Changelog

3.2.0

Sharpens the red-team gate, validated by a defect-injection eval (11 arms: A/B on planted and clean diffs, plus a self-application arm). The eval also refuted two candidate additions (an external-compat angle and blindspot-probe handling): the shipped gate already caught those defect classes, so neither shipped. The eval's self-application arm also surfaced that this repo never committed its own .devloop/ config, fixed here.

Added

• Landing site under docs/. A GitHub Pages site (docs/index.html, docs/llms.txt, and an FAQ) describing the loop, the gates, and install, served from the repo. Shipped after 3.1.1 and carried in this release.
• devloop dogfoods its own config convention. The README instructs users to commit .devloop/config.md and domain.md, but this repo gitignored the whole .devloop/ directory, so red-team's conventions angle returned nothing on every self-review it had ever run. Both files are now committed (.gitignore un-ignores them and keeps trackers/ local state), alongside a short harness-native AGENTS.md that points at them so the rules keep one canonical home. config.md's §Standards to Verify deliberately carries only the rules validate.sh cannot mechanize, each traceable to the release that fixed it; the deterministic checks are not restated, since a mechanical gate beats an advisory rule.
• validate.sh section 19 asserts both files exist, carry the same schema headings required of the examples/ configs, that AGENTS.md points at config.md, and that .gitignore still un-ignores them.
• validate.sh section 20 derives the red-team angle counts from agents/red-team.md (5 correctness + 4 cleanup) and fails if the README, docs/llms.txt, skills/implement/SKILL.md, or the red-team mode table state a different number, closing the doc-drift gap the self-application eval flagged.
• Vendor-neutral Agent Plugins manifest. Added a root plugin.json in the Agent Plugins 1.0.0 format (the cross-vendor standard backed by Amazon, Cursor, Microsoft, OpenAI, and Vercel; Kiro has announced support) alongside the existing Claude Code and Codex manifests. devloop's skills/ layout was already conformant; this adds the manifest, validated against the published schema, so the three skills are structured to install on any Agent Plugins client. Claude Code and Codex are unaffected: each reads its own manifest and ignores the root one (verified with claude plugin validate --strict).
• validate.sh section 21 validates the root plugin.json against the Agent Plugins 1.0.0 fatal rules (schema const, name pattern, closed field set, author fields) and pins version and description identical across all three manifests, closing a drift surface that was previously unchecked even between the two existing manifests. The suite is now 244 checks, up from 219.

Changed

• Convention faults now block the gate. The Phase 4 verdict ladder in red-team FAILs on a CONFIRMED conventions finding whose failure scenario is a genuine fault (security, data loss, a broken documented invariant, or a robustness gap such as a missing rate limit or an unbounded query). Purely stylistic conventions findings (naming, commit-message format, doc location) stay at WARN. Previously only CONFIRMED correctness findings could FAIL, so a real standards violation with no correctness face could pass the /implement gate. The gate wording in the /implement skill matches, and the README and docs/ now describe the conventions angle as a gate-blocking check that runs in every mode, not a cleanup-only one.
• No fixed cap on findings. Removed the ~6 per angle / up to 8 additional numeric limits in red-team, replaced with "surface every candidate with a nameable failure." Aligns the prompt with its own surface-everything-then-verify stance; the no-pad discipline is unchanged.

3.1.1

Documentation and consistency fixes surfaced by two independent release-readiness reviews. No behavior change: the loop, gates, and schemas are identical to 3.1.0.

Fixed

• Examples are concrete, not templates. The README described the examples/<stack>/ files as YOUR_*_HERE placeholder templates, but they are concrete, stack-specific instances. Reworded to "copy the closest to .devloop/ and adapt it."
• Examples surface the directory settings. Each examples/*/config.md now shows the spec-directory / tracker-directory settings (with defaults docs/specs/ and .devloop/trackers/), matching what the README says config.md owns.
• Accurate config cross-references. quality-checklist.md now cites the real config headings (§Standards to Verify, §Blindspots to Check); examples/*/domain.md attributes the Domain-Specific Concerns section to the Clarification Taxonomy (§9), not the Validation Checklist (which has only §1-8); and chunk-template.md's state-holder pointer now resolves to implement SKILL.md §2.2 instead of a non-existent §3.1.
• /implement Process Overview now states that a passing plan_review is PASS or PASS-WITH-WARNINGS (it previously named only PASS).
• Marketplace descriptions de-drifted. Both marketplace.json catalog entries now match their plugin.json description verbatim, so there is one canonical blurb per harness.
• validate.sh housekeeping. The dash guard now also scans examples/*/config.md and examples/*/domain.md (published files a user copies into .devloop/); a stale comment referencing an uncommitted evals/ directory was removed.

