ export class ScheduledLoopManager {
     async attempt(definition, state, scheduledAt) {
         const runId = `sl_${randomUUID()}`;
         const origin = { kind: 'scheduled-loop', loop: definition.name, runId };
-        const prompt = this.envelope(definition, runId, scheduledAt);
+        // The gap is read here and cleared only if the turn is actually admitted:
+        // an attempt that ends `skipped_busy` or `unavailable` reported it to
+        // nobody, so it has to still be there for the attempt that succeeds.
+        const gap = state.missedGap;
+        const prompt = this.envelope(definition, runId, scheduledAt, gap);
         let claimed = false;
         const result = await this.arbiter.tryScheduled(prompt, origin, () => {
             claimed = true;
+            state.missedGap = null;
             state.activeRunId = runId;
             state.lastRunId = runId;
             state.lastStartedAt = new Date(this.deps.now()).toISOString();
@@ -265,7 +270,15 @@ export class ScheduledLoopManager {
         state.nextScheduledAt = new Date(next).toISOString();
         state.nextDueAt = new Date(next + deterministicJitter(this.role, definition.name, next, definition.jitterMs)).toISOString();
     }
+    /**
+     * Coalesce a backlog into one skip. The counters alone say how many
+     * occurrences were lost but never when or for how long, so the window is
+     * recorded too and carried on the state until a run is actually told about it
+     * — a dropped pass has to stay visible to the next one, not just to whoever
+     * was reading the log at the time.
+     */
     skipMissed(definition, state, now) {
+        const from = state.nextScheduledAt;
         let missed = 0;
         while (Date.parse(state.nextDueAt) <= now) {
             this.advance(definition, state);
@@ -275,14 +288,43 @@ export class ScheduledLoopManager {
         state.counts.skippedMissed = increment(state.counts.skippedMissed, missed);
         state.lastOutcome = 'skipped_missed';
         state.lastFinishedAt = new Date(now).toISOString();
+        // Successive outages before any run lands merge into one gap: the earliest
+        // start wins, so the window always spans the whole silence.
+        const previous = state.missedGap;
+        state.missedGap = {
+            count: increment(previous?.count ?? 0, missed),
+            fromAt: previous?.fromAt ?? from,
+            throughAt: state.lastScheduledAt ?? from,
+            detectedAt: new Date(now).toISOString(),
+        };
         this.store.persist();
-        this.deps.log(`[${this.role}] loop ${definition.name} skipped_missed count=${missed}`);
+        this.deps.log(`[${this.role}] loop ${definition.name} skipped_missed count=${missed} `
+            + `gap=${from}..${state.missedGap.throughAt} `
+            + `unreported=${state.missedGap.count}`);
     }
-    skipRestartMisses() {
+    /**
+     * Restart is not, by itself, a reason to lose an occurrence a running manager
+     * would still have run. `poll` tolerates lateness up to one full interval and
+     * runs the tick late; this path used to drop anything already due however
+     * recently, so a role restarted seconds after its own tick came due lost it
+     * outright. For an oversight role that is precisely the pass which would have
+     * recorded why it restarted, so the failure erased its own witness.
+     *
+     * The tolerance is the only thing shared with `poll`. A backlog at least one
+     * interval deep is still coalesced into a single skip and never replayed —
+     * after a long outage exactly one occurrence survives, and `schedule` then
+     * arms it through the ordinary path rather than firing a burst here.
+     *
+     * Running the survivor late cannot outpace the configured cadence: `advance`
+     * moves the cursor by exactly one `intervalMs` per occurrence from the nominal
+     * time, so a loop that keeps restarting still runs at most once per interval.
