v0.14.0

Changed

• BREAKING: the systemd templates run the packaged dsh runtime. scripts/install-systemd.sh now installs a dsh-web.service that runs node "$DSH_BIN" web --no-open with DSH_BIN=~/.local/share/dsh/current/bin/dsh and WorkingDirectory=~/.local/share/dsh, instead of booting a deepseek-harness checkout through tsx. Install @deepseek-ai/dsh under ~/.local/share/dsh/<version> and link current before running the installer. The supervisor unit passes the same DSH_BIN. The working directory holds no .env, which dsh would otherwise load as a project env layer.
• restart-dsh-web.sh no longer requires --repo: a systemd-managed unit needs no checkout, and without systemd the helper launches DSH_BIN. --repo / DSH_REPO remain as the source-launch fallback.

Fixed

• Every path that boots dsh web follows DSH_BIN when it is set (resolveDshLaunch): the dsh_web_restart / dsh_web_dryboot gate, the dsh_web_gc orphan fingerprint, the daemon nohup fallback and the debug-agent dry-boot. They previously always booted the source checkout, so on a packaged host the gate verified code the host does not run. Without DSH_BIN the source launch is unchanged.

What's Changed

• fix(supervisor): run and restart dsh from the packaged runtime by @ddtcorex in #131
• chore(release): 0.14.0 by @ddtcorex in #132

Full Changelog: v0.13.0...v0.14.0

v0.13.0

Removed

• Last-known-good snapshots and crash rollback are gone. The supervisor daemon no longer writes ~/.dsh/.supervisor/lkg/ snapshots and no longer restores files on crash — a broken boot surfaces through the crash report (reports/report-<ts>.md) instead of an automatic restore. Health polling, restart on crash, Telegram notify, session auto-resume and the debug-agent still run as before. The dsh_web_restart gate now always dry-boots since there is no snapshot baseline to compare against.

What's Changed

• feat(supervisor): remove LKG snapshots and crash rollback by @ddtcorex in #129
• chore(release): 0.13.0 by @ddtcorex in #130

Full Changelog: v0.12.1...v0.13.0

v0.12.1

Removed

• The guard no longer lists maestro_read_file and maestro_write_file as file tools. No plugin has registered them since the monolith was retired; the generic read_file, write_file, fs_read and fs_write names stay covered.

What's Changed

• docs(core): fix rules-audit findings by @ddtcorex in #123
• chore: follow the maestro-reviewer rename by @ddtcorex in #124
• test(config): read maestro-reviewer from the meta root by @ddtcorex in #125
• test(core): read the review client from dsh-maestro-review by @ddtcorex in #126
• refactor(guard): drop the retired maestro file tool names by @ddtcorex in #127
• chore(release): 0.12.1 by @ddtcorex in #128

Full Changelog: v0.12.0...v0.12.1

v0.12.0

Removed

• BREAKING: schema v1 guard config is no longer translated. mapLegacyConfig, loadGuardConfigWithMigration, journalConfigMigration and the config-legacy journal note are gone. A persisted domains.guard that still carries gitProtection, publishBlocked, cwdContainment or credentialPaths now has those keys ignored, so protection falls back to the built-in defaults until the schema v2 keys (rules, protectedBranches, protectedPaths) are set. The Settings tab no longer renders a "legacy" tier source.
• BREAKING: the retired ticket store is no longer moved aside. retireLegacyStore and guard/migrate.ts are removed; pending.json is left where it is. The guard still treats legacy-pending.json as a protected path.
• BREAKING: <dsh home>/maestro/settings.json is no longer read. The supervisor consults only <dsh home>/dsh-maestro-config/settings.json.
• BREAKING: the mtime-only ~/.dsh/.supervisor/planned-restart file is no longer honoured. Only the JSON planned-restart.json marker counts. restart-dsh-web.sh now writes that JSON marker (default path planned-restart.json, 180s ttl); a custom DSH_SUPERVISOR_MARKER must point at a JSON marker too.
• BREAKING: SyncService.pull() and SyncService.push() and the PullResult and PushResult types. Use preview({ direction }) then apply({ previewId, direction, confirm: true }).
• BREAKING: the maestro_sync_pull and maestro_sync_push tools and the pull and push sync RPC methods. They were preview-only aliases that never wrote. Agents call maestro_sync_preview then maestro_sync_apply; the Sync tab already uses previewStart and apply.
• The duplicate MaestroMark in the sync UI kit; the sync surface imports the one in config/components/BrandMark.tsx.