3.1.0

A single project-local home for devloop's per-project files: .devloop/. This fixes config discovery under a plugin install (where the skills live in a read-only shared cache and a PROJECT.md beside SKILL.md never resolves in the user's project) and gives trackers a stable home outside docs/.

Added

• .devloop/ convention: skills and agents read project config from .devloop/ in the project root:
  • config.md, engineering config (build/test/lint commands, architecture rules, standards, blindspots, commit conventions, and the spec/tracker directory settings), read by /plan, /implement, review-plan, review-impl, red-team, and /spec (for the spec-directory setting).
  • domain.md, pure domain knowledge (domain context, architecture overview, domain-specific concerns, existing patterns, quality standards), read by /spec.
  • trackers/, home for impl-tracker-<feature>.json, written by /plan.
• Config discovery: each skill and agent resolves config as .devloop/<file> in the project, else generic mode. This is why a plugin install now works: the read-only cache holds the skills, but they read .devloop/ from the project. There is no copied-in PROJECT.md fallback (the old template files are removed); .devloop/ is the only project-config source.
• validate.sh section 17: fails if a core file reintroduces the plugin-cache config pointer ("plugin's skill directory") and requires each skill and review agent to name the .devloop/ home.
• validate.sh section 16 now also rejects en dashes (not just em dashes), closing a gap in the standard-punctuation guard.
• Plugin marketplaces: .claude-plugin/marketplace.json (Claude Code) and .agents/plugins/marketplace.json (Codex, its native catalog location) so devloop installs via /plugin marketplace add KashZod/devloop then /plugin install devloop@kashzod, and the codex plugin marketplace add / codex plugin add equivalents.

Changed

• Config ownership: config.md owns the operational paths (spec directory, tracker directory) alongside the engineering settings; domain.md is now purely domain knowledge. Commit conventions live only in config.md (read by the skills that commit). /spec reads its output path from config.md and its domain context from domain.md.
• Tracker home: /plan writes trackers to .devloop/trackers/ by default (was docs/); /implement, review-plan, and review-impl look there.
• Example configs live in one place per stack under a top-level examples/<stack>/ (typescript-node, python, rust, android-kotlin), each holding a config.md and a domain.md; copy the closest directory to .devloop/. This replaces the split skills/spec/project-configs/ (domain) and skills/implement/project-configs/ (engineering) layout, and validate.sh now checks the examples in a single section (the former duplicate example-config check is removed).
• /implement Phase 3 sizes the red-team half by diff size, the same way /plan sizes work (its Trivial / Small / Medium+ / Large table). A single-file change (or a trivial one with no new logic) runs one red-team in mode: both, unchanged from before. A broader, multi-file or cross-cutting diff (/plan Medium+ and Large) splits the red-team half into parallel mode: bugs and mode: cleanup runs so neither family crowds the other out. review-impl runs alongside in every case. Because red-team in mode: cleanup can apply fixes, the split invokes the cleanup run report-only, so all three concurrent agents only report and the parallel gate stays read-only. No new red-team mode was added; report-only is an invocation instruction inside cleanup mode.
• validate.sh section 18 asserts the Phase 3 spawn stays size-adaptive (it names the mode: both, mode: bugs, mode: cleanup, and report-only markers), so a future edit can't silently revert to the fixed single-agent gate.

Migration

• Move in-flight trackers. Trackers previously written under docs/ now live in .devloop/trackers/, and this release drops the docs/ read-fallback. Move any existing docs/impl-tracker-*.json into .devloop/trackers/, or pass an explicit tracker path when invoking /implement or the review agents.
• Migrate an old PROJECT.md. The copied-in PROJECT.md fallback is gone; .devloop/ is the only project-config source. Split any old PROJECT.md into .devloop/config.md (engineering settings and paths) and .devloop/domain.md (domain knowledge), or copy the closest examples/<stack>/ directory as a starting point.

Fixed

• Valid Claude Code manifest. .claude-plugin/plugin.json no longer enumerates skills/agents as arrays of objects, a shape the current schema rejects (claude plugin validate reported skills: Invalid input / agents: Invalid input). Claude Code auto-discovers skills/ and agents/, so the keys are dropped; the manifest now passes claude plugin validate --strict. The Claude manifest also gains repository and license, matching the Codex manifest.

3.0.0

Three-command split, a harness-agnostic rewrite, and a de-overlapped review layer. The loop is now /spec -> /plan -> /implement, mapping to three gates (spec validation, plan review, code review) one gate per command, with a convergence back-edge where /implement re-runs the plan-review gate in-session. This release rolls up every change since 2.5.0. Breaking change: /implement no longer plans.

