@ours.network/fleet 0.17.9 → 0.17.11

package/README.md

@@ -214,6 +214,22 @@ unknown evidence; a newly launched harness follows its generated first-boot
 instructions to choose or create and bind the identity. The console never
 claims that the host created an identity and never deletes one.
 
+ACP tool diffs are normalized before they enter the durable conversation
+ledger. Small diffs retain their existing before/after representation. When an
+adapter reports an oversized whole-file snapshot, fleet stores only the changed
+region with path, operation, original byte counts, digest, and boundedness
+metadata. Each retained changed side is capped at 64 KiB as a UTF-8-safe,
+newest-content tail. The tail advances to a line boundary when a complete line
+fits. If one logical line alone exceeds the cap, fleet retains its newest
+UTF-8-safe suffix and records that it starts mid-line, together with omitted-byte
+count and digest. Paths retain at most a 4 KiB suffix with the same explicit
+size/digest/omission provenance, and a 320 KiB cap covers the complete normalized
+update. An append therefore cannot replay a large historical file while hiding
+the current appended text. The live web-console transcript projects only events
+from the current runner generation and excludes adapter `session/load` replay;
+those replay events remain in the durable ledger with `agent_replay` provenance
+for diagnosis and recovery rather than appearing as current work.
+
 For a temporary role, those first-boot instructions preserve and bind an
 existing identity when one is present. If the assigned identity is missing,
 the role capability-detects the ours MCP `create_temporary_identity` tool and
@@ -344,6 +360,10 @@ defaults:
   max_tokens: 500000                    # session cap (harness-interpreted)
   monitor:
     mode: fleet                         # fleet (default) | native
+  worklog:                              # built-ins shown; set false to opt out
+    max_kb: 1024                        # rotate only above this active-log size
+    keep_tail_kb: 256                   # UTF-8 tail; line-aligned when one fits
+    max_archives: 12                    # recent beside log; older preserved cold
 roles:
   Name:                                 # [A-Za-z0-9_-]+
     harness: claude-code
@@ -1120,7 +1140,7 @@ rollout, anchors, aliases, explicit tags, non-scalar keys, and multiple document
 produce source-positioned warnings; opt into enforcement with
 `--yaml-mode strict`. Strict mode will become the next-major default.
 
-Long-running roles may opt into bounded durable logs:
+WORKLOG rotation is enabled for roles by default with conservative built-ins:
 
 ```yaml
 worklog:
@@ -1130,8 +1150,23 @@ worklog:
 ```
 
 Rotation is conservative: a concurrent change aborts the attempt and retries at
-a later fleet lifecycle point. Archives remain in the role state directory and
-may contain the same sensitive material as `WORKLOG.md`.
+a later fleet lifecycle point. The active log retains a bounded UTF-8 tail,
+advancing to a line boundary when a complete line fits. If one logical line
+alone exceeds the tail budget, its newest suffix remains and
+`.worklog-rotation.json` explicitly records the mid-line start and omitted byte
+count. It also records SHA-256 digests for the archive and retained live bytes
+observed when the manifest is written. The complete prior inode is published
+under a collision-safe UTC name.
+`max_archives` bounds recent archives beside `WORKLOG.md`; older complete
+archives move to `WORKLOG.archives/` and are never deleted. Use `worklog: false` on a role or in
+`defaults` to opt out. Archives may contain the same sensitive material as
+`WORKLOG.md`. Rotation refuses a symlinked/non-regular live log or a symlinked
+cold-archive boundary before replacing the live path, and best-effort removes a
+duplicate link left by a detected failure while the original inode is still
+available. These are ordinary path/error safeguards, not a security boundary
+against a malicious concurrent process with the same Unix authority: intentional
+symlink swaps or archive-directory renames between checks are outside the threat
+model and require OS-level isolation from that process.
 