+     */
+    skipRestartBacklog() {
         const now = this.deps.now();
         for (const definition of this.definitions.values()) {
             const state = this.store.state.loops[definition.name];
-            if (definition.enabled && !state.operatorDisabled && Date.parse(state.nextDueAt) <= now)
+            if (definition.enabled && !state.operatorDisabled
+                && now >= Date.parse(state.nextDueAt) + definition.intervalMs)
                 this.skipMissed(definition, state, now);
         }
     }
@@ -340,17 +382,38 @@ export class ScheduledLoopManager {
             this.deps.clearTimer(this.timer);
         this.arm(backoffMs(this.pollFailures));
     }
-    envelope(definition, runId, scheduledAt) {
+    /**
+     * The envelope is the only channel a scheduled pass has for learning about
+     * the passes that did not happen. A gap stated here is what lets an oversight
+     * role report its own outage instead of resuming as if nothing was missed.
+     */
+    envelope(definition, runId, scheduledAt, gap) {
+        const lateBy = Math.max(0, this.deps.now() - scheduledAt);
         return [
             '[fleet-loop]',
             `loop: ${definition.name}`,
             `run: ${runId}`,
             `scheduled_at: ${new Date(scheduledAt).toISOString()}`,
+            ...(lateBy > 0 ? [`started_late_by_ms: ${lateBy}`] : []),
+            ...(gap ? [
+                `missed_occurrences: ${gap.count}`,
+                `missed_window: ${gap.fromAt}..${gap.throughAt}`,
+                `missed_gap_ms: ${Math.max(0, Date.parse(gap.detectedAt) - Date.parse(gap.fromAt))}`,
+            ] : []),
             'origin: local-trusted-config',
             '',
             'This is a scheduled internal maintenance turn, not an owner message and not ordinary ours mail.',
             'Perform one bounded pass. Do not wait for the next tick. Do not report to an owner unless your',
             'configured policy and an existing authenticated proactive-report route authorize a material report.',
+            // Same single route as the owner-request prompt, and for the same reason.
+            'To send a file, call ours `send_file` with the recipient and the path — to your owner-channel',
+            'identity if this role has one, otherwise directly to the contact who should receive it.',
+            'A file written anywhere else is not delivered and nothing will report that it was not.',
+            ...(gap ? ['',
+                'This loop did not run for the window above: those occurrences were coalesced away while the role',
+                'was unavailable, and this pass is the first since. Treat the gap as part of what you are reporting',
+                'on — it is the record of your own outage, and no later pass will be told about it.',
+            ] : []),
             '',
             definition.prompt,
         ].join('\n');

package/dist/loops/state.d.ts

@@ -8,11 +8,29 @@ export interface LoopCounts {
     skippedBusy: number;
     skippedMissed: number;
 }
+/**
+ * A coalesced run of occurrences that were never submitted, held until a run
+ * actually starts and can be told about it. Without it a dropped occurrence
+ * survives only as a counter, which says how many were lost but never when or
+ * for how long — and an oversight role cannot report an outage it cannot date.
+ */
+export interface LoopMissedGap {
+    /** Occurrences coalesced away, summed across every skip since the last run. */
+    count: number;
+    /** Nominal time of the earliest occurrence in the gap. */
+    fromAt: string;
+    /** Nominal time of the latest occurrence in the gap. */
+    throughAt: string;
+    /** When the manager noticed — the end of the outage, not of the last skip. */
+    detectedAt: string;
+}
 export interface LoopRuntimeState {
     definitionHash: string;
     promptHash: string;
     enabled: boolean;
     operatorDisabled: boolean;
+    /** Unreported gap, cleared by the first run that carries it. */
+    missedGap: LoopMissedGap | null;
     nextScheduledAt: string;
     nextDueAt: string;
     lastScheduledAt: string | null;

package/dist/loops/state.js

@@ -100,6 +100,9 @@ export class ScheduledLoopStateStore {
             if (old?.definitionHash === definition.definitionHash) {
                 next[definition.name] = {
                     ...old, promptHash: definition.promptHash, enabled: definition.enabled,
+                    // A file written before this field existed restores as undefined; an
+                    // unreported gap is absent, not lost, so normalize rather than trust.