Fixed

• Connection row injects webStartup. DSH 0.2.1-alpha.x removed the webRuntime service; a connection entry still injecting it waits for a service that never mounts and dsh web never boots.
• Journal stays silent on the rotation chmod race. Boot-time rotation can rename the live file away between appendFile and the hardening chmod; the entry is already preserved, so only ENOENT is tolerated and any other chmod failure still reports.

What's Changed

• docs(core): document requirements, pinning and the shipped connection patch by @ddtcorex in #117
• chore(core): remove dead code and legacy shims by @ddtcorex in #118
• fix(supervisor): defer RPC registration until connection is active by @ddtcorex in #119
• chore(core): drop unused markerPath parameter by @ddtcorex in #120
• DSH 0.2.1 compat: webStartup inject + journal chmod race tolerance by @ddtcorex in #121
• chore(release): 0.12.0 by @ddtcorex in #122

Full Changelog: v0.11.2...v0.12.0

v0.11.2

Fixed

• RPC rows activate on DSH 0.2.x. The bundle patch declares webServer on the connection entry, so rows that call rpc.handle no longer fail with cannot get property "webServer" without inject and their channels no longer answer 405.

What's Changed

• fix(core): declare webServer on the connection entry by @ddtcorex in #115
• chore(release): 0.11.2 by @ddtcorex in #116

Full Changelog: v0.11.1...v0.11.2

v0.11.1

Fixed

• Settings fields share one box. Fields adopt the host's own ConfigField declaration plus a 44px touch floor, so the dialog no longer shows several geometries. The SSH target buttons stay on a row of their own below the field, the Maestro tab dropdowns keep 13px text, and the SSH label states a whole-pixel line-height so the hairline border cannot land on a half pixel.

What's Changed

• docs(snapshot): say "harness home" instead of the private tunnel prefix by @ddtcorex in #111
• fix(supervisor): register tools and skills in a child fiber by @ddtcorex in #112
• fix(core): adopt the shared settings field box by @ddtcorex in #113
• chore(release): 0.11.1 by @ddtcorex in #114

Full Changelog: v0.11.0...v0.11.1

v0.11.0

Security

• The /dsh-maestro-config channel no longer serves other plugins' secrets.
  list/get/set/unset read and wrote every domain of the shared
  settings store, so any logged-in browser session — including one arriving
  through the PIN-gated public tunnel — could read the GitLab token, the
  webhook secret, the Telegram bot token and the tunnel credentials path in
  clear text, and could write or unset another plugin's keys without that
  plugin's validator. All four endpoints now accept only guard,
  guardBlacklist and supervisor, the domains this package's Settings card
  uses; list no longer reveals foreign domain names. In-process consumers of
  ctx.maestroConfig keep the full service.

Fixed

• The boot resume scan fires at boot, not minutes later (0.10.0 and earlier).
  The window was anchored on Date.now() - withinMs, which drifts minutes into
  the session, so a session interrupted right before a restart was excluded from
  its own recovery scan. The anchor is now the process start time, and
  runAutoResume takes the boundary as an injectable option so the behaviour is
  testable rather than clock-dependent.
• The Maestro Settings card renders again (0.10.0 and earlier). A refactor
  trimmed the tab set and removed the error state its surviving handlers still
  called, so the whole section threw at render time and came up empty.

Changed

• The store cache validates mtime, inode and size. Two copies writing inside
  one filesystem timestamp tick, or an out-of-band edit that restored the mtime,
  served a stale document.
• Lock stealing is safe. A lock carries an owner token and is released only
  by its owner; a stale lock must be older than 30s (longer than the 5s acquire
  timeout) and is stolen with an atomic rename that is rolled back if the moved
  file turns out to be a different inode. The settings directory is created 0700
  and the lock 0600.
• A throwing onChange callback can no longer fail a write that already
  landed, or drop a sibling listener.
• fs.watch has an error listener, so a watcher failure is logged instead
  of crashing the process.
• A failed store write removes its temp file instead of leaving it behind.

