argent 2026.7.7

Highlights

• The AI workforce is real. Worker Runtime v2 lands end-to-end: blank-slate
  ephemeral worker sessions with compiled role profiles (#452), a full lease
  lifecycle — claim/heartbeat/halt with true runNow semantics (#453), a
  work_report completion contract that lets a conductor demo COMPLETE on a
  local gemma model (#450), and a grading store with scorecards and a
  quantified promotion gate so a worker earns autonomy with evidence, not
  vibes (#463). The worker wizard mounts straight from the Workforce Board
  (#440), and local-model workers are viable thanks to native tool-call
  support with blank-slate prompts (#444, #447).
• One source of truth again. ArgentOS-Business merges back into core —
  business logic, dashboard surfaces, and runtime live in a single repo and
  ship together (#435).
• Approvals you can trust. Deny now denies, timeouts actually fire, failed
  sends fail loud, and toolsAllow is enforced end-to-end (#462). Cron
  approval buttons render and resolve (#457, #459), with a Telegram text-reply
  fallback when buttons aren't available, and the approval text-command path
  is refactored to be provably testable (#458).
• Rust daemons graduate from experiment to credible. Fail-closed bearer
  token auth on both shadow daemons — no token, no start, no route (#470);
  argentd answers canary receipt methods over WebSocket (#468); and the
  authority delegation seam is designed and locked after adversarial review
  (#469).
• An idle gateway now burns zero inference. Deterministic salience gates
  keep the kernel (#451), heartbeat, contemplation, and closet runners (#454)
  silent unless something actually changed, and the personal-skill block is
  off the turn critical path (#436) — idle boxes stay cool and cheap.
• Install and update paths harden. Unattended installs never block on the
  master-key prompt (#434 closes #429), argent update runs native rebuilds
  under the gateway's own pinned Node and verifies the ABI before restart
  (#448 closes #416), and a fresh PostgreSQL self-bootstraps its full schema
  and agent rows at first init (#439).
• Security and CI baseline. All 11 critical Dependabot alerts cleared
  (vitest, protobufjs, shell-quote, baileys) (#456), and the unit suite is
  green with a new blocking CI unit-test job so it stays that way (#461).

Changes

• feat: merge ArgentOS-Business back into core — one source of truth (distillation merge) (#435)
• feat(worker): Worker Runtime v2 P2 — blank-slate ephemeral sessions + compiled role profiles (#452)
• feat(worker): Worker Runtime v2 P3 — lease lifecycle (claim/heartbeat/halt) + full runNow semantics (#453)
• feat(worker): work_report completion contract — conductor demo COMPLETES on gemma (v2 D5) (#450)
• feat(workforce): WR2 P5 — grading store, scorecards, quantified promotion gate (D10) (#463)
• feat(dashboard): mount the worker wizard from the Workforce Board (#440)
• feat(demo): conductor demo kit + local-provider runtime auth + bootstrap markers (#441)
• feat(argentd): canary receipt methods over ws — status + generateLocalProof (#468)
• perf(kernel): deterministic salience gate — idle gateway burns zero kernel inference (LIMBIC law 3) (#451)
• perf(runners): idle-salience gates for heartbeat + contemplation + closet quick wins (LIMBIC law 3) (#454)
• perf(#405): take the personal-skill block off the turn critical path (writes + reads) (#436)
• refactor(workflows): testable runApprovalTextCommand (proves #2 resolve+resume) (#458)
• test: green the unit suite + add blocking CI unit-test job (#461)
• chore(deps): clear all 11 critical Dependabot alerts (vitest, protobufjs, shell-quote, baileys) (#456)
• docs(rust): authority delegation seam design — locked after adversarial review (#469)

Fixes

• fix(rust): fail-closed token auth on both shadow daemons (delegation seam P1) (#470)
• fix(workflows): approval integrity — deny denies, timeouts fire, sends fail loud, toolsAllow enforced (#462)
• fix(workflows): cron approval buttons + Telegram text-reply fallback (#457)
• fix(telegram): wire sendPayload into extension outbound (approval buttons render) (#459)
• fix(workforce): D9 stubbed calls no longer trip the simulate violation wire (#467)
• fix(worker): textual pseudo-tool-call detection + format correction (#442) (#444)
• fix(worker): native tool calls + blank-slate prompt — local-model workers viable (#442, #407 worker lane, #445 P0) (#447)
• fix(operator): builder tool-name validator, Discord crash guards, pnpm resolver + Rust shadow evidence (#465)
• fix(operator): calendar failure cooldown + workflows.resume replay guard (#466)
• fix(dashboard-api): prefer gateway.auth.token over stale env token (DocPanel 401) (#460)
• fix(compact): resolve the session's agent; sweep deprecated CLI profiles from all agent stores (#455)
• fix(installer): never block unattended installs on the master-key prompt (closes #429) (#434)
• fix(update): run native rebuilds under the gateway's pinned Node + verify ABI (#416) (#448)
• fix(pg): fresh databases self-bootstrap full schema + agent rows at init (#439)
• fix(memu): keep singleton identity slots canonical under noisy queries (#437)
• fix(demo): conductor SOP completes its own job task (tickets read-only) (#443)

argent 2026.5.6.7

Fixes

• Public curl | bash install no longer dies on the pinned pnpm. On a Mac
  with no compatible system Node, the installer provisions its own private Node
  runtime — but the pinned-pnpm activation invoked corepack/npm (both
  #!/usr/bin/env node scripts) without that Node on PATH, so they failed
  with env: node: No such file or directory and the install fell back to a
  mismatched global pnpm and aborted. The activation now puts the runtime Node
  on PATH, disables corepack's non-interactive download prompt
  (COREPACK_ENABLE_DOWNLOAD_PROMPT=0), and surfaces the corepack error instead
  of swallowing it.
• Dashboard (port 8080) builds again. The dashboard's vite build imports
  shared UI source that depends on @noble/ed25519; under the installer's
  --ignore-workspace install it was unresolved and the build failed. Declared
  @noble/ed25519 so it resolves for the dashboard bundle.

