[0.8.0] — 2026-06-10
Trustworthy execution — multi-provider executors with full run observability. (Jumps from 0.3.3:0.4.x–0.7.0 were consumed by pre-reset publishes and
versions are forward-only — see RELEASING.md.)
Added
- DeepSeek provider via aider delegation (#822) —
provider: deepseekin agent frontmatter (or--providerflag, or squadproviders.default) delegates execution to aider with--model deepseek/deepseek-chat. File-based roles only (no web tools); the OpenAI-compatible seam makes further providers a config swap. - Per-run outcome capture (#818) — observability records now capture what each run actually produced: actions, commits, PRs, issues.
- Provider run observability (#826) — every foreground provider run writes an observability record; real token/cost figures parsed from executor output via the new
CLIConfig.parseUsageseam (implemented foraider/deepseek); agentmodel:frontmatter is now parsed so records carry the agent’s real model.
Fixed
- Executor work can never be lost (#825) — provider-executor output is harvested from the isolated worktree (commit → ff-merge into the project root, guarded by the secret/PII staged-diff scan) instead of being destroyed with it. On divergence the
agent/*branch is preserved with the manual-merge command printed; harvest runs on failed exits too, so partial work survives. - Tag pushes produce GitHub Releases again (#819) —
release.ymldropped its always-failing npm-publish step (onlypublish.ymlis the OIDC trusted publisher), which had been blocking GitHub Release creation.
Docs
RELEASING.md(#827) — in-repo release procedure: publish path, version ladder, squash-divergence recipe, known traps.
[0.3.3] — 2026-06-08
Runtime reliability —squads run is safe and pleasant to leave unattended.
Fixed
- Per-agent timeout now actually bounds agents —
--timeout <min>was ignored in conversation mode; it now caps each agent (precedence:SQUADS_AGENT_TIMEOUT_MINUTESenv >--timeout> default). Default lowered 30 → 15 min so a hung agent can’t burn half an hour (#806). - Founder-context refresh no longer blocks the run — when context is stale, the digest now refreshes in the background while the run proceeds with the current copy (was a multi-minute synchronous Pass-1 over the whole session history).
--force/SQUADS_DIGEST_SYNC=1still refresh synchronously (#807).
Added
- Per-squad-run worktree isolation — each squad run executes in its own git worktree, so agents never switch branches, drop files, or open PRs in your working checkout. Graceful fallback to in-place if the dir isn’t a git repo;
SQUADS_NO_WORKTREE=1to disable (#808).
Changed
squads run(no target) now lists squads and surfaces aRun all squads: squads run --orghint; corrected the misleading “autopilot mode” command description (#805).
[0.3.2] — 2026-06-08
Agent runtime, founder-context, and a full safety/governance layer.Added
- Founder-context layers —
founder-context.md(universal) + per-squadfounder-alignment.mdinjected first into every agent’s context, so squads run aligned with the operator’s live pipeline. squads brief— distills founder intentions from recent sessions into GitHub issues.- Live
--verbosestreaming —squads run --verbosestreams each agent’s output as it works (not just post-run). - Per-agent
max_context_tokens— cap an agent’s context-assembly budget in YAML frontmatter. - Agent Contract — schema + CI validator for agent capabilities (tool grants, write scope, credential scope, resource ceilings).
- OS sandbox (opt-in) — run agent sessions in Claude Code’s OS sandbox with an egress allowlist.
- API type codegen — generated client types from the squads-api OpenAPI spec.
Changed
- Roadmap-bounded autonomy — leads plan, delegate, update state, and land reviewed PRs, but cannot author or ship code themselves; workers do the building.
- Role-based timeouts + anti-collision rules in the conversation engine.
- Requires Node ≥ 20 — dropped EOL Node 18 from the test/release matrices; the bundled
vitest/rolldowntoolchain importsstyleTextfromnode:util, available only on Node ≥ 20.
Fixed
- Release pipeline —
release.yml/publish.ymlno longer fail on Node 18, which had silently blocked the last two tagged releases from publishing to npm. - Telemetry write-key restored (broken since 2026-03-14).
- Services made path-agnostic (no hardcoded paths).
- Agent guardrail Bash denylist now actually fires.
- UX: prerequisites check, no-args squad list, schedule hint.
Security
- Secret/PII guardrail — blocks agent auto-commits that would leak secrets or PII.
- Governance deny-rules — agents can’t edit
goals/priorities/directives/SQUAD.mdduring runs. - Defensive validation hardening across contract / secret-scan / brief / sandbox.
0.3.1 — 2026-04-24
First stable v0.3.x release on@latest. Same code as 0.3.0-rc.1 (burned in on @next).
Note: 0.3.0 was skipped because that version slot is reserved by a deprecated historical pre-release (Jan 2026) and npm enforces version immutability.
Added
- Conversation protocol — agents talk to each other and use tools mid-conversation.
squads run <squad>now drives a lead → scan → work → review → verify cycle. - Org cycle —
squads runwith no target runs all squads in waves, with smart-skip for converged work. - New commands —
review,credentials,goals,log, plus minor refinements toinit,status, andrun. - Project config system —
.squads/config.ymlfor per-project settings (agent_timeout_minutes,token_budget,cost_ceiling,company_name,compose_file,telemetry). Resolution: env var > config file > defaults. - PreToolUse guardrail hooks — agent sessions can be gated by user-defined safety hooks.
- Demo agent scaffold —
squads initnow includes starter agents and “what’s next” guidance. - Growth squad template — added to
squads initseed templates. - Tier 2 documentation — guides for local-services mode (Postgres, Redis, API, Bridge).
Changed
- Run engine rewrite — decomposed into smaller modules (
conversation.ts,workflow.ts, context helpers). Foundation for future cloud execution. - Role-based timeouts — workers, reviewers, and leads have appropriate per-role timeouts (replaces hardcoded 8-minute ceiling).
- Anti-collision rules — multiple squads no longer race to create the same release PR or duplicate issues.
- Prompts extracted — lead briefings, planning instructions, and orchestrator prompts moved from TypeScript into
templates/prompts/*.md. - Services command — agnostic compose-file discovery (no hardcoded internal paths).
- OIDC trusted publishing —
release.ymlandpublish.ymlnow publish via GitHub OIDC instead ofNPM_TOKEN. No long-lived secret to rotate. - Audit remediation — removed hardcoded values, parameterized company name, extracted internal prompts.
Fixed
- Telemetry write-only key — restored after being incorrectly removed in March (telemetry has been silent since 2026-03-14).
- First-run UX — prerequisites check, helpful empty-state for
squads listwith no squads, schedule hint after first run.
Infrastructure
@nextdist-tag channel — pre-release tags (v0.3.0-rc.1,v0.4.0-beta.1, etc.) auto-publish to@nextfor burn-in. Clean semver tags publish to@latest.- npm install via
npm i -g squads-cli@nextfor early access.
0.2.2 — 2026-03-28
- IDP (Internal Developer Platform), observability infrastructure, tiered architecture, org cycle scaffolding.
0.2.1 — 2026-03-13
- First-run experience reset.
[0.2.0] and earlier
See GitHub Releases for the full history.Versions0.3.0,0.4.0–0.4.13,0.5.0–0.5.1,0.6.0–0.6.2, and0.7.0were experimental pre-releases published in early 2026 and have been deprecated on npm. Do not install them. Start at0.2.2or0.3.1+.