 Claude roles can use a credential-free loopback proxy:
 

package/dist/briefing.js

@@ -183,8 +183,10 @@ export function generateBriefing(role, v, opts) {
     L.push('it survives restarts.');
     if (role.worklog) {
         L.push(`Fleet rotates it above ${role.worklog.max_kb} KiB, keeps approximately the newest ` +
-            `${role.worklog.keep_tail_kb} KiB here, and retains ${role.worklog.max_archives} archives ` +
-            'beside it. Continue writing only WORKLOG.md.');
+            `${role.worklog.keep_tail_kb} KiB here, and keeps ${role.worklog.max_archives} recent archives ` +
+            'beside it. Older complete archives move to WORKLOG.archives without deletion; ' +
+            '.worklog-rotation.json identifies the latest archive for restart provenance. ' +
+            'Continue writing only WORKLOG.md.');
     }
     L.push('', '## Routines');
     L.push(`If \`${opts.routinesPath}\` exists, re-read it at the START of every wake — before acting`);

package/dist/build-info.json

@@ -1,9 +1,9 @@
 {
-  "version": "0.17.9",
-  "buildId": "6f77d2fb4e88",
-  "commit": "94827734295ad22ecb6691c099e960413a322970",
+  "version": "0.17.11",
+  "buildId": "1f620d26f86b",
+  "commit": "2ce98775d61c1d06f5d1b86290a8cd6a8af0afdd",
   "dirty": false,
-  "builtAt": "2026-08-19T09:41:14.713Z",
+  "builtAt": "2026-08-20T15:57:33.324Z",
   "capabilities": [
     "monitor.interrupt.after_tool"
   ]

package/dist/config.d.ts

@@ -11,6 +11,9 @@ export interface WorklogPolicy {
     keep_tail_kb: number;
     max_archives: number;
 }
+export type WorklogPolicyInput = Partial<WorklogPolicy> | false;
+/** Conservative built-in policy; `worklog: false` is the explicit opt-out. */
+export declare const DEFAULT_WORKLOG_POLICY: Readonly<WorklogPolicy>;
 export interface AuthProxyConfig {
     kind: 'anthropic';
     base_url: string;
@@ -130,10 +133,10 @@ export interface RoleConfig {
     isolation?: IsolationConfig;
     monitor?: Partial<MonitorConfig>;
     owner_channel?: OwnerChannelConfigInput;
-    worklog?: WorklogPolicy;
+    worklog?: WorklogPolicyInput;
     auth_proxy?: Partial<AuthProxyConfig>;
 }
-export interface ResolvedRole extends Omit<RoleConfig, 'model' | 'owner_channel'> {
+export interface ResolvedRole extends Omit<RoleConfig, 'model' | 'owner_channel' | 'worklog'> {
     name: string;
     harness: string;
     session: SessionBackendId;
@@ -193,7 +196,7 @@ export declare function canonicalCid(value: string): string;
 export declare function resolveOwnerChannelConfig(defaults: unknown, role: OwnerChannelConfigInput | undefined, session: SessionBackendId, file?: string, name?: string): OwnerChannelConfig | undefined;
 export declare function resolveModelChain(model: string | undefined, chain: string[] | undefined, file?: string, name?: string): string[] | undefined;
 export declare function resolveAuthProxy(defaults: unknown, role: Partial<AuthProxyConfig> | undefined, file?: string, name?: string): AuthProxyConfig | undefined;
-export declare function resolveWorklogPolicy(defaults: unknown, role: WorklogPolicy | undefined, file?: string, name?: string): WorklogPolicy | undefined;
+export declare function resolveWorklogPolicy(defaults: unknown, role: WorklogPolicyInput | undefined, file?: string, name?: string): WorklogPolicy | undefined;
 export declare function resolvePermissions(defaults: unknown, role: Partial<CommonPermissions> | undefined, file?: string, name?: string): CommonPermissions;
 /**
  * Merge `defaults.monitor` under the role's own `monitor:` key-by-key, validate the