argent 2026.5.6.6

Highlights

• Argent's kernel learns to notice itself. New kernel-fitness substrate
  tracks engagement, surfaces "system unavailable" on shutdown, detects
  artifact-open via atime, and writes a stall composite ledger — the
  foundation for Argent to know when it's actually being used vs idle.
  (#388, #390, #391, #392, #393, #394, #395)
• Installer Mac path matures. install.sh now auto-installs Ollama on
  macOS (#397), the wizard writes kernel.localModel after a local-runtime
  pick (#398), and argent update rebuilds native modules after dependency
  installs (#396) so the post-update ABI-mismatch class of error stops
  showing up.
• Auth recovery hardened across both halves. Dashboard surfaces a
  recoverable banner instead of silently 401-ing when localStorage hits
  quota or the api-server rejects a rotated token (#417). Gateway-side
  tools (doc_panel, workflow-runner, workflow-execution-service)
  fall back to gateway.auth.token from argent.json so they stop
  failing with "Invalid token" after a token rotation (#422).
• Memory recall stops burying exact answers. New dominance guardrail
  in the MemU LLM reranker prevents the LLM from demoting a clearly
  dominant pre-rerank top in favor of broad identity facts that only
  sound related (#420 closes #418).
• Performance + correctness wave. AppForge bases.list cache stops
  the per-web-turn cascade (#409); sessions skillsSnapshot deduplication
  • workflow auto-prune cut sessions.json from ~98 MB to <16 MB on real
    installs (#412); argent update clean-check is now honest about
    regenerated artifacts and surfaces dirty paths (#414); tinyfish_search
    becomes a first-class tool (#415); Ollama pings now use IPv4 to avoid
    the localhost IPv6-first resolution failure (#385); the wizard's
    Ollama + MLX model catalog refreshed to May 2026 (#387); a new
    read-only System Health Panel M1 lands in the dashboard (#399).

Changes

• feat(kernel): kernel-fitness Phase 1 — extract inner-loop prompt to scaffold + bump tickMs default (#388)
• feat(kernel): kernel-fitness Phase 2 — fitness writer + stall composite ledger (#390)
• feat(kernel): kernel-fitness Phase 3a — engagement tracker backend + fitness-writer integration (#391)
• feat(kernel): kernel-fitness Phase 3b.1 — notifier writes surface_emitted to engagement-ledger (#392)
• feat(kernel): kernel-fitness Phase 3b.2 — system_unavailable on shutdown + startup recovery (#393)
• feat(kernel): kernel-fitness Phase 3b.3 — artifact-open atime detection (#394)
• feat(kernel): kernel-fitness Phase 3b.4 — engagement.record gateway handler (#395)
• feat(installer): Phase B — install.sh auto-installs Ollama on Mac (#397)
• feat(installer): wizard writes kernel.localModel after local runtime pick (Phase C) (#398)
• feat(dashboard): System Health Panel M1 (read-only diagnostics) (#399)
• feat(tools): first-class tinyfish_search tool (#415)
• feat(wizard): default Ollama+MLX models to current May 2026 catalog (#387)
• perf(gateway): cache appforge.bases.list 60s + invalidate on writes (#409)
• perf(sessions): dedup skillsSnapshot to sidecar + auto-prune stale workflows (#412)
• docs: restore release-core + release-swift-mac runbooks (#419)

Fixes

• fix(memory): reranker guardrail — prevent LLM from burying dominant pre-rerank top (closes #418, #420)
• fix(gateway): dashboardApiHeaders falls back to gateway.auth.token from argent.json (#422)
• fix(dashboard): surface auth-lost banner instead of silent 401 (closes #400, #417)
• fix(dashboard): fetchJsonWithRetry attaches auth header via fetchLocalApi (closes #403, #411)
• fix(update): honest clean-check + ignore regenerated artifacts + show dirty paths (closes #413, #414)
• fix(update): rebuild native modules after deps install (closes #386, #396)
• fix(gateway): pingOllama uses IPv4 literal to avoid IPv6-first localhost resolution (#385)

v2026.5.6.3 — TinyFish full stack + 1Password + pi-bridge cluster

2026.5.6.3

Third micro release on top of v2026.5.6. 17 commits / 11 issues closed,
combining the full TinyFish web toolkit (search + fetch + browser + agent),
1Password Service Account as a first-class secrets backend, and the
pi-bridge cluster that unblocks the long-pending pi-ai 0.73.1+ bump (#182).
No schema changes, no breaking API changes. Safe drop-in upgrade from
v2026.5.6.2.

Highlights

• TinyFish becomes the recommended default web stack. Free Search +
  Fetch ship as the first-class default for every account (no API-key
  wrangling, no credits), and the paid Browser API + Agent API now ship
  as first-class agent tools alongside them. Operators get a working
  web stack in 30 seconds; agents can drive a real browser
  (Playwright/CDP) or describe goals in natural language and have
  TinyFish navigate / click / fill on their behalf.
• 1Password Service Accounts as a first-class secrets backend.
  Per-key op://Vault/Item/field references on any entry in
  service-keys are now detected at resolve time and resolved through
  the op CLI. Existing literal values continue to work unchanged —
  fully additive and opt-in per key. Tokens are never logged, output
  is always masked, and 5-minute in-memory caching keeps the resolver
  cheap. argent secrets backend 1password setup, doctor, and
  test <variable> ship with the CLI; argent doctor reports
  1Password health only when refs exist.
• Pi-bridge cluster closes — pi-ai 0.73.1+ bump (#182) is now
  unblocked. Six bridge PRs land in this release (#321, #322, #323,
  #324, #327 plus the foundation cluster #299/#308/#309/#310). Argent's
  runtime now consumes pi-coding-agent's AgentSession, Agent.state,
  AgentMessage, BashExecutionMessage, supportsXhigh, Transport,
  AuthStorage, ModelRegistry, SessionCompactionResult, and the
  typebox Tool/Context types through structural bridge interfaces
  in pi-bridge/. The next pi-ai bump becomes a routine version-pin
  rather than a multi-day migration.
• SetupWizard cards derived from the provider seed. The dashboard's
  first-run LLM provider cards (anthropic, openai, minimax, zai, local)
  are now generated from src/agents/provider-registry-seed.ts at
  build time, eliminating the drift between the runtime registry and
  the onboarding wizard that #295 catalogued.

