Changelog

Releases are published to npm and GitHub Releases.

​

[0.9.0] — 2026-07-07

The loop release: the CLI teaches the loop, presents it, and enforces its trust properties.

​

The loop, taught and presented

• The seed skill teaches the operating rhythm — intent → execute → decide → learn — with the human-five split from agent working verbs (#1019)
• 62 visible commands become 25 canonical verbs; absorbed names keep working, hidden from --help, each printing its canonical replacement on stderr so piped output and --json stay clean (#1020)
• docs/architecture.md updated to the shipped loop: containment, bounded dispatch, the decision gate, measured feedback, initiative — with the canonical execution-loop diagram (light/dark)

​

Trust enforcement

• Inbox is truthful (#1021): headless/autonomous decisions stamp agent:<name>, never a login identity the human didn’t exercise; run-artifact rows reconcile against live PR state — landed work stops showing as “waiting”
• The trunk is sacred (#966): squad-level --task on a non-tool-use provider lane refuses with remedies instead of silently reinterpreting; provider harvest never fast-forwards the operator’s checkout — work lands on the agent branch and integrates only through the inbox gate
• Fresh bases (#1014): run worktrees are cut from freshly-fetched origin/develop, never a stale local trunk (offline falls back with a visible warning)
• One autonomy story (#970): the orphaned in-CLI autopilot loop is deleted — autonomous cadence belongs to the platform heartbeat and the opt-in ambient squads heartbeat (upcoming)

​

Behavior changes

• Provider-executed agent work is no longer auto-merged into your checkout on run completion — review it via squads inbox (this closes the path that could write to a repo’s main without a human decision)
• --help lists the canonical surface only; squads commands --json --all still enumerates everything

​

0.8.3 — 2026-07-07

The first-run journey release: squads now works on a stranger’s machine, verified by clean-room persona tests.

​

Journey & truthfulness

• Fresh squads init scaffolds agents that can actually run — init/runtime provider vocabulary unified (#955)
• Bare machines (no AI CLI installed) init in planning-only mode with the full provider install list — never a silent claude scaffold, never a wall (#977)
• squads propose — one bounded background run turns repo intent into a reviewable proposal branch + inbox card (#983)
• Every command reports usage via one root hook — command path + flag names + duration + success ONLY; this pass also removed 29 legacy events that leaked squad names and search queries in violation of the telemetry disclosure (#1009)
• Unauthenticated or missing claude fails LOUD with the exact remedy and a nonzero exit (#956, #957); doctor reports real auth state and exits 1 on missing core tools
• README quickstart works verbatim: git init step, auth note, demo verification; deprecated autopilot example removed (#958, decision tracker #970)

​

Local-first trust

• Fresh installs default to the local environment; auto-update is notify-only (opt-in SQUADS_AUTO_UPDATE=1); the optional init email stays on your machine (#959)
• squads never touches your personal Claude settings: the ~/.claude.json trust mutation is gone, generated MCP configs live in your project, session-usage reads are project-scoped (squads usage --all-claude for machine-wide) (#960)
• Init lists the session hooks it installs and asks before enabling push-on-session-end (#963)
• Anonymous usage telemetry disclosed at init; journey events with a blocked-reason taxonomy; CI/test runs never phone home (#964)

​

Works with just git — no GitHub required

• squads inbox approve lands work locally (squash-merge to your trunk) when there’s no remote or gh (#979)
• Machines with no git identity get a commit-scoped fallback — init commits for real and run isolation survives (#980)

​

Smarter conversations (Missions-informed)

• Structured handoffs: turns report commands + exit codes + undone work; self-contradicted DONE claims are rejected (#990)
• Validation contracts: plans define done before code; the verifier checks every assertion (#989, persisted per run #995)
• Bounded remediation: a REJECTED verdict triggers exactly one scoped fix round (#994)
• Deliver-and-stop gate matches closing keywords only (#971); scoped --task runs use a lead+delegate roster (#951)

​

Testing

• Clean-room persona battery (test/e2e/personas/): synthetic users — newcomer, unauthenticated, no-git-identity, no-remote — assert the journey on every change (#997)

​

[0.8.2] — 2026-07-01

Trustworthy execution — per-squad pause/resume enforcement, no silent loss of deliverables, a verified context-injection layer (strategy.md as L1), and a quota-aware org runner that survives quota walls.

​

Added

• Per-squad pause / resume enforcement (#877) — squads pause <squad> makes run, --org, and cron dispatch refuse that squad until squads resume <squad>; the runner prints how to resume or override with --force. Activation state is enforced by the runner and honored by the org planner, so a paused squad can’t be dispatched by accident.
• Context loader: strategy.md is the L1 company layer (#876) — the Squad Context System now reads memory/company/strategy.md as the primary “why” layer (falling back to company.md → directives.md), matching the single-strategy-file model. Context-layer docs updated.
• Quota-aware org runner (#861) — squads run --org probes quota before dispatching; --wait-for-quota polls until the session window reopens instead of stopping. Pre-flight --dry-run prints which squads would run without spending quota.
• Post-run ingest ping (#870) — after a run’s execution record lands in executions.jsonl, the CLI fires a fire-and-forget POST /ingest/trigger to the local squads-api so usage analytics update immediately instead of waiting for the periodic sweep. Silent no-op when the API isn’t running.

