refactor(framework): executor/checker agents + SDLC orchestration redesign (recipe-not-spawn) by blafourcade · Pull Request #314 · ai-driven-dev/framework

blafourcade · 2026-06-20T05:38:18Z

🎯 What & why

Redesign the SDLC flow and its agents around one rule: a recipe skill never spawns; only the orchestrator does. The trio planner/implementer/reviewer collapses to two leaf agents, executor + checker, and the SDLC owns planning itself. Every flow skill is renormalized to the skill-authoring contract, made tool-agnostic, and taught to defer to project memory. Net +673/-1269 — the redesign removes more than it adds.

🛠️ How it works

Agents: trio → two leaves + orchestrator-owned planning.

Delete planner.md, implementer.md, reviewer.md.
Add executor.md (runs 02-implement) and checker.md (runs 05-review, embedded checklist per feat(framework): add review checklist consumed by the reviewer agent #253).
00-sdlc is the only orchestrator: it authors the plan inline (01-plan in its own context) and spawns exactly two workers. executor never holds 01-plan — structural immutability, it can't rewrite the contract it builds.

recipe-not-spawn. Spawning is an orchestration decision, encoded in docs/ARCHITECTURE.md. The recipe skills (01-plan, 02-implement, 03-assert, 05-review) no longer mention spawning, other skills, or the SDLC — they focus on themselves. 00-sdlc names its own-plugin skills by address and discovers cross-plugin VCS capabilities by description (cross-plugin orthogonality), so the ship step never names aidd-vcs.

3-axis review into one report. 05-review runs code, behavior-vs-plan, and a new review-relevancy axis (fit / conform / rot) into a single review.md written to the feature folder.

Project-memory deference. commit, pull-request, and assert treat project memory (vcs.md, coding-assertions.md) as the source of truth and the bundled templates as the fallback. The SDLC ship step no longer dictates a commit format.

Tool-agnostic. pull-request reads the VCS tool and request template from project memory; no gh/GitHub literal in skill logic.

Norm + cleanup. All flow skills follow the prose action anatomy (singular Input, one-line Output, one-word bold step labels, no frozen YAML, no empty sections). Orphan assets dropped; a stray user-stories WIP that duplicated 02-user-stories-create removed. An independent audit against the contract found 8 issues, all fixed.

🧪 How to verify

Headless, all passing on this branch:
- commit follows a custom aidd_docs/memory/vcs.md format over conventional.
- pull-request reads tool + base + body sections from memory (not main, not the bundled template).
- assert reads aidd_docs/memory/coding-assertions.