TinyFish web toolkit

• TinyFish first-class default — web_search + web_fetch (PR
  #317):
  tinyfish joins SEARCH_PROVIDERS; resolveSearchProvider picks
  TinyFish whenever a TINYFISH_API_KEY is resolvable (env, dashboard
  service-keys, or tools.web.search.tinyfish.apiKey). web_fetch
  gains an opt-in backend: "tinyfish" param that posts JS-heavy or
  anti-bot pages through api.fetch.tinyfish.ai. CLI wizard surfaces
  TinyFish first in promptWebToolsConfig; dashboard onboarding
  reorders SEARCH_PROVIDER_CARDS with TinyFish at the top labelled
  "Free Search & Fetch (Recommended)". Brave + Perplexity remain
  fully supported. Strict-Zod gap closed for firecrawl, readability,
  and tinyfish sub-objects so valid argent.json configs are no
  longer rejected. 30 tests in web-search.test.ts cover every
  smart-default branch, resolver units, and 401 auth-failure transport.
• TinyFish Browser API — remote Playwright/CDP session (PR
  #325, closes
  #320):
  New tinyfish_browser_open / tinyfish_browser_close agent tools.
  open posts to api.browser.tinyfish.ai and returns
  {session_id, cdp_url, base_url} — the agent (or a higher-level
  tool) connects via chromium.connect_over_cdp(cdp_url). Argent
  itself does not wrap Playwright. close is acknowledgement-only
  (sessions auto-terminate after 1 hour of inactivity; the API has no
  explicit delete endpoint). Paid-tier walls surface as
  tinyfish_browser_paid_tier_required with an upgrade message; 401
  is mapped separately so the agent doesn't suggest an upgrade on
  auth failures. 21 tests cover registration, schema, auth, transport,
  and error mapping.
• TinyFish Agent API — natural-language web automation (PR
  #326, closes
  #319):
  New tinyfish_agent tool. Operators describe a goal in natural
  language plus a starting URL; TinyFish navigates / clicks / fills
  on real sites and returns structured results. Optional
  max_steps, browser_profile (lite/stealth), screenshots,
  snapshots, recording, webhook_url, timeout_seconds.
  Hard caps: 500 steps, 600s timeout. New config surface
  tools.web.agent.*. Opt-in (defaults enabled: false) because
  Agent is a paid tier. Error mapping for INSUFFICIENT_CREDITS /
  paid-feature wall, RATE_LIMIT_EXCEEDED, auth failures, and
  client-side timeout. 35 tests covering config, auth, URL guards,
  registration, transport, and error mapping.

1Password Service Accounts (first-class secrets backend)

• op:// references resolved at service-key resolution time (PR
  #318):
  Per-key op://Vault/Item/field references in any service-keys entry
  are detected at resolve time and resolved through the op CLI.
  New src/infra/onepassword-resolver.ts exposes isOnePasswordRef,
  parseOnePasswordRef, maskRef, and resolveOnePasswordRef with
  a 5-minute in-memory cache (TTL configurable per call; ttl=0
  disables). OP_SERVICE_ACCOUNT_TOKEN is read from process.env
  or from argent's encrypted secret store and is never logged or
  printed — all op output passes through redactToken first, and
  the subprocess env is whitelist-built so unrelated env doesn't leak
  into op. Graceful fallback when the op CLI is absent
  (op_cli_missing). New CLI: argent secrets backend 1password setup [--token ops_...] [--migrate-existing], doctor [--sample op://...], and test <variable>. argent doctor reports a
  1Password block only when refs are present, with clear remediation
  steps; never fails the parent doctor run. Dashboard exposes
  GET /api/settings/service-keys/1password/health with the same
  probe data (UI toggle is a follow-up). 20 tests cover parse, mask,
  redact, token resolution priority, cache TTL boundary (hit / miss
  / ttl=0), op-missing fallback, ref detection, and env-fallback
  on resolver failure. Ships flavor A (per-key refs) from the
  spec; flavor B (global secrets.backend = "1password" flag) is a
  follow-up.

Pi-bridge cluster (unblocks #182)

The pi-coding-agent 0.70.2+ surface added ~78 private members,
removed replaceMessages, privatized AuthStorage / ModelRegistry
constructors, switched to typebox@1.x, and reshaped
BashExecutionMessage and supportsXhigh. PR #275 / #276 attempted
forward-direction unification (argent's AgentSession impl wrapping
pi's class) and correctly stopped — class extension can't satisfy
private members it doesn't own. This release ships the structural
bridge instead: argent's runtime consumes pi's surface through
interfaces in pi-bridge/ that capture only the public shape
argent reads. Future pi bumps absorb drift at one chokepoint per type.

• Foundation + AuthStorage / ModelRegistry / SessionCompactionResult /
  typebox Tool+Context (PRs
  #299 /
  #308 /
  #309 /
  #310, partial
  #286;
  closes #300,
  #303,
  #305):
  Establishes src/argent-agent/pi-bridge/ as the single chokepoint
  for argent ↔ pi-coding-agent type identity. Migrates the remaining
  AuthStorage and ModelRegistry call sites off direct pi imports;
  adds a SessionCompactionResult mapper; bridges the typebox
  Tool/Context surface so 0.34 → 1.x drift is absorbed in one
  place.
• supportsXhigh + Transport cleanup (PR
  #321, closes
  #306):
  Drift-absorbing supportsXhigh(model) wrapper in pi-bridge (local
  re-implementation, not a re-export — pi 0.73+ removed the named
  export and replaced it with
  getSupportedThinkingLevels().includes("xhigh")). Transport type
  alias forwards pi's "sse" | "websocket" | "auto" union; audited
  src/ for the removed "websocket-cached" variant and confirmed
  zero occurrences. The two #182-catalogued sites (agent-core/ai.ts
  re-export and pi-embedded-runner/run/attempt.ts
resolveOpenAICodexTransport) migrated through the bridge.
• AgentSessionLike structural bridge (PR
  #323, closes
  #301):
  pi-bridge/agent-session.ts exposes AgentSessionLike and
  AgentSessionAgentLike — interfaces capturing only the public
  surface pi-embedded-runner actually reads. Both argent's
  AgentSession interface and pi's class satisfy the read subset;
  argent's interface additionally satisfies agent.setSystemPrompt
  (the runtime source is always argent — agent-core/coding.ts
  rebinds createAgentSession → createArgentAgentSession).
  attempt.ts:1130 TS2352 cast and system-prompt.ts:94 TS2339
  setSystemPrompt-on-Agent error retire. Baseline: 227 → 225.
• Agent.replaceMessages migration (PR
  [#324](https://...

v2026.5.6.2 — 18 issue fixes + AOS subctl connector

2026.5.6.2

Second micro release on top of v2026.5.6. 18 issues closed, all bug
fixes, install/update hardening, and ergonomics — no schema changes, no
breaking API changes. Safe drop-in upgrade.