package/dist/config.js

@@ -4,10 +4,17 @@ import { agentDir, defaultConfigPath, fleetDDir, home } from './paths.js';
 import { parseFleetDocument, } from './config-yaml.js';
 import { harnessRuntimeDir, resolveIsolation, validateIsolationConfig, } from './isolation/policy.js';
 import { getAdapter } from './harness/registry.js';
+import { resolveRoleModelEnv } from './model-env.js';
 import { resolveWatchdogs } from './watchdog/config.js';
 import { resolveLoops } from './loops/config.js';
 import { CAPABILITIES, CAP_MONITOR_INTERRUPT_AFTER_TOOL } from './capabilities.js';
 import { runningLabel } from './provenance.js';
+/** Conservative built-in policy; `worklog: false` is the explicit opt-out. */
+export const DEFAULT_WORKLOG_POLICY = Object.freeze({
+    max_kb: 1024,
+    keep_tail_kb: 256,
+    max_archives: 12,
+});
 /** The 8 content-free event types the ours daemon appends to notifications.log. */
 export const NOTIFY_EVENT_TYPES = [
     'message_received', 'file_received', 'sibling_contact_added', 'local_contact_request',
@@ -217,17 +224,23 @@ export function loadConfig(configPath, options = {}) {
             const harness = r.harness ?? defaults.harness ?? 'claude-code';
             const defaultHarness = defaults.harness ?? 'claude-code';
             const inheritsModelDefaults = harness === defaultHarness && r.model !== null;
-            const model = resolveRoleModel(r.model, r.harness, defaults);
+            if (authProxy && harness !== 'claude-code')
+                throw new ConfigError(`${file}: role '${name}' auth_proxy is supported only by claude-code`);
+            // Environment and runtime model are resolved together so `model:` and the
+            // harness's model pin can never disagree (see src/model-env.ts).
+            const modelEnv = resolveRoleModelEnv({
+                harness,
+                model: resolveRoleModel(r.model, r.harness, defaults),
+                modelWasExplicit: r.model !== undefined,
+                defaultsEnv: (defaults.env ?? {}),
+                roleEnv: r.env,
+                ...(authProxy ? { authProxyBaseUrl: authProxy.base_url } : {}),
+            }, message => new ConfigError(`${file}: role '${name}' ${message}`));
+            const env = modelEnv.env;
+            const model = modelEnv.model;
             const modelChain = resolveModelChain(model, r.model_chain ?? (inheritsModelDefaults
                 ? defaults.model_chain
                 : undefined), file, name);
-            if (authProxy && harness !== 'claude-code')
-                throw new ConfigError(`${file}: role '${name}' auth_proxy is supported only by claude-code`);
-            const env = {
-                ...(defaults.env ?? {}),
-                ...(r.env ?? {}),
-                ...(authProxy ? { ANTHROPIC_BASE_URL: authProxy.base_url } : {}),
-            };
             roles.push({
                 ...r,
                 name,
@@ -461,15 +474,16 @@ export function resolveAuthProxy(defaults, role, file = 'config', name = 'role')
     };
 }
 export function resolveWorklogPolicy(defaults, role, file = 'config', name = 'role') {
-    if (defaults === undefined && role === undefined)
+    if (role === false || (role === undefined && defaults === false))
         return undefined;
-    if (defaults !== undefined && !isPlainObject(defaults))
-        throw new ConfigError(`${file}: defaults.worklog must be a map`);
+    if (defaults !== undefined && defaults !== false && !isPlainObject(defaults))
+        throw new ConfigError(`${file}: defaults.worklog must be a map or false`);
     if (role !== undefined && !isPlainObject(role))
-        throw new ConfigError(`${file}: role '${name}' worklog must be a map`);
+        throw new ConfigError(`${file}: role '${name}' worklog must be a map or false`);
     const merged = {
-        ...(defaults ?? {}),
-        ...(role ?? {}),
+        ...DEFAULT_WORKLOG_POLICY,
+        ...(defaults === false || defaults === undefined ? {} : defaults),
+        ...(role === undefined ? {} : role),
     };
     const bad = Object.keys(merged).filter(key => !['max_kb', 'keep_tail_kb', 'max_archives'].includes(key));
     if (bad.length)