Added

• systemd/ and scripts/ ship in the npm tarball. The manifest's files
  list excluded them, so a fresh npm install @ddtcorex/dsh-maestro-core could
  not install the supervisor daemon it publishes — scripts/install-systemd.sh
  was reachable only from a git checkout.

What's Changed

• fix(client): derive the browser loader id from the manifest by @ddtcorex in #103
• fix(config): render the Maestro settings tab again by @ddtcorex in #104
• fix(supervisor): make the boot resume scan fire at boot, not minutes later by @ddtcorex in #106
• docs: record what the absorption actually preserved by @ddtcorex in #105
• docs(core): align README, AGENTS and skill with the code by @ddtcorex in #107
• fix(core): restrict the config RPC to owned domains and harden the store by @ddtcorex in #108
• fix: ship systemd/ and scripts/ in the npm tarball by @ddtcorex in #109
• chore(release): 0.11.0 by @ddtcorex in #110

Full Changelog: v0.10.0...v0.11.0

v0.10.0

Changed

• One package, four host rows. The settings store, the Maestro settings
  card service, the guard and the sync engine now live in this repository as
  src/host/{store,config,guard,sync}, loaded by four rows in
  cordis.patch.yml instead of four separately installed packages. Each
  repository's history is preserved and its changelog is kept under
  docs/history/.
• The store registers no domain validator. store/index.ts is the file
  and the atomic write; the legacy flat-key adapter moved to store/legacy.ts
  and each domain owner registers its own validator through defineDomain.
  The guard validators ship with the guard module; the review and notifier
  validators are declared by those plugins.
• onChange fires for writes made by another copy. An embedded store copy
  watches the settings file while it has listeners, so a module that embeds the
  store keeps hearing Settings writes. The watcher starts with the first
  listener and stops with the last.
• One client bundle. src/client/index.tsx composes the auto-reload, the
  settings card and the Sync card, each guarded so one failure cannot cost the
  others their registration.

Added

• scripts/vendor-store.mjs generates a consumer's embedded store copy with a
  body hash, and verifyVendored rejects a hand-edited or stale copy.
• exports["./store"] publishes the store as a subpath.

Removed

• The @ddtcorex/dsh-maestro-config-lib dependency, the sibling checkout in
  pnpm-workspace.yaml and the sibling-repos entry in CI.

Changed

• The package is published as @ddtcorex/dsh-maestro-core — renamed from
  @ddtcorex/dsh-maestro-supervisor in one sweep, because a profile carrying
  half the rename resolves nothing. The repository and the workspace directory
  were renamed with it. The three subpath row name values in
  cordis.patch.yml, the client loader id, the source-kind label the
  supervisor persists into session logs, and every runnable path in README.md
  and AGENTS.md moved in the same commit. The RPC channel names
  (/dsh-maestro-supervisor-resume, -session-health, -reload) and the
  dsh-web-supervisor binary name are wire protocol and did not change, so
  the earlier sections of this file keep the old spelling on purpose.

What's Changed

• fix(supervisor): catch the fire-and-forget start by @ddtcorex in #93
• fix(supervisor): drop the ignored authority argument from rpc.handle by @ddtcorex in #94
• fix(supervisor): discover the session log generation instead of listing it by @ddtcorex in #95
• docs(supervisor): point skill references at dsh-safe-restart by @ddtcorex in #96
• fix(supervisor): parse CRLF skill frontmatter and attribute injected messages by @ddtcorex in #97
• chore: drop the dead dsh-client-runtime client inject edge by @ddtcorex in #98
• docs(skill): add a bundle-resolution arm to the restart preflight by @ddtcorex in #99
• fix(supervisor): report the real restart budget and daemon staleness by @ddtcorex in #100
• feat(core): absorb config-lib, config, guard and sync into one package by @ddtcorex in #101
• build(core): publish under @ddtcorex/dsh-maestro-core by @ddtcorex in #102

Full Changelog: v0.9.0...v0.10.0

v0.9.0

Fixed

• Supervisor fallback reads the shared settings store — the four
  domains.supervisor.* reads pointed at ~/.dsh/maestro/settings.json,
  which never exists on a real install where the config lib writes
  ~/.dsh/dsh-maestro-config/settings.json; they now resolve the shared
  path honouring DSH_HOME (#91).
• A fresh install defers to the store — the hardcoded
  autoResumeEnabled: true left cordis.patch.yml, so the UI toggle is no
  longer silently overridden (#91).