Highlights

• Auto-close-on-dev-merge pipeline is live. Issues referenced with
  closes/fixes/resolves in commits that land on dev now auto-close,
  even though dev isn't the default branch. Catches the long-standing
  pain of having to manually close issues after PR merges
  (#180).
• AOS subctl connector lands. Bridges ArgentOS workflows and the AOS
  worker harness to subctl
  — multi-account AI subscription control plane. Mirrors the
  aos-telegram connector pattern. Read commands work in readonly
  mode; write commands gated behind mode=write (PR #192).
• Install + update get serious hardening. Five compound install
  fixes: legacy pnpm symlink sweep, postgres@17 restart-not-start +
  port-conflict diagnosis with ARGENT_PG_REPURPOSE_EXISTING opt-in,
  pg_dump snapshot before takeover, dashboard node_modules/react
  postinstall sweep, and dashboard provider-catalog populated from
  seed at build time.
• Type-identity work continues. AgentMessage/StreamFn between
  argent's local pi-types and @mariozechner/pi-agent-core are
  unified — tsc baseline shrinks 267 → 249. The band-aid in #256
  stays as defensive (orthogonal to drift).
• Z.AI thinking-only responses surface correctly. Two-layer fix:
  fallback retry to glm-4.7 when glm-5* returns empty, plus
  adapter-layer reasoning_content extraction in
  openai-completions so the wasted round-trip goes away over time.

Bug fixes — Install / Update / Gateway

• Legacy pnpm symlink sweep
  (#168):
  argent update now sweeps 7,892+ legacy pnpm symlinks that pointed
  at /Users/sem/argentos/ (the pre-canonical install path) and
  rewrites them to the canonical install root. argent doctor --repair
  invokes the same sweep so existing affected installs self-heal in
  place. Idempotent.
• Postgres@17 already-running handling
  (#109):
  Hosted installer detects postgresql@17 already running, uses
  brew services restart (not start) so config rewrites actually
  take effect, polls psql for 20s after restart, and reports a clear
  port-conflict diagnosis with the listener identity if 5433 is held
  by something else. Data-loss guard: aborts with a warning if the
  existing postgres has user data unless
  ARGENT_PG_REPURPOSE_EXISTING=1 is set.
• Postgres pre-takeover snapshot
  (#278):
  When ARGENT_PG_REPURPOSE_EXISTING=1 triggers an actual takeover,
  pg_dumpall snapshots the existing data to
  ~/.argentos/backups/postgres-pre-takeover-<timestamp>.sql before
  the port rewrite. Auth-required clusters fail fast (-w) with a
  pgpass/PGPASSWORD hint. Any dump failure aborts the takeover and
  leaves the port untouched.
• Dashboard react symlink postinstall sweep
  (#264):
  Pre-existing dashboard/node_modules/react symlinks that pointed at
  removed worktrees (a pnpm hoist artifact) are auto-rewritten by a
  postinstall script. Fresh worktrees pass React tests on first
  install. Silent on healthy installs.
• Dashboard provider-catalog populated from seed
  (#268):
  dashboard/provider-catalog/index.cjs is now generated from
  provider-registry-seed.ts at build time. Previously shipped with
  providers: {}, causing the Models page to render blank on
  headless launches before the user's ~/.argentos/provider-registry.json
  existed. 12 providers ship in the fallback.

Bug fixes — Providers / Models / Auth

• Provider registry seed adds zai
  (#269):
  zai was offered in CLI + dashboard onboarding with 8 GLM models in
  models-db.ts, but provider-registry-seed.ts had no entry — so
  ensureArgentModelsJson never wrote a zai block. Z.AI worked only
  via pi-coding-agent's anthropic-compat fallback. Now seeded
  directly. SEED_VERSION bumped to 9 so existing installs re-seed.
• Chutes reachable from grouped wizard
  (#267):
  chutes was in the flat buildAuthChoiceOptions list but missing
  from AUTH_CHOICE_GROUP_DEFS, so the grouped wizard couldn't reach
  it. Added as a standalone group modeled on qwen.
• zai/glm-5-turbo empty response fallback
  (#254):
  Generalized the empty-response fallback retry to cover
  zai/zai-coding providers. When glm-5-turbo defaults to
  thinking-on and emits reasoning-only content (a known Z.AI quirk)
  via the openai-completions path, the runtime now retries on
  glm-4.7 rather than surfacing a blank assistant message.
• openai-completions surfaces reasoning_content
  (#280):
  Adapter-layer fix paired with the retry-fallback in #254. When
  content is empty AND toolCalls is empty AND reasoning_content
  is non-empty, the openai-completions adapter now uses
  reasoning_content as visible text. Provider-agnostic conditional
  gate — OpenAI-proper (always emits content) and tool-call responses
  are unaffected.

Bug fixes — Channels / Plugins

• Agent-tag transform extends to all text channels
  (#202):
  The MOOD/TTS tag transform added in #198 for Telegram + Discord now
  also strips/transforms tags for Signal, Slack, iMessage, WhatsApp,
  and MS Teams. Channels not covered in core are explicitly
  documented in the PR body.
• Mood→emoji map via plugin config
  (#203):
  channels.defaults.agentTags.moodEmojiMap is now a typed,
  zod-validated plugin-config field. Plugins can override individual
  mood entries or supply a full replacement map for deployment
  branding. Empty-string values suppress a default mood.
  Case-insensitive keys. Backward-compatible — default map kept for
  deployments without overrides.