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      mcp_servers:                       # claude-code: per-role MCP servers, additive by default\n        ours: { command: ours-mcp, args: [proxy] }\n      mcp_servers_only: false            # true = ONLY these; drops user/project/plugin servers\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`,\n`mem_palace_midsession_autosave`, `mcp_servers` and `mcp_servers_only`.\n\n`mcp_servers` declares MCP servers for the role, in `.mcp.json`'s own shape\n(a map of name to `{ command, args, env }`, or `{ type: http|sse, url,\nheaders }`). By default they are ADDED to whatever the OS user running the role\nalready has configured, on both session types: tmux passes `--mcp-config`, and\nACP sends them in `session/new`.\n\n`mcp_servers_only: true` makes the declared set EXCLUSIVE \u2014 `--strict-mcp-config`\non tmux, `strictMcpConfig` on ACP. It is all-or-nothing and it ignores every\nother MCP configuration: project `.mcp.json`, user settings, and **plugins**.\nThe ours connector is normally installed as a plugin, so a strict role that does\nnot re-declare it has no `send_message` and no `get_messages` \u2014 it cannot even\nreport that it has gone mute. Fleet therefore refuses a strict role whose\n`mcp_servers` does not name the connector; declare it explicitly, e.g.\n`ours: { command: ours-mcp, args: [proxy] }`.\n\nBoth options, and `plugins`, reach an ACP session through the bundled Claude ACP\nagent's `_meta` vocabulary. A role that sets `session_options.acp.command` runs\nan agent fleet did not choose and cannot be promised them, so that combination is\nrefused at validation rather than accepted and dropped. This narrows a role's\ntool surface; it does not stop the harness deferring tool schemas, which is the\nharness's own decision.\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";
+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      mcp_servers:                       # claude-code: per-role MCP servers, additive by default\n        ours: { command: ours-mcp, args: [proxy] }\n      mcp_servers_only: false            # true = ONLY these; drops user/project/plugin servers\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`,\n`mem_palace_midsession_autosave`, `mcp_servers` and `mcp_servers_only`.\n\n`mcp_servers` declares MCP servers for the role, in `.mcp.json`'s own shape\n(a map of name to `{ command, args, env }`, or `{ type: http|sse, url,\nheaders }`). By default they are ADDED to whatever the OS user running the role\nalready has configured, on both session types: tmux passes `--mcp-config`, and\nACP sends them in `session/new`.\n\n`mcp_servers_only: true` makes the declared set EXCLUSIVE \u2014 `--strict-mcp-config`\non tmux, `strictMcpConfig` on ACP. It is all-or-nothing and it ignores every\nother MCP configuration: project `.mcp.json`, user settings, and **plugins**.\nThe ours connector is normally installed as a plugin, so a strict role that does\nnot re-declare it has no `send_message` and no `get_messages` \u2014 it cannot even\nreport that it has gone mute. Fleet therefore refuses a strict role whose\n`mcp_servers` does not name the connector; declare it explicitly, e.g.\n`ours: { command: ours-mcp, args: [proxy] }`.\n\nBoth options, and `plugins`, reach an ACP session through the bundled Claude ACP\nagent's `_meta` vocabulary. A role that sets `session_options.acp.command` runs\nan agent fleet did not choose and cannot be promised them, so that combination is\nrefused at validation rather than accepted and dropped. This narrows a role's\ntool surface; it does not stop the harness deferring tool schemas, which is the\nharness's own decision.\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\nWORKLOG rotation is enabled by default with\n`worklog: { max_kb: 1024, keep_tail_kb: 256, max_archives: 12 }`. Maps may\noverride individual values; `worklog: false` on a role or in defaults opts out.\nFleet rotates only at that role's launch/resume lifecycle boundary. Concurrent\nchanges defer rotation. The active file keeps a bounded UTF-8 tail and advances\nto a line boundary when a complete line fits. If one logical line alone exceeds\nthe budget, its newest suffix remains and the rotation manifest records the\nmid-line start and omitted byte count. The complete prior inode receives a\ncollision-safe UTC archive name, and\n`.worklog-rotation.json` records restart provenance. `max_archives` bounds\nrecent archives beside WORKLOG.md; older complete archives move to\n`WORKLOG.archives/` without deletion. All archives share the role's sensitive\nstate boundary. Fleet refuses a symlinked/non-regular live log or a symlinked\ncold-archive boundary before replacing the live path and best-effort removes a\nduplicate publication left by a detected failure while the original inode is\nstill available. The manifest records SHA-256 digests for the archive and live\nbytes observed when it is written. These checks address ordinary path hazards,\nnot intentional path mutation by a malicious concurrent process with the same\nUnix authority; that is outside the threat model and requires OS-level isolation.\n\nACP tool diffs are bounded before entering web conversation events. Existing\nsmall before/after diffs are unchanged. Oversized whole-file snapshots are\nreduced to the actual changed region plus path, operation, original byte counts,\ndigest, and omission metadata. Each retained side is a newest-content UTF-8 tail\nof at most 64 KiB, advanced to a line boundary when a complete line fits. An\noverlong single line keeps its newest suffix and explicitly records a mid-line\nstart. Paths retain at most a 4 KiB suffix with byte count, digest, and omitted\nprefix metadata; the complete normalized update is capped at 320 KiB. A large\nappend therefore retains current appended content, not the historical prefix.\nThe live web-console transcript includes only the current runner generation and\nexcludes adapter session/load replay. Replayed events remain durable with\nagent_replay provenance for diagnosis and recovery.\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