Added

• /plan skill: decomposes a spec into an ordered, dependency-aware chunk plan, writes the JSON tracker, and runs the review-plan gate before any code is written. This is the old /implement Phases 1-2.5 (analysis, chunk decomposition, dependency graph, tracker creation, plan-review gate), promoted from a buried mid-/implement checkpoint to a first-class command. The plan-review gate is the most important checkpoint in the loop, now its own visible step.
• Convergence back-edge: when a code-time finding (Phase 3) reveals that the plan was wrong (not just the code), /implement appends corrective chunks to the tracker and re-gates in-session by spawning the review-plan agent directly (not by re-invoking /plan, which would regenerate the tracker), preserving completed chunks and looping under a bounded guard until the plan and code converge. /plan gained a /spec-style detect-existing-tracker branch so a re-run merges into the existing tracker instead of resetting completed work.
• spec_doc tracker field: /plan records the source spec path so /implement and review-impl bind to the exact spec instead of globbing the spec directory (sharpens spec -> tracker traceability).
• Shared concern vocabulary: the seven concerns common to review-plan (plan-time) and the Phase 3 checklist (code-time) are now documented as one vocabulary in quality-checklist.md, so the reviewers speak the same language at both altitudes (the eighth concern is phase-specific: TDD Quality of the plan vs Blindspots in the code).
• Three-state acceptance-criteria verification: review-impl classifies each criterion CONFIRMED / PLAUSIBLE / REFUTED, each backed by a quoted line, recall-biased (default PLAUSIBLE; only CONFIRMED when a real test would go red on regression). Ported from the red-team verification model.
• Spec validation evidence rule: every WARN/FAIL in the spec validation checklist must quote the exact spec line it refers to, the same discipline the review agents apply to code.
• validate.sh checks: a harness-agnosticism check fails if any harness-specific mechanic (Shift+Tab, Ctrl+G, /compact, /rename, --resume, and similar) reappears in the skill or agent prose; the structural suite (~220 checks) also enforces the three-skill layout, phase sequencing, JSON validity, no project-specific leaks, and no em dashes in any published file.
• Empirical validation: the higher-risk changes, the review-impl/red-team de-overlap (does a defect ever fall between them?) and the three-state false-positive catch, were validated with an A/B eval harness over seeded fixtures rather than by inspection alone.

Changed

• /implement is now build-only (3 phases): Load the Plan (locate the tracker, hard-stop unless its plan_review gate passed, orient on the next chunk), TDD Cycle per chunk (red/green/verify), and Quality Verification (8-point checklist + parallel review-impl + red-team gate). It refuses to start the TDD cycle on a tracker whose plan_review is missing or FAIL, telling the user to run /plan first.
• /spec hands off to /plan instead of /implement; its downstream mapping now routes spec sections to /plan (analysis, chunking) and /implement (tests) phases.
• Plan artifacts moved with the plan: tracker-schema.md and chunk-template.md now live under skills/plan/references/; quality-checklist.md stays under skills/implement/references/ (it is the Phase 3 code checklist). Each skill cross-references the one shared file it needs.
• review-impl narrowed to a conformance gate: it verifies plan match, acceptance criteria (with quoted test evidence), test quality, and regression only. Adversarial correctness, robustness/blindspots, standards violations, and cleanup are deferred to red-team, which already does them better. This removes the overlap between the two Quality-Verification reviewers while preserving their conformance-vs-correctness separation.
• Harness-agnostic instructions: removed terminal-specific mechanics from the skill prose in favor of portable behavior. Plan presentation states the principle (planning is read-only; present a plan; get explicit approval) and lets the harness supply the mechanism; context management and session resumption describe the intent instead of naming specific keystrokes or commands. Exploration and check-running steps use conditional phrasing: use a subagent or parallel-tool capability if the harness has one, otherwise sequential is the default. The workflow tables are retitled "Mapping to the Explore -> Plan -> Code Loop" with no harness brand in the header.
• review-plan / review-impl repointing: review-plan's description now says "in the /plan skill"; both agents read "the project's engineering PROJECT.md" rather than "PROJECT.md in the skill directory" (there are now three skills); review-impl's Criterion 5 and the checklist reference /implement Phase 3 (Quality Verification). Agent names are unchanged (review-plan, review-impl, red-team).
• review-plan Criterion 1 renamed "Scope, Completeness & Traceability" with explicit spec -> plan -> tracker forward/backward traceability language.
• Scaffolding trim: default to continuing multi-chunk work in one session rather than resetting between chunks; reset only when context degrades. Chunk decomposition prefers the fewest independently-testable chunks.