+                    missedGap: old.missedGap ?? null,
                 };
             }
             else {
@@ -119,6 +122,7 @@ export class ScheduledLoopStateStore {
                     activeRunId: old?.activeRunId ?? null,
                     counts: old?.counts ?? zeroCounts(), lastError: old?.lastError ?? null,
                     operatorDisabled: old?.operatorDisabled ?? false,
+                    missedGap: old?.missedGap ?? null,
                 };
             }
             if (recoverActive && next[definition.name].activeRunId) {

package/dist/owner-channel/channel.js

@@ -709,7 +709,7 @@ export class OwnerChannel {
             outbox = this.outboxDir(originWireId);
             await mkdir(outbox, { recursive: true, mode: 0o700 });
             const activityCursor = this.latestEventSeq(this.options.session.eventsSince(0));
-            const queued = await this.options.session.queuePrompt(this.ownerAttachmentPrompt(sender, originWireId, requestId, outbox, admitted, group.caption), {
+            const queued = await this.options.session.queuePrompt(this.ownerAttachmentPrompt(sender, originWireId, requestId, admitted, group.caption), {
                 interrupt: this.options.config.interrupt,
                 ...(this.options.config.interrupt ? { interruptSource: 'owner' } : {}),
                 origin: { kind: 'owner', requestId,
@@ -838,7 +838,7 @@ export class OwnerChannel {
         let queued;
         const activityCursor = this.latestEventSeq(this.options.session.eventsSince(0));
         try {
-            queued = await this.options.session.queuePrompt(this.ownerPrompt(sender, text, wireId, outbox), {
+            queued = await this.options.session.queuePrompt(this.ownerPrompt(sender, text, wireId), {
                 interrupt: this.options.config.interrupt,
                 ...(this.options.config.interrupt ? { interruptSource: 'owner' } : {}),
                 origin: { kind: 'owner', requestId, displayText: text },
@@ -1453,7 +1453,7 @@ export class OwnerChannel {
             throw new Error('commentary appears to contain secret, reasoning, or raw tool content');
         return message;
     }
-    ownerAttachmentPrompt(sender, wireId, requestId, outbox, files, caption) {
+    ownerAttachmentPrompt(sender, wireId, requestId, files, caption) {
         const lines = [
             '[fleet-owner]',
             `Authenticated owner ${safeField(sender.name, 160)} (${sender.id}) sent owner-channel attachment request ${wireId}.`,
@@ -1477,10 +1477,13 @@ export class OwnerChannel {
                     lines.push(`- voice transcript status: ${transcription?.status ?? 'unavailable'}`, `- voice transcript fallback: audio path above; category ${transcription?.errorCategory ?? 'not_provided'}`);
             }
         }
-        lines.push('Answer in your final assistant response; fleet routes it only to the authenticated sender and correlates it to the originating file wire.', `To attach response files, write regular files only to: ${outbox}`);
+        lines.push('Answer in your final assistant response; fleet routes it only to the authenticated sender and correlates it to the originating file wire.', 
+        // send_file is the only delivery route an agent is given: a tool call either
+        // delivers or reports an error, where a file written to disk does neither.
+        'To send the owner a file — now or later in this turn — call ours `send_file`:', `contact: ${this.options.config.identity}`, 'and the path of the finished file. Fleet routes it to the authenticated owner.', 'A file written anywhere else is not delivered and nothing will report that it was not.', 'Use descriptive unique filenames. Send nothing the owner did not request or should not receive.');
         return lines.join('\n');
     }
-    ownerPrompt(sender, text, wireId, outbox) {
+    ownerPrompt(sender, text, wireId) {
         return [
             '[fleet-owner]',
             `Authenticated owner ${sender.name} (${sender.id}) sent owner-channel message ${wireId}.`,
@@ -1494,10 +1497,13 @@ export class OwnerChannel {
             ] : [
                 'Managed-agent outbound relay is not configured; do not send intermediate or proactive owner-channel messages.',
             ]),
-            'To attach files to your response, copy each finished file directly into this fleet outbox:',
-            outbox,
-            'Fleet sends every regular file in that directory to the authenticated owner, correlated to this request.',
-            'Use descriptive unique filenames. Put nothing there that the owner did not request or should not receive.',
+            // send_file is the only delivery route an agent is given: a tool call either
+            // delivers or reports an error, where a file written to disk does neither.