@@ -697,10 +697,38 @@ escape hatch.
 
 ## Bounded worklogs, auth proxy, and model recovery
 
-An optional \`worklog: { max_kb, keep_tail_kb, max_archives }\` policy rotates a
-stable snapshot at fleet-owned lifecycle points. Concurrent changes defer
-rotation. Archives remain beside WORKLOG.md with the same sensitive-state
-boundary; retention deletes only recognized fleet archive names.
+WORKLOG rotation is enabled by default with
+\`worklog: { max_kb: 1024, keep_tail_kb: 256, max_archives: 12 }\`. Maps may
+override individual values; \`worklog: false\` on a role or in defaults opts out.
+Fleet rotates only at that role's launch/resume lifecycle boundary. Concurrent
+changes defer rotation. The active file keeps a bounded UTF-8 tail and advances
+to a line boundary when a complete line fits. If one logical line alone exceeds
+the budget, its newest suffix remains and the rotation manifest records the
+mid-line start and omitted byte count. The complete prior inode receives a
+collision-safe UTC archive name, and
+\`.worklog-rotation.json\` records restart provenance. \`max_archives\` bounds
+recent archives beside WORKLOG.md; older complete archives move to
+\`WORKLOG.archives/\` without deletion. All archives share the role's sensitive
+state boundary. Fleet refuses a symlinked/non-regular live log or a symlinked
+cold-archive boundary before replacing the live path and best-effort removes a
+duplicate publication left by a detected failure while the original inode is
+still available. The manifest records SHA-256 digests for the archive and live
+bytes observed when it is written. These checks address ordinary path hazards,
+not intentional path mutation by a malicious concurrent process with the same
+Unix authority; that is outside the threat model and requires OS-level isolation.
+
+ACP tool diffs are bounded before entering web conversation events. Existing
+small before/after diffs are unchanged. Oversized whole-file snapshots are
+reduced to the actual changed region plus path, operation, original byte counts,
+digest, and omission metadata. Each retained side is a newest-content UTF-8 tail
+of at most 64 KiB, advanced to a line boundary when a complete line fits. An
+overlong single line keeps its newest suffix and explicitly records a mid-line
+start. Paths retain at most a 4 KiB suffix with byte count, digest, and omitted
+prefix metadata; the complete normalized update is capped at 320 KiB. A large
+append therefore retains current appended content, not the historical prefix.
+The live web-console transcript includes only the current runner generation and
+excludes adapter session/load replay. Replayed events remain durable with
+agent_replay provenance for diagnosis and recovery.
 