Bug fixes — UI / Build / CI

• fetchLocalApi migration completes
  (#173):
  All 23 grandfathered dashboard files migrated from direct fetch()
  to fetchLocalApi(). INV-2 baseline shrunk to empty — the
  invariant now enforces uniformly. ChatPanel's external-URL fetches
  (lines 95/136 for blob downloads) correctly preserved.
• pi-types unify with pi-agent-core
  (#257):
  AgentMessage/StreamFn/CustomAgentMessages now forwarded
  directly from @mariozechner/pi-agent-core so type identity
  unifies across the runtime. Three latent type-soundness fixes
  surfaced and corrected:
  Extract<AgentMessage, { role: "user" }> after a role filter in
  create-agent-session.ts; null narrowing in estimateTokens;
  onPayload(payload, model) contract in anthropic-payload-log.ts
  (pi's contract was always 2-arg). Baseline 267 → 249. Type-only —
  no runtime behavior change.
• Oxlint shadow-warnings cleared
  (#204):
  Cleared the 2 oxlint shadow errors in
  src/infra/outbound/deliver.ts. 3-line diff with no semantic change.
• Oxlint per-rule sweep begins
  (#285):
  First per-rule safe --fix (no-useless-escape, 12 files, +20/-20).
  Methodology fix documented: pnpm exec oxlint -A all -D <rule> --fix
  to isolate per-rule changes (the naive -D <rule> flag ADDS to
  enabled categories rather than isolating). eslint(curly) found
  unsafe in this codebase (defeats TS narrowing on single-statement
  guards) — documented in the closed PR #288.

CI / Workflow infrastructure

• Auto-close on dev-merge (#180):
  New .github/workflows/auto-close-on-dev-merge.yml. On every push
  to dev, scans each commit's message for verb-prefixed issue
  references (closes/fixes/resolves #N, including multi-issue
  groups), and auto-closes them with a back-linked comment. Verb
  required, so trailing (#PR-N) squash artifacts don't trigger
  false closures. Self-validating cutover: PR #266's own merge
  closed #180.

New features

• AOS subctl connector (PR
  #192):
  New tools/aos/aos-subctl/ module mirroring aos-telegram. Bridges
  ArgentOS workflows + AOS worker harness to subctl's HTTP API +
  notify shell. Read commands: capabilities, health, doctor,
  state.get, orchestration.list/status, notify.inbox. Write commands
  (gated): orchestration.spawn/msg/kill, notify.send,
  notify.inbox_ack. Configurable via SUBCTL_API and SUBCTL_BIN
  env vars. No service keys required (subctl is localhost-only).

Known carry-overs

• #276
  AgentSession identity unify — deferred to #182 (broader pi-ai bump);
  the structural fix touches the wider StreamFn/Agent/AgentSession
  cluster simultaneously, not a one-off.
• #182
  pi-ai 0.73.1+ bump — meta-tracker; intentionally not attempted as a
  single-PR task because the current pi-coding-agent API has 78 new
  private members on AgentSession, removed replaceMessages,
  privatized AuthStorage/ModelRegistry constructors, and switched
  to typebox@1.x. Routine bumps ...

v2026.5.6.1 — 29 issue fixes (micro release)

2026.5.6.1

A micro release on top of v2026.5.6. 29 issues closed, all bug fixes and
ergonomics — no schema changes, no breaking API changes. Safe drop-in upgrade.

Highlights

• argent update is fully unblocked. The tsc --noEmit pre-bundle gate
  introduced in v2026.5.6 was rejecting updates past dev.36 because its
  identity comparison didn't account for tsc's environment-dependent error
  message truncation. The gate now uses a path-length-independent fingerprint,
  so updates run from any checkout depth.
• Install + update hygiene pass. Eight issues that made fresh installs and
  in-place updates fragile are fixed: missing LaunchAgents, port-collision
  during gateway restart, legacy env vars in the gateway plist, residual
  admin surfaces in public Core, doctor inconsistency during transitions,
  Google Workspace preflight resolution, gateway-token rotation on update
  (now preserved), and localStorage quota issues from cron-jobs cache.
• Local-model UX truth-telling. LM Studio dropdown now distinguishes
  loaded vs catalog-only models (previously listed everything as if
  usable). New "LM Studio Models" section on the Models page parallels
  Ollama. Routing-suggestion engine filters to currently-loaded models
  only. agent_status distinguishes "not-configured" from
  "configured-but-down" for Ollama.
• Auth that self-heals. auth-profiles.json now repairs itself when the
  order array references a missing profile. --set-default for
  plugin-backed providers (zai, openrouter, etc.) now actually sticks via
  a recommendedModel manifest field instead of being silently dropped.
• AppForge progress. New rating cell field type (star/heart/thumb/flame
  glyphs) and click-to-sort grid column headers with visual indicators.

Bug fixes — Update / Install / Gateway

• argent update LaunchAgent install
  (#175):
  argent update now installs missing ai.argent.dashboard-ui and
  ai.argent.dashboard-api LaunchAgents for new install paths instead of
  leaving them out.
• argent update gateway-restart cycle
  (#155):
  Wait for port 18789 to release before restarting the gateway, and let
  argent doctor detect transition-state rather than falsely reporting
  "Gateway not running" mid-cycle.
• argent update gateway-token preservation
  (#167): Stop
  rotating gateway.auth.token on every argent update. Token rotation
  is now gated behind an explicit user action — open sessions don't get
  silently invalidated by routine updates.
• Legacy plist env vars purged
  (#169):
  ai.argent.gateway.plist no longer ships with ARGENT_GIT_DIR or
  ARGENTOS_GIT_DIR pointing at the legacy /Users/sem/argentos
  location. New installs get a clean plist.
• argent doctor plist binary validation
  (#172):
  argent doctor now validates ai.argent.*.plist BinaryPath against
  the canonical install location and flags drift.
• Google Workspace connector setup
  (#128):
  Resolve preflight_gws.py from the package root in the installed CLI
  rather than the source tree — setup no longer 404s on installed
  Argent.
• Public Core admin-surfaces hidden
  (#105):
  Public Core dashboard no longer exposes the residual internal/admin
  surfaces (broken routes that previously rendered a confusing 404 in
  the sidebar).
• Dashboard cron-jobs cache quota
  (#157):
  Cap the cron-jobs localStorage cache to prevent QuotaExceededError
  on accounts with long heartbeat history.