+            'To send the owner a file — now or later in this turn — call ours `send_file`:',
+            `contact: ${this.options.config.identity}`,
+            'and the path of the finished file. Fleet routes it to the authenticated owner.',
+            'A file written anywhere else is not delivered and nothing will report that it was not.',
+            'Use descriptive unique filenames. Send nothing the owner did not request or should not receive.',
             '',
             text || '(empty message)',
         ].join('\n');

package/dist/session/acp.d.ts

@@ -5,6 +5,21 @@ import type { ConversationSnapshot, PromptOrigin, PromptReceipt, SubmitPromptCom
 import type { ConversationHandlePage, ExitRecord, InterruptOutcome, QueuedPrompt, SessionEvent, RuntimeSelectorMetadata, SessionHandle, SessionSnapshot, SubmitPromptOptions, TurnCancellationSource, TurnOutcome, TurnResult } from './types.js';
 /** Bound safe-boundary waiting without turning a hung tool into cancellation. */
 export declare const AFTER_TOOL_BOUNDARY_TIMEOUT_MS = 120000;
+/**
+ * How long a steering-started turn is presumed to still own the adapter after
+ * its last update. Such a turn has no prompt id, so it never reports a
+ * stopReason and there is no exact end to observe — silence is the only signal
+ * available, and this is the bound that turns it into a decision.
+ *
+ * Sized from the fleet's own scheduled-run history: across 1513 completed
+ * scheduled runs the longest silence WITHIN a working turn was 120.2 s (p99
+ * 41.0 s; 5 runs above 60 s). A shorter grace would release the lease while the
+ * adapter is still working and re-admit a prompt into a busy turn, which is the
+ * FLEET-003 failure itself. The costs are deliberately asymmetric: holding too
+ * long skips one best-effort maintenance tick, releasing too early SIGTERMs a
+ * live role.
+ */
+export declare const STEERING_OCCUPANCY_IDLE_MS = 150000;
 /** Server-generated typed provenance followed by the exact human-authored body. */
 export declare function promptContentBlocks(text: string, origin?: PromptOrigin): acp.ContentBlock[];
 export declare function runtimeSelector(options: acp.SessionConfigOption[] | null | undefined, category: string): RuntimeSelectorMetadata | undefined;
@@ -33,6 +48,8 @@ export interface AcpSessionOptions {
     controllerGraceMs?: number;
     /** Test seam; production uses AFTER_TOOL_BOUNDARY_TIMEOUT_MS. */
     afterToolBoundaryTimeoutMs?: number;
+    /** Test seam; production uses STEERING_OCCUPANCY_IDLE_MS. */
+    steeringOccupancyIdleMs?: number;
 }
 /**
  * Classify an ACP `stopReason` into a terminal outcome. A refusal and a
@@ -81,6 +98,13 @@ export declare class AcpSession implements SessionHandle {
     private cancelEscalation?;
     private cancelForceKill?;
     private cancelRecoveryReason?;
+    /**
+     * Held while a steering-started turn is believed to own the adapter. It is a
+     * lease, not a latch: `steeringRelease` always fires, so the role can never be
+     * stranded busy by a wake whose turn ended without telling anyone.
+     */
+    private steeringOccupied;
+    private steeringRelease?;
     /**
      * Rejects the moment the adapter process is gone. Every in-flight ACP request
      * races it, so a dead adapter can never leave a turn — and therefore a
@@ -102,6 +126,20 @@ export declare class AcpSession implements SessionHandle {
      */
     private recoverOpenPrompts;
     isAlive(): boolean;
+    /**
+     * Take the occupancy lease for a turn the adapter started on its own behalf.
+     * Refreshed by every adapter update, so it tracks work actually happening
+     * rather than a fixed guess at how long a wake takes.