 \`auth_proxy: { kind: anthropic, base_url, required, health_url }\` is Claude-only
 and loopback-only. Fleet injects only ANTHROPIC_BASE_URL and doctor rejects

package/dist/model-env.d.ts

@@ -0,0 +1,71 @@
+import type { ResolvedRole } from './config.js';
+/**
+ * Which environment variable a harness reads to pin the model it RUNS.
+ *
+ * This is not a convenience: for `claude-code` it is the only channel that
+ * reaches the ACP backend at all. `buildLaunch` (tmux) passes `--model`, but
+ * `buildAcpLaunch` launches the ACP adapter with no model argument, and that
+ * adapter resolves its model in this order — ANTHROPIC_MODEL, then
+ * `settings.model`, then a resumed session's live model, then its first
+ * catalogue entry. A role's declared model was therefore invisible to every
+ * ACP role, and a fleet-wide `defaults.env.ANTHROPIC_MODEL` silently outranked
+ * an explicitly requested one.
+ */
+export declare const MODEL_ENV_BY_HARNESS: Readonly<Record<string, string>>;
+/** The model-pin variable for a harness, or undefined if it pins no model by env. */
+export declare function modelEnvVar(harness: string | undefined): string | undefined;
+export interface RoleModelEnvInput {
+    harness: string;
+    /** Already resolved by `resolveRoleModel` — may come from the fleet default. */
+    model: string | undefined;
+    /** True when the role (or `--model`) named a model, including `model: null`. */
+    modelWasExplicit: boolean;
+    defaultsEnv?: Record<string, string>;
+    roleEnv?: Record<string, string>;
+    authProxyBaseUrl?: string;
+}
+export interface RoleModelEnv {
+    env: Record<string, string>;
+    /**
+     * The model the harness will actually run. Equal to `env[pin]` for a harness
+     * that pins by env, so anything reporting this value reports the runtime.
+     */
+    model: string | undefined;
+}
+/**
+ * Resolve a role's environment and its runtime model TOGETHER, so the two can
+ * never disagree.
+ *
+ * Precedence, highest first:
+ *   1. an explicit `model:` / `--model` on the role
+ *   2. the role's own `env:` pin
+ *   3. the fleet `defaults.model`
+ *   4. the fleet `defaults.env` pin
+ *
+ * Inheriting the fleet default remains correct when the role names no model
+ * (2, 3, 4); an explicitly named one wins (1). Where both are explicit and they
+ * disagree, there is no defensible winner, so this refuses rather than picking
+ * one silently — the silence is what let a day of "Fable" work run on Opus.
+ *
+ * `model: null` explicitly asks for no fleet-chosen model, so it also clears an
+ * inherited pin instead of leaving one in place to act as a hidden default.
+ */
+export declare function resolveRoleModelEnv(input: RoleModelEnvInput, describe?: (message: string) => Error): RoleModelEnv;
+/**
+ * The model a role will actually run, read back from the environment it was
+ * resolved with. Use this wherever a model is reported to a human.
+ */
+export declare function effectiveRoleModel(role: ResolvedRole): string | undefined;
+/**
+ * Move a role's env pin onto a new model. Anything that changes the model a
+ * role runs after resolution — model-chain recovery is the live example — must
+ * go through this, or it changes only the label.
+ */
+export declare function repinModelEnv(role: ResolvedRole, model: string | undefined): Record<string, string> | undefined;
+/**
+ * Last line of defence, at the exact point a child's environment is composed:
+ * refuse to launch a role whose child would run a model other than the one the
+ * role declares and the banner reports. A spawn that cannot keep those two in
+ * agreement must fail loudly, not start and be believed.
+ */
+export declare function assertModelPinReachesChild(role: ResolvedRole, childEnv: Record<string, string | undefined>): void;