Added

• isAutoResumePinned() and a status RPC endpoint, so a settings UI can
  render an honest locked/disabled toggle when an install-supplied Cordis
  config pins the value (#91).

Changed

• Follow @ddtcorex/dsh-maestro-config-lib to ^0.3.0.

What's Changed

• docs: declare package license, repository and node engine by @ddtcorex in #90
• fix(supervisor): correct settings path, remove pinned autoResume default, expose pin status by @ddtcorex in #91
• chore: release v0.9.0 by @ddtcorex in #92

Full Changelog: v0.8.4...v0.9.0

v0.8.4

Fixed

• A resumed session keeps its agent preset — agents.resume() composes an
  agent's scoped world only through ResumeAgentOptions.setup, and the resume
  path passed none, so every auto-resumed session was published joined to no
  preset: its tool view collapsed to the deployment-global layer and every
  preset tool (bash, read, write, edit, glob, grep, subagent,
  skill, todo_write, workflow, cordis_*, goal/plan/jobs) answered
  Error: unknown tool "bash" for the rest of that agent's life. Measured on
  the live incident: the resume header carried 76 tools instead of 110, and
  agentPresets.composedPreset(agent.ctx) was undefined where a healthy
  session answers cordis. Activation now delegates to
  sessionController.resolveAgent — the entry point the Web API itself uses,
  which composes the preset, de-duplicates concurrent resumes and returns the
  live agent when the browser already re-opened the session — and falls back to
  agents.resume with setup: agentPresets.mount(agentCtx, presetId), reading
  the preset the session records (projection, then persisted header, then raw
  log header). A session that records no preset still resumes bare, which is
  correct for agents created outside the Web app.
• A preset that cannot be composed no longer degrades into a preset-less
  resume — the setup failure is wrapped in PresetComposeError, that one
  session is skipped with a resume-failed journal entry and an operator
  notification, and the remaining sessions still resume.
• An unreadable tool registry is no longer reported as healthy — the probe
  returned {missing: [], visible: 0} both when nothing was missing and when it
  could not read the registry at all, which is how the loss above stayed
  invisible (maestro_resume_tool_health printed missing=none visible=0).
  Probes now carry registry: reachable | unreachable, and the health output
  renders registry=… composed=… repaired=… beside the tool view.
• Removed the pre-delivery bash readiness wait — it polled the global
  registry for bash, which never holds a preset tool (measured), so it could
  only exhaust its budget and continue anyway; composition is now verified
  against the preset roster instead.

Added

• maestro_repair_session_preset — re-links one live session's agent to the
  preset that session records (agentPresets.recompose, which for an agent with
  no binding is its first bind). It repairs an agent that lost its preset before
  this release without waiting for a restart, and reports repaired,
  already-composed, no-preset-recorded, no-live-agent or
  recompose-failed.
• Automatic repair on resume (resumeAutoRepair, default true) — after a
  resumed agent is activated the roster is asked
  (composedPreset(agent.ctx)) and a preset-less agent is re-linked before the
  recovery prompt is delivered, so the turn the prompt starts already carries
  the preset's tools. The check is skipped when it cannot be made (no roster
  service, or an agent without its own context), and the config key is the kill
  switch for this live re-link (setup/tools/change); with it off the loss is
  still notified and parked.

What's Changed

• docs: correct the lib/ build-output claim by @ddtcorex in #80
• docs: drop hard-coded test counts by @ddtcorex in #81
• fix(supervisor): let the resume-tool-health tool report a null probe by @ddtcorex in #82
• fix(supervisor): detect an interrupted long turn by the log mtime by @ddtcorex in #83
• fix(supervisor): retry a resume blocked by another owner's write handle by @ddtcorex in #84
• fix(supervisor): continue a session the restarted host already owns by @ddtcorex in #85
• fix(supervisor): gate both resume paths on bash being registered by @ddtcorex in #86
• chore: release v0.8.3 by @ddtcorex in #87
• fix(supervisor): resume a session under its agent preset by @ddtcorex in #88
• chore: release v0.8.4 by @ddtcorex in #89

Full Changelog: v0.8.2...v0.8.4