+     */
+    private holdSteeringOccupancy;
+    private refreshSteeringOccupancy;
+    /**
+     * Every exit from occupancy comes through here, including the ones that are
+     * not the timer: a real turn boundary, close, and adapter exit. A lease that
+     * can leak is worse than the bug it fixes — it would leave the role reporting
+     * `running` forever and starve scheduled admission permanently.
+     */
+    private releaseSteeringOccupancy;
     snapshot(): SessionSnapshot;
     private toolCall;
     private reserveTool;

package/dist/session/acp.js

@@ -16,6 +16,21 @@ const PERMISSION_TIMEOUT_MS = 10 * 60_000;
 const CONTROLLER_GRACE_MS = 12_000;
 /** Bound safe-boundary waiting without turning a hung tool into cancellation. */
 export const AFTER_TOOL_BOUNDARY_TIMEOUT_MS = 120_000;
+/**
+ * How long a steering-started turn is presumed to still own the adapter after
+ * its last update. Such a turn has no prompt id, so it never reports a
+ * stopReason and there is no exact end to observe — silence is the only signal
+ * available, and this is the bound that turns it into a decision.
+ *
+ * Sized from the fleet's own scheduled-run history: across 1513 completed
+ * scheduled runs the longest silence WITHIN a working turn was 120.2 s (p99
+ * 41.0 s; 5 runs above 60 s). A shorter grace would release the lease while the
+ * adapter is still working and re-admit a prompt into a busy turn, which is the
+ * FLEET-003 failure itself. The costs are deliberately asymmetric: holding too
+ * long skips one best-effort maintenance tick, releasing too early SIGTERMs a
+ * live role.
+ */
+export const STEERING_OCCUPANCY_IDLE_MS = 150_000;
 const TERMINAL_TOOL_STATUSES = new Set(['completed', 'failed']);
 const SCHEDULED_LOOP_REDACTION = '[scheduled-loop content redacted]';
 const OWNER_COMMENTARY_REDACTION = '[assistant commentary redacted]';
@@ -223,6 +238,13 @@ export class AcpSession {
     cancelEscalation;
     cancelForceKill;
     cancelRecoveryReason;
+    /**
+     * Held while a steering-started turn is believed to own the adapter. It is a
+     * lease, not a latch: `steeringRelease` always fires, so the role can never be
+     * stranded busy by a wake whose turn ended without telling anyone.
+     */
+    steeringOccupied = false;
+    steeringRelease;
     /**
      * Rejects the moment the adapter process is gone. Every in-flight ACP request
      * races it, so a dead adapter can never leave a turn — and therefore a
@@ -253,6 +275,7 @@ export class AcpSession {
             if (this.cancelForceKill)
                 clearTimeout(this.cancelForceKill);
             this.cancelForceKill = undefined;
+            this.releaseSteeringOccupancy('adapter exited');
             // Record the child's real exit code/signal. The tmux path can only see a
             // shell's `$?`; here the truth is available, so keep it.
             const classified = classifyChildExit(code, signal);
@@ -353,11 +376,50 @@ export class AcpSession {
         // terminal fact (signal exits deliberately leave exitCode null).
         return this.child.exitCode === null && (this.child.signalCode ?? null) === null;
     }
+    /**
+     * Take the occupancy lease for a turn the adapter started on its own behalf.
+     * Refreshed by every adapter update, so it tracks work actually happening
+     * rather than a fixed guess at how long a wake takes.
+     */
+    holdSteeringOccupancy() {
+        if (this.closing || !this.isAlive())
+            return;
+        this.steeringOccupied = true;
+        this.refreshSteeringOccupancy();
+    }
+    refreshSteeringOccupancy() {
+        if (!this.steeringOccupied)
+            return;
+        if (this.steeringRelease)
+            clearTimeout(this.steeringRelease);
+        this.steeringRelease = setTimeout(() => this.releaseSteeringOccupancy('adapter silent'), this.options.steeringOccupancyIdleMs ?? STEERING_OCCUPANCY_IDLE_MS);
+        this.steeringRelease.unref?.();
+    }
+    /**
+     * Every exit from occupancy comes through here, including the ones that are
+     * not the timer: a real turn boundary, close, and adapter exit. A lease that
+     * can leak is worse than the bug it fixes — it would leave the role reporting
+     * `running` forever and starve scheduled admission permanently.