Bug fixes — Models / Providers / Auth

• --set-default for plugin-backed providers
  (#190):
  argent models auth login --set-default --provider <zai|openrouter|...>
  now actually sets the default model. Previously honored only for
  openai-codex; other providers silently dropped the flag.
• auth-profiles.json self-heal
  (#193):
  argent agent no longer throws No API key found for provider 'openai-codex' when auth-profiles.json has a valid profile but an
  order array pointing at a missing alias. The store now self-repairs
  on load.
• LM Studio loaded vs catalog
  (#220):
  The LM Studio model dropdown now lists currently-loaded models, not
  the full catalog. Catalog entries no longer mislead users into
  selecting a model that isn't actually loaded in LM Studio.
• LM Studio Models page parity
  (#219):
  Models page now has an "LM Studio Models" section parallel to the
  "Ollama Models" section instead of asymmetric coverage.
• Routing suggestion filter
  (#218):
  The routing-suggestion engine no longer recommends unloaded LM Studio
  models or Ollama models when Ollama is down.
• agent_status reflects configuration
  (#214):
  agent_status now distinguishes "not-configured" from
  "configured-but-down" for Ollama so DOWN messages aren't shown for
  providers the user never set up.

Bug fixes — Channels / Telegram / Discord

• MOOD/TTS tag leak fix
  (#198):
  Agent [MOOD:X] and [TTS:Y] tags no longer leak into Telegram and
  Discord channel output as raw text. They are stripped from outbound
  text and translated into channel-appropriate metadata where supported.
• Telegram 409 logging + UX
  (#194):
  argent-gateway now includes the bot ID in 409-conflict logs and
  surfaces a persistent-conflict UX warning so users see which bot is
  fighting for the long-poll lock when there are multiple bots
  configured.

Bug fixes — UI / Chat

• Re-auth command preserved in chat error
  (#224):
  When an OAuth token is invalidated mid-chat, the error message now
  includes the literal re-auth command (e.g.
  argent models auth login --provider openai-codex) so the user can
  copy-paste rather than guess.
• Personal skills promotion
  (#210):
  promoteCandidate now increments recurrenceCount on dedup-skip so
  the "promoted" bucket fills correctly. Previously, repeat invocations
  of the same candidate were treated as no-ops and the bucket stayed
  empty.
• Connectors catalog popup spam
  (#152):
  Cache the connectors.catalog probe so the Settings → System tab
  stops triggering a node popup every render.

Bug fixes — Build / Type / CI

• tsc-since.mjs identity hardening (this release's headline,
  #255):
  tsc's per-message length cap truncates error tails with ...' based
  on the un-elided string length, which differs by checkout path.
  argent update's preflight worktree at /private/var/folders/.../
  has paths ~50 chars longer than the baseline-snapshot worktree at
  /Users/sem/code/argent-core/worktrees/<name>/, so the same
  conceptual error got truncated at a different point and missed the
  baseline as net-new. The fix collapses 'import("…")…' blocks to a
  stable placeholder during normalization, making identity
  path-length-independent.
• tsc pre-bundle gate added
  (#178):
  Run tsc --noEmit before tsdown bundles, comparing against a
  267-entry baseline at ops/known-failing.json. Catches net-new type
  errors before they reach dist/.
• Bundle emits composio + builtin-profiles
  (#196):
  tsdown now emits dist/connectors/composio and
  dist/models/builtin-profiles. These features were dead at runtime
  because the bundler was skipping them.
• pi-embedded-runner symlinked worktrees
  (#211):
  tsconfig.json resolution now works when type-checking from a
  worktree whose node_modules is a symlink (the common bump-worktree
  case).
• Worktree-aware pre-commit hook
  (#176):
  Pre-commit hooks no longer fail in bump worktrees that lack their
  own node_modules. The hook resolves oxfmt from the canonical
  install instead of requiring a per-worktree symlink.
• Label-issues workflow auth fix
  (#199):
  Use GITHUB_TOKEN for the label / label-issues workflow instead of
  the broken GitHub App private-key secret. Stops the workflow from
  failing on every PR.

Refactors

• isHeadlessSession consolidation
  (#206):
  isHeadlessSession now delegates to isRemoteEnvironment with a
  documented override, removing duplicate detection logic.

Documentation

• CLI docs depth audit
  (#189):
  Sibling docs/cli/* pages audited and expanded — shallow stubs
  filled in.
• Marketplace + license CLI entries
  (#191):
  docs/cli/index.md now includes marketplace + license commands;
  personal-skills page ...

ArgentOS Core 2026.4.30

2026.4.30

Highlights

• A more reliable personal AI operating system. This release promotes the
  active Core dev line to the public main branch, bringing the newest
  gateway, workflow, dashboard, memory, connector, and installer improvements
  into one stable release.
• Workflows now feel closer to an operator-ready automation surface. The
  canvas has better run visibility, JSON/YAML import/export, owner-operator
  templates, richer binding reports, connector-backed destinations, approval
  records, run history, and clearer failure feedback.
• Reminders become part of the schedule system instead of a separate timer
  stack. Schedule items can now carry reminder intent and delivery targets,
  including in-app, voice, Telegram, Slack, email, and DocPanel routes where
  configured.
• AppForge and Argent Tables move toward a true personal workspace builder.
  The shell now supports more honest Live/Preview/Declaration labels, native
  base and table onboarding, typed field editing, saved table views, record
  add/edit/delete flows, and reload-persistence proof.
• AOS connectors get a large service-key-first expansion. Core now includes
  readiness metadata and harness-backed connector work across tools such as
  Slack, Teams, Discord Workflow, Airtable, Buffer, Hootsuite, Monday, Zapier,
  WordPress, n8n, HubSpot, Make, PayPunch, CallScrub, and more.
• Realtime voice and browser automation foundations move forward. This
  includes browser diagnostics, OpenAI Realtime adapter support, local audio
  smoke tools, operator voice-alert routing foundations, and Google Meet
  setup/status plus browser-only recovery for already-open Meet tabs.