package/dist/model-env.js

@@ -0,0 +1,106 @@
+/**
+ * Which environment variable a harness reads to pin the model it RUNS.
+ *
+ * This is not a convenience: for `claude-code` it is the only channel that
+ * reaches the ACP backend at all. `buildLaunch` (tmux) passes `--model`, but
+ * `buildAcpLaunch` launches the ACP adapter with no model argument, and that
+ * adapter resolves its model in this order — ANTHROPIC_MODEL, then
+ * `settings.model`, then a resumed session's live model, then its first
+ * catalogue entry. A role's declared model was therefore invisible to every
+ * ACP role, and a fleet-wide `defaults.env.ANTHROPIC_MODEL` silently outranked
+ * an explicitly requested one.
+ */
+export const MODEL_ENV_BY_HARNESS = {
+    'claude-code': 'ANTHROPIC_MODEL',
+};
+/** The model-pin variable for a harness, or undefined if it pins no model by env. */
+export function modelEnvVar(harness) {
+    return harness === undefined ? undefined : MODEL_ENV_BY_HARNESS[harness];
+}
+/**
+ * Resolve a role's environment and its runtime model TOGETHER, so the two can
+ * never disagree.
+ *
+ * Precedence, highest first:
+ *   1. an explicit `model:` / `--model` on the role
+ *   2. the role's own `env:` pin
+ *   3. the fleet `defaults.model`
+ *   4. the fleet `defaults.env` pin
+ *
+ * Inheriting the fleet default remains correct when the role names no model
+ * (2, 3, 4); an explicitly named one wins (1). Where both are explicit and they
+ * disagree, there is no defensible winner, so this refuses rather than picking
+ * one silently — the silence is what let a day of "Fable" work run on Opus.
+ *
+ * `model: null` explicitly asks for no fleet-chosen model, so it also clears an
+ * inherited pin instead of leaving one in place to act as a hidden default.
+ */
+export function resolveRoleModelEnv(input, describe = message => new Error(message)) {
+    const env = {
+        ...(input.defaultsEnv ?? {}),
+        ...(input.roleEnv ?? {}),
+        ...(input.authProxyBaseUrl ? { ANTHROPIC_BASE_URL: input.authProxyBaseUrl } : {}),
+    };
+    const pin = modelEnvVar(input.harness);
+    if (!pin)
+        return { env, model: input.model };
+    const rolePin = input.roleEnv?.[pin];
+    if (input.modelWasExplicit) {
+        if (rolePin !== undefined && rolePin !== input.model)
+            throw describe(`model '${input.model ?? '(none)'}' contradicts env.${pin} '${rolePin}'; `
+                + `remove one — ${pin} is what the harness actually runs`);
+        if (input.model === undefined)
+            delete env[pin];
+        else
+            env[pin] = input.model;
+        return { env, model: input.model };
+    }
+    // Not explicit: a role-level pin is the most specific thing said about this
+    // role, so it decides — and the reported model follows it.
+    if (rolePin !== undefined)
+        return { env, model: rolePin };
+    if (input.model !== undefined)
+        env[pin] = input.model;
+    return { env, model: input.model ?? env[pin] };
+}
+/**
+ * The model a role will actually run, read back from the environment it was
+ * resolved with. Use this wherever a model is reported to a human.
+ */
+export function effectiveRoleModel(role) {
+    const pin = modelEnvVar(role.harness);
+    return (pin ? role.env?.[pin] : undefined) ?? role.model;
+}
+/**
+ * Move a role's env pin onto a new model. Anything that changes the model a
+ * role runs after resolution — model-chain recovery is the live example — must
+ * go through this, or it changes only the label.
+ */
+export function repinModelEnv(role, model) {
+    const pin = modelEnvVar(role.harness);
+    if (!pin)
+        return role.env;
+    const env = { ...(role.env ?? {}) };
+    if (model === undefined)
+        delete env[pin];
+    else
+        env[pin] = model;
+    return Object.keys(env).length ? env : undefined;
+}
+/**
+ * Last line of defence, at the exact point a child's environment is composed:
+ * refuse to launch a role whose child would run a model other than the one the
+ * role declares and the banner reports. A spawn that cannot keep those two in
+ * agreement must fail loudly, not start and be believed.
+ */
+export function assertModelPinReachesChild(role, childEnv) {
+    const pin = modelEnvVar(role.harness);
+    if (!pin || role.model === undefined)
+        return;
+    const actual = childEnv[pin];
+    if (actual === role.model)
+        return;
+    throw new Error(`[${role.name}] refusing to launch: role model is '${role.model}' but the child's `
+        + `${pin} is ${actual === undefined ? 'unset' : `'${actual}'`} — the session would run a `
+        + 'different model than the one reported');
+}