+     */
+    releaseSteeringOccupancy(reason) {
+        if (this.steeringRelease)
+            clearTimeout(this.steeringRelease);
+        this.steeringRelease = undefined;
+        if (!this.steeringOccupied)
+            return;
+        this.steeringOccupied = false;
+        this.options.log(`[${this.options.name}] steering-started turn no longer holds the adapter (${reason})`);
+    }
     snapshot() {
         return {
             backend: 'acp',
             alive: this.isAlive(),
-            readiness: this.readiness,
+            // A steering-started turn is real work with no prompt id. Reporting the
+            // session idle while it runs is what let the arbiter admit a scheduled
+            // prompt into a busy adapter, whose `session/prompt` then never returned
+            // a stopReason and ended in a cancellation deadline and a SIGTERM.
+            readiness: this.readiness === 'idle' && this.steeringOccupied
+                ? 'running' : this.readiness,
             sessionId: this.sessionId,
             lastError: this.lastError,
             pendingPermissionId: this.pendingPermissions.keys().next().value,
@@ -865,6 +927,7 @@ export class AcpSession {
         if (this.controllerGrace)
             clearTimeout(this.controllerGrace);
         this.controllerGrace = undefined;
+        this.releaseSteeringOccupancy('session closed');
         for (const [permissionId, pending] of [...this.pendingPermissions])
             this.settlePendingAutomatically(permissionId, pending, 'cancelled', undefined, 'the session closed while this request was pending');
         this.releaseAllTools();
@@ -1023,6 +1086,10 @@ export class AcpSession {
         }
         finally {
             this.releaseAllTools();
+            // A turn this client owned has ended, so the adapter has reported a
+            // boundary: whatever a steering call started before it is over too. This
+            // is the release path that does not depend on the silence timer.
+            this.releaseSteeringOccupancy('turn boundary');
             if (this.activeTurn?.id === turnId) {
                 this.activeTurn.settle();
                 if (this.cancelEscalation)
@@ -1045,6 +1112,12 @@ export class AcpSession {
             ]);
             if (response.outcome === 'failed')
                 return turnResult(false, 'failed', 'ACP steering failed');
+            // `injected` joined a turn this client already owns and will settle.
+            // `startedNewTurn` created one nobody owns: the adapter is working and
+            // will never answer for it, so admission has to learn about it here or
+            // not at all.
+            if (response.outcome === 'startedNewTurn')
+                this.holdSteeringOccupancy();
             return turnResult(true, 'inconclusive', response.outcome);
         }
         catch (error) {
@@ -1230,9 +1303,13 @@ export class AcpSession {
     }
     recordUpdate(update) {
         // Replayed history is not current activity: `session/load` would otherwise
-        // make a cold session look like it had just been working.
-        if (!this.replaying)
+        // make a cold session look like it had just been working. The same reason
+        // keeps it from extending the steering lease, which is evidence the adapter
+        // is working right now — for a steering-started turn, the only evidence.
+        if (!this.replaying) {
             this.lastUpdateAt = new Date().toISOString();
+            this.refreshSteeringOccupancy();
+        }
         const scheduled = this.activeTurn?.origin?.kind === 'scheduled-loop';
         const messagePhase = update.sessionUpdate === 'agent_message_chunk'
             ? this.codexMessagePhase(update) : undefined;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ours.network/fleet",
-  "version": "0.17.6",
+  "version": "0.17.8",
   "description": "Harness-agnostic fleet of persistent, identity-bound AI agents. Declarative fleet.yaml, tmux or ACP sessions, supervision, and ours.network messaging.",
   "type": "module",
   "license": "FSL-1.1-Apache-2.0",