Changes

• Gateway/runtime: pin LaunchAgent and bundled service execution to the
  selected Node 22 runtime, reducing native module mismatches and making update
  behavior more predictable across local machines.
• Gateway status: stop warning about unrelated LaunchAgents as if they were
  duplicate gateways; status output now focuses on actual gateway-like services.
• Workflow scheduling: new scheduled workflow copies start inactive by
  default, and validation warns when multiple active workflows share the same
  cron expression and timezone.
• Workflow payloads: podcast_generate wiring is now validated so missing
  or incorrectly shaped podcast_plan payloads are caught before a bad workflow
  run creates confusing output.
• DocPanel output hygiene: placeholder object envelopes such as demo
  name/body payloads are rejected instead of being saved as operator-facing
  documents.
• Changelog intelligence: agents can now see changelog context directly,
  making it easier for them to answer what changed without guessing from memory.
• Memory readiness: memory recall and category cleanup now surface thinner
  or degraded evidence more honestly, so agents can report when memory context
  is unavailable instead of pretending it worked.
• Personal skills and agent family defaults: coding-family skill mappings
  and persona defaults are now more auditable, with clearer provenance for
  promoted behavior.
• Security readiness: marketplace tool claims now require evidence, security
  audit coverage is broader, and failed memory recall is treated as degraded
  evidence instead of a silent success.
• Telegram stability: repeated polling ownership conflicts are guarded
  against so duplicate pollers stop instead of fighting the same bot token.
• Z.AI subscriptions: operators can choose General/API or Coding subscription
  endpoints with models.providers.zai.baseUrl; ArgentOS appends
  /chat/completions when needed and applies the override to built-in,
  fallback, and discovered zai/* models.

Fixes

• Fixed noisy DocPanel artifacts from workflow result envelopes.
• Fixed duplicate Telegram pollers and repeated Telegram ownership conflicts.
• Fixed false gateway-health warnings after updates and restarts.
• Fixed Node-version drift that could break native modules such as
  better-sqlite3 after local runtime changes.
• Fixed model-drift blind spots by exposing requested model, allowed model, and
  allowlist diagnostics when a subagent model patch is rejected.
• Fixed workflow duplicate scheduling behavior that could let morning workflow
  variants overlap unexpectedly.
• Fixed A2UI missing-asset behavior so the canvas host no longer ships a blank
  or confusing surface when the bundle is unavailable.

Testing Notes

• Workflows: test import, save, reload, bind, validate, schedule, Run Now, and
  failure inspection with the Daily Marketing Brief, VIP Email Alert, Client
  Onboarding, Operations Cleanup, Newsletter Builder, and podcast templates.
• Reminders: test one-time and recurring reminders through the Schedule surface;
  routes should report delivered, skipped, or config-missing rather than
  silently failing.
• AppForge/Argent Tables: test Create Base -> Create Table -> Add Field -> Add
  Record -> Edit Record -> Delete Record -> Reload.
• AOS: configure service keys before expecting live connector reads or writes.
  Harness tests cover the connector packets, but most production accounts still
  need operator-specific smoke tests.
• Realtime voice: set OPENAI_API_KEY before live realtime voice smoke. Google
  Meet recovery requires a signed-in Chrome profile and an already-open Meet
  tab; create/join/leave remain follow-up work.

Known Gaps

• Workflow browser smoke remains the main gate before calling the full canvas
  broadly operator-ready.
• AppForge is not yet a full TableForge/Airtable replacement; relational fields,
  permissions enforcement, connector-backed tables, interface building, and deep
  automations are still follow-up work.
• AOS connector claims are truth-labeled from manifests and harness tests; live
  external-account smoke tests still require operator credentials.
• Browser voice control has a gateway/API substrate, not a finished dashboard
  voice button yet. Remote WebRTC browser smoke and Gemini/Google Live provider
  parity remain follow-up work.
• The Rust gateway remains in shadow mode; TypeScript remains the live gateway
  authority until parity, restart recovery, and protocol drift checks are proven
  over time.

v2026.4.25.2

Supplemental patch for the April Core stabilization release.

Fixes:

• Backports Z.AI provider endpoint selection to main.
• Preserves explicit Z.AI General vs Coding API lane URLs during model compatibility normalization.
• Updates public docs and the bundled Obsidian docs mirror with the Coding subscription setup.
• Aligns package and bundled extension metadata to 2026.4.25.2.

Verification:

• pnpm vitest run src/agents/pi-embedded-runner/model.test.ts src/agents/model-compat.test.ts src/argent-ai/providers/zai.test.ts
• pnpm release:check
• pnpm docs:vault:check
• pnpm protocol:check

v2026.4.25.1

April Core stabilization release

This release moves the last several days of Core stabilization work onto the stable update rail. It is bigger than the short changelog summary: the range from v2026.4.18.3 to v2026.4.25.1 includes Marketplace becoming a real Core capability, the public install/update path being repaired, Core dashboard surfaces being restored, a large docs vault being published, provider routing fixes, migration tools, and the first observable Rust shadow runtime lane.

Marketplace and Core tools

• Marketplace is now a first-class Core tool surface, not a docs-only or legacy skill path.
• Marketplace browse/install claims now require real marketplace evidence, preventing the agent from claiming package availability without tool-backed proof.
• Legacy ClawHub paths were retired from the Core tool flow in favor of argent marketplace.
• Marketplace package installs now support connector harnesses.
• Cognee ships as an AOS Core connector with a machine-readable connector artifact and runtime validation.
• Personal Skills are visible in public Core again.
• Core family agents are exposed without requiring manual subagent allowlists.
• Public Core tools now default to allowed unless the operator explicitly denies or marks them ask-first.