Migration

• The workflow gains one user-invoked step: after /spec, run /plan, then /implement. Trackers created by an older /implement run without a plan_review field will be refused by the new /implement; run /plan (pointed at the existing tracker/spec) to gate them, or set plan_review manually if the plan was already reviewed. Hand-setting plan_review: "PASS" bypasses the review-plan gate: /implement trusts the field and cannot tell a gate-written verdict from a typed one.

Fixed (design stress-test)

Hardening from an adversarial review of the whole v3.0.0 design:

• Plan-time gate crash-safety, the symmetric twin of the convergence fix. /plan creates the tracker with plan_review: "PENDING" (never a pre-stamped PASS), writes FAIL to disk before re-running on a gate FAIL, and bounds the FAIL/re-run loop, so a crash mid-review can no longer leave a stale PASS that /implement would build against.
• error and in_progress chunks are no longer dead-ends. error is documented as non-terminal (re-entered like in_progress); /implement Phase 1.3 validates the chunk graph (rejecting depends_on cycles and dangling ids); resumption re-enters an unfinished chunk before searching for the next pending one, so a blocked feature is surfaced, not silently left with the Phase 3 gate un-triggered.
• Convergence re-gate loop is now counted. convergence_rounds is bumped before each review-plan re-gate (not only when chunks are appended), so the two-round cap bounds the re-gate loop too; a bail-out cleanup path is documented.
• Spec back-edge. A finding that an acceptance criterion itself is wrong now routes to /spec (update mode) instead of into the plan gate built to reject it.
• Honest degradation without subagents / without a project rule file. The gates document that a harness with no subagent capability degrades to a non-isolated self-check; /implement Phase 3 covers standards and architecture with a self-check when no PROJECT.md/CLAUDE.md exists (where red-team's conventions angle would otherwise return nothing).
• Docs. Softened the "1:1 gates" phrasing (/implement touches two gates via convergence); README's table notes review-plan's convergence spawn; tracker writes documented as atomic.

2.5.0

Changed

• red-team mode: cleanup can now apply fixes: added Edit/Write to its tools so the tidy pass can edit files, not just report.
• red-team is no longer git-only: Phase 0 shows git as the common case but instructs substituting another VCS (hg, jj, Perforce) or asking the caller for the changed set, and clarifies that a tracker or plan path is context, not the review target.
• /implement degrades gracefully with no PROJECT.md: infers the test/build commands, confirms them with the user, notes the miss in the tracker, and suggests creating one, matching how /spec already behaves.
• Wired the conventions angle into every red-team mode and disambiguated "all modes" from the mode literally named both.

Fixed

• review-impl no longer treats the post-implementation document as mandatory; its absence on a small self-contained change is no longer a false finding, matching the implement skill's conditional-docs rule.
• The plan-review gate branches on the verdict review-plan actually emits (PASS-WITH-WARNINGS) instead of a WARN value it never produces.
• /implement reads the spec directory from PROJECT.md (default docs/specs/), matching where /spec writes, instead of a hard-coded path.
• README install now documents both PROJECT.md templates, points at the project-configs/ examples, and clarifies where PROJECT.md lives for plugin installs.
• Removed every em dash from the repository in favor of standard punctuation.

2.0.0

Renamed from ai-agent-dev-workflow to devloop.

Added

• red-team agent, adversarial diff reviewer that hunts correctness bugs (5 angles) and flags cleanup (reuse, simplification, efficiency, altitude), verifies each finding (recall-biased, 3-state), then sweeps for gaps. Modes: bugs, cleanup, both. The cleanup mode is a standalone tidy pass.

Changed

• Renamed the tdd skill to implement.
• Phase 6 quality gate now spawns review-impl + red-team in parallel instead of review-impl + /code-review. A skill runs in the main loop and cannot invoke another skill or slash command, so it could not trigger /code-review. red-team ports the same finder-angle engine into an agent the skill can spawn via the Agent tool.
• Post-implementation documentation is now conditional (write it when the work outlives the session or has deferred follow-ups; skip it for small closed fixes) instead of mandatory for every change.
• Sharpened the scaffolding-calibration guidance for strong-instruction-following models: prefer fewer/larger chunks and fewer resets; keep the tracker and review gates; make bug review recall-biased-then-verified rather than conservative single-pass.

Removed

• Dependency on /code-review from within the implement skill (a skill cannot invoke it).
• The extension.yml spec-kit manifest. .claude-plugin/plugin.json is the single source of truth; the spec-kit convention added a second manifest to keep in sync with no consumer in this project.