​

Fixed

• No silent loss of run deliverables (#875) — a squad run whose lead ended BLOCKED on git/gh write-approval left its deliverable uncommitted in the per-run worktree, which cleanup then destroyed with git worktree remove --force. Cleanup now auto-commits any uncommitted/untracked work to the run branch (recoverable from the shared .git) and best-effort pushes it before removing the directory; if the work can’t be preserved, the worktree is left in place instead of deleted.
• Stale memory no longer reads as current (#893) — feedback.md was injected under “act on this first” with no age caveat (only state.md had one), so a months-old correction looked current. A shared staleness helper now caveats both layers (Last updated N days ago — verify before relying on this). Adds real fixture-based tests for the context loader’s layer order, role gating, strategy.md-as-L1, and budget behavior (previously asserted nothing).
• squads run SQUAD AGENT now routes to the agent (#866) — passing the agent as a second positional (squads run engineering code-review) was silently ignored and ran the whole squad. All three notations now produce identical results: SQUAD/AGENT, SQUAD AGENT, and SQUAD -a AGENT.
• Session-limit quota variant detected (#860) — loud failure printed when quota hits mid-conversation instead of a silent empty result.
• Detached runs pinned to their own session id (#862) — background runs that escaped their session were attributed to the wrong squad’s usage budget.
• Agents can run their own shell scripts (#900) — Bash(bash:*) and Bash(sh:*) added to the agent tool allowlist in both spawn paths (single-agent and conversation), so an agent invoking a co-located helper script no longer stalls on a permission it can never grant.

​

[0.8.1] — 2026-06-11

Run containment + introspection — background runs are recorded, bounded, and controllable; the command tree is machine-readable.

​

Added

• Detached-run observability (#849) — background/scheduled runs write an atomic done-file on exit; the next CLI invocation reconciles it into executions.jsonl with real token/cost usage. Detached runs were previously invisible to squads usage and budgets.
• Watchdog timeout (#850) — detached executors are reaped at their deadline (SQUADS_AGENT_TIMEOUT_MINUTES > --timeout > 15 min); the wrapper survives the kill, so a reaped run still harvests its work and reports status: timeout. Built from live evidence of an executor deadlocking post-completion.
• squads runs / squads kill (#852) — live background-run inventory across all squad repos; graceful stop (executor first, so the run still reports); runs --clean salvages crashed runs and clears stale pid files.
• squads commands --json (#842) — the live command tree as data, from the Commander registry; feeds the docs site, the seed skill, and agent discovery.
• SQUADS_AIDER_MAP_TOKENS (#847) — cap the aider executors’ repo-map token budget (measured ~4.6k overhead on a small repo, far more on monorepos).
• Release → docs dispatch (#843) — releases notify the docs repo to regenerate its CLI reference (requires DOCS_DISPATCH_TOKEN; weekly cron otherwise).

​

Fixed

• Provider runs route to the squad’s bound repo (#846) — single-agent provider runs ignored SQUAD.md repo: and worktree’d/harvested onto whatever repo dispatched them. Live-validated both directions.
• One canonical commit identity per AI provider (#839) — Claude <noreply@anthropic.com> everywhere; detached harvests author as the user’s git identity instead of a hardcoded bot email (was inflating contributor counts with phantom entries).
• Seed skill reference generated, never hand-written (#848, #851) — built from squads commands --json with a CI drift-guard; ships in squads init so new users get a current capability map.

​

Security

• Cleared all 9 dependabot alerts (6 high) — lockfile-only transitive bumps (#833).

​

Docs

• README rewritten: 640-line manual → 105-line front door + 8-page docs/ (#836); openapi-ts config co-located with its spec (#835); AGENTS.md is the single agent-instructions source, CLAUDE.md imports it (#841).

​

[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: deepseek in agent frontmatter (or --provider flag, or squad providers.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.parseUsage seam (implemented for aider/deepseek); agent model: 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.yml dropped its always-failing npm-publish step (only publish.yml is 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_MINUTES env > --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=1 still 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=1 to disable (#808).

​

Changed

• squads run (no target) now lists squads and surfaces a Run all squads: squads run --org hint; 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-squad founder-alignment.md injected 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 --verbose streaming — squads run --verbose streams 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/rolldown toolchain imports styleText from node:util, available only on Node ≥ 20.

​

Fixed

• Release pipeline — release.yml/publish.yml no 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.md during 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 run with no target runs all squads in waves, with smart-skip for converged work.
• New commands — review, credentials, goals, log, plus minor refinements to init, status, and run.
• Project config system — .squads/config.yml for 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 init now includes starter agents and “what’s next” guidance.
• Growth squad template — added to squads init seed 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.yml and publish.yml now publish via GitHub OIDC instead of NPM_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 list with no squads, schedule hint after first run.

​

Infrastructure

• @next dist-tag channel — pre-release tags (v0.3.0-rc.1, v0.4.0-beta.1, etc.) auto-publish to @next for burn-in. Clean semver tags publish to @latest.
• npm install via npm i -g squads-cli@next for 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.

Versions 0.3.0, 0.4.0–0.4.13, 0.5.0–0.5.1, 0.6.0–0.6.2, and 0.7.0 were experimental pre-releases published in early 2026 and have been deprecated on npm. Do not install them. Start at 0.2.2 or 0.3.1+.