v0.2.1 — deep-dive hardening

Hotfix + hardening from the first adversarial deep-dive of the system itself (5 finders → 48 findings → 2 independent verifiers). The theme of what it caught: in a handful of places, "machine-checkable" was an instruction rather than a mechanism.

• Installer fix — the documented multi-harness command (--claude --codex --opencode) silently installed nothing for Codex: --claude was never read and --opencode overwrote the harness choice last. The installer now installs to every requested harness root; --global is accepted as the (default) documented spelling.
• files_touched reality check — new scripts/files-audit.js: Wave 2 diffs the actual changed files against declared card ownership and reported files_touched, so an edit a worker omitted from its report can no longer hide. The one hole in the wave-parallelism safety story, closed with an exit code.
• Scoped redo cap survives sessions — STATE.md gains a "Scoped redo spent" line written at fix-worker dispatch; the one-scoped-attempt-ever rule no longer lives in anyone's memory across a session boundary.
• PARTIAL joins the Wave-2 mechanical flags — execution.md's flag list now matches the integrator agent and the relay rule; an unfinished relay story can't be merged by an integrator following that section literally.
• Split statement wired as a required step — /aa-new-project and /aa-import now name the "my part / your part" narration explicitly instead of relying on playbook prose an agent might skim past.
• DESIGN.md catches up to v0.2 — research fan-out subsystem and both researcher agents documented; PREREQS description gains the "When needed" column.
• Plugin manifests bumped to match the released version (missed in the 0.2.0 release pass).

v0.2.0 — machine-checkable handoffs

Agents fail at handoffs, so this release makes every handoff machine-checkable. Grounded in Anthropic's CwC 2026 material (a subagent returning "fairly confident" instead of 0.72 failed an eval; 216k→7.6k tokens by scripting instead of reading) and a full survey of the lane — every steal-worthy idea from GSD, BMAD, gstack, spec-kit, Kiro, Agent OS, superpowers, and Ralph is either in here or documented as rejected.

• Typed story reports — SN.report.md is now a fixed schema the integrator parses, not reads: status, files touched, acceptance command + verbatim output, deviations, contract-change requests, dead ends. Prose demoted to a notes: field.
• Contracts pin exact shapes — every cross-story interface needs the exact return shape, one populated example, and the failure shape. Signatures alone are not a contract.
• Graders + control cases — acceptance checks declare one of 5 grader types; every epic re-runs a previously-passing control case in Wave 2 so a "fix" that regresses something is caught mechanically; repair-loop fixes mint new control cases (capped).
• Bounded repair loops — story self-verify and Wave 2 verify run up to 3 repair rounds against the deterministic check before FAIL; unsatisfiable-check diagnosis (same finding + unchanged files after a redo = the check is wrong) stops the expensive failure spiral.
• DECISIONS.md ledger + ambiguity protocol — settled questions are never re-asked across sessions; ~70–80% of ambiguities auto-resolve via encoded principles; security/paid/irreversible always escalate.
• Readiness dashboard — a mechanical gate table before Wave 1: contracts frozen, every card typed, every check graded, control case declared.
• Design layer (playbooks/design.md) — UI stories carry a data-verify-* DOM contract with headless __verify.runAll(); verifier walks a 10-item falsifiable design audit; every epic's review gate renders a clickable DEMO.html from parsed reports; 2–4 HTML design directions pre-Wave-0.
• The agent tree (playbooks/system.md) — two-level tree, depth by sequencing never nesting; four-part dispatch briefs (inputs by path, never pasted); typed ~1–2K compressed returns; dead_ends survive handoffs; one writer per artifact; STATE.md externalized at every wave boundary.
• Context budget doctrine — orchestrator operates in the front half of its window (≤40–50% fill); predictive check at wave boundaries (fill + next-wave estimate ≥60% → hand off first); workers can return PARTIAL for one orchestrator-dispatched continuation; auto-compact is an emergency brake, never a plan.
• Steal-list v2 — EARS grammar for acceptance checks/KRs, discover-standards step in /aa-import, counter-metrics ("do not optimize") per objective, completion-promise gate on autopilot exit, quality/balanced/budget cost profiles, constitution versioning on PROJECT.md, broad-regression cadence at panel-refresh epics.
• Research fan-out + risk fixtures — multi-epic roadmaps dispatch two parallel smart-tier researchers (domain edges, ecosystem prior art) after the first roadmap draft, producing a RESEARCH.md risk register. Risks ranked fixture become fixture stories pinned before feature stories build on the edge; an ignored fixture row is a BLOCK.
• Panel-pass scaling — the critic-panel gate scales with roadmap size: 1 epic skips, 2–4 get one full pass, 5+ add a spec+execution panel refresh at every third slicing. Bounded, position-derived, no stored state.
• QUICKSTART.md — step-by-step first run, and the README quickstart no longer skips /aa-new-project.

12 of these land below the user surface — no new commands, no new flags. The command surface stays at ~11.

v0.1.2 — stress-test hardening

Fixes what the head-to-head stress test surfaced the day v0.1.0 shipped.

• Scoped redo: patch-sized redo-lists take a cheap path — one fix worker + re-verification instead of a full wave re-run (~275–300k tokens saved per cycle). Breaker ladder bounded at fail → scoped → fail → full → fail → stop.
• Contracts own their invariants: every data-store invariant names an owner — writer guarantees it or readers tolerate violations.
• --codex --local installs now resolve: ./.agents/agent-agile/playbooks joined the resolution rule (project-local before home-level).
• /aa-panel wording matches the critic agents (Panel protocol + own section).
• npm hygiene: tarball allowlist, repo metadata, Windows-safe npm test.

Full notes: RELEASE-NOTES.md

v0.1.0 — first public release

First public release. Every adopted planning system for coding agents starts after someone decided the project is worth building — this one makes that decision part of the system.

• The Grill (/aa-grill) — idea interrogation that can reject the idea before any planning tokens are spent
• OKR → Initiative → Epic → Story hierarchy with the worker-readiness test (cheap models execute; if a card needs a smart model, the card is wrong)
• Adversarial critic panel (/aa-panel) with a market critic holding kill authority
• PREREQS.md human-shopping gate — preflight refuses to launch on unverified prerequisites
• Wave execution with contract freeze + file-ownership collision refusal; /aa-autopilot with gate modes and circuit breakers
• One skill tree, three harnesses: Claude Code / Codex CLI / OpenCode

Install: /plugin marketplace add Peeyushmeher/agent-agile → /plugin install agent-agile