Install, update, and release rail

• Hosted installs now stay outside pnpm workspace resolution, keep clean source checkouts, and identify the current Core build correctly.
• Stable, beta, and dev update channels resolve more consistently from source checkouts.
• argent update now repairs runtime snapshot drift, avoids removing live chunks, rebuilds dashboard bundles when needed, and restarts dashboard/gateway services after public updates.
• Post-update doctor now runs through the fresh CLI so repairs happen against the updated code.
• Skipped-version updates can still apply safe repairs.
• Added scripts/out-of-sync-patch.sh for installs stuck on a pre-fix update runner.
• Release checks now agree with generated protocol artifacts and docs-vault metadata.
• The signed Mac app/public release rail was kept intact, and a false “argent CLI not found in PATH” warning in the macOS app was fixed.

Dashboard and setup experience

• Public Core installs now default to the public Core dashboard surface instead of accidentally exposing full/Business-only controls.
• Core Settings navigation was restored, including Operations, Vault, MemU, Personal Skills, service-key policy routes, and other Core-safe panels.
• Advanced Config tabs are hidden by default behind an Advanced toggle.
• Setup wizard auth tiles gained Groq and Kimi K2 options, plus direct “Get your key” docs links.
• First-run model selection and dashboard model dropdowns now use fresher live catalog data.
• Dashboard status now tracks Core runtime config more accurately.
• Pi/LAN dashboard access was improved with CORS, Vite host filtering, and an opt-in Pi AEVP profile.

Models, providers, and media handling

• Automatic reasoning, routed defaults, fallback streams, and Deep Think now stay on configured compatible providers instead of silently falling back to unconfigured Bedrock or Anthropic paths.
• Deep Think honors the powerful model tier.
• MiniMax and GLM routing were kept on compatible runtime paths.
• Stale MiniMax configs no longer break Anthropic routing.
• Image fallback has more time to describe attachments and can use MiniMax auth profiles.
• Pi streaming was hardened against partial-stream crashes.
• Malformed transcript state and compaction summaries are normalized before model calls.
• Tool-event WebSocket broadcast is decoupled from verbose logging level.
• Prompt-budget diagnostics were added as an opt-in audit path.

Migration, credentials, and service keys

• Added argent migrate export and argent migrate import.
• Migration export now includes models, workspace-specific data, nudges, license state, cron jobs, widgets, and env secrets where appropriate.
• Runtime provider keys now use the auth storage API.
• Connectors can use Argent service keys at runtime.
• Encrypted service-key recovery is more resilient when master keys are available.
• Channel setup now saves usable allowlists.

Memory, database, and runtime repairs

• Core startup repairs Postgres schema drift and memory entity-link schema drift.
• Memory timeline handling now tolerates undefined entity names.
• Hosted git status reports from the source checkout correctly.
• Core kernel, screenshots, Vault, and MemU surfaces were kept usable in public Core.

Docs and operator visibility

• Published the expanded public Core docs mirror and local operator docs vault.
• Added docs for update distribution, Core/Business boundaries, Rust shadow runtime, provider setup, gateway behavior, channels, workflows, and troubleshooting.
• Stabilized docs-vault generation so generated metadata does not drift on repeated checks.
• Added release-readiness and stabilization incident notes for the April Core lane.

Rust shadow runtime

• Added the Rust executive shadow runtime as an observable, contract-tested lane without promoting it ahead of the Node gateway.
• Status reporting now surfaces executive-shadow state.
• Gateway cleanup warnings no longer fire on expected Rust shadow processes.
• Startup waits for shadow services to report running before treating them as ready.

Security and trust boundaries

• Public Core/Business boundaries are documented and enforced in the dashboard surface.
• Marketplace tool-call blocks are validated so fake marketplace evidence is rejected.
• Service-key policy modules are packaged for dashboard API use.
• Secret scanning CI passed for the release-prep PR.

MCP and integrations

• MCP OAuth well-known discovery paths no longer fall into the root-mounted Control UI HTML fallback, so external MCP clients can reach /mcp cleanly.
• Connector catalog/runtime checks were tightened.
• Native Telegram command coverage and channel setup behavior were improved.

Release scope note

This GitHub release publishes the source/tag release consumed by the stable source-checkout update rail. No notarized macOS .dmg or .zip assets are attached to v2026.4.25.1 yet; macOS app packaging remains a separate artifact step.

v2026.4.18.3

Summary

This release ships a major upgrade to the Argent macOS app and Core runtime, including Personal Skills, a cleaner Core Intent experience, and a more stable day-to-day local runtime. The result is a sharper operator experience, better learned behavior over time, and a signed/notarized Mac app release that is ready to install.

Highlights

• Personal Skills
  Argent can now learn operator-specific procedures as reusable Personal Skills. These skills can be captured from real work, reviewed, reinforced through successful usage, and surfaced back into runtime behavior so the system gets more aligned over time.

• Intent Lite for Core
  Core now includes the slimmed-down Safety Rules intent experience, giving operators a simpler, more approachable control surface without requiring the full advanced governance console.

• macOS App Improvements
  The Swift menubar app has been rebuilt and shipped as a signed and notarized release. This update improves launch reliability, menu bar behavior, dashboard access, and local gateway attachment.

Improvements

• Improved Personal Skill lifecycle behavior, including candidate review, procedural structure, usage reinforcement, and decay.
• Improved runtime handling of learned operator procedures so relevant Personal Skills can influence execution more consistently.
• Improved packaged macOS app connection behavior when local auth state has gone stale.
• Improved dashboard/settings experience for Core operators with the restored Safety Rules surface.

Fixes

• Fixed packaged macOS app left-click/menu bar behavior.
• Fixed stale packaged-app gateway auth fallback so local reconnects recover more reliably.
• Fixed false-positive gateway restarts when toggling the consciousness kernel.
• Fixed assistant crash paths that could surface as Cannot read properties of undefined (reading 'filter') during normal conversation.
• Fixed several release-lane stability issues in the local runtime and dashboard path.

macOS Release

• App: Argent
• Bundle ID: ai.argent.mac
• Signed with: Developer ID Application: Jason Brashear (F2DH8T4BVH)
• Notarized and stapled

Install / Update

• GitHub release: https://github.com/ArgentAIOS/argentos-core/releases/tag/v2026.4.18.3
• Website installer manifest: https://argentos.ai/releases/macos/latest.json
• Website artifact: https://argentos.ai/releases/macos/2026.4.18.3/Argent-2026.4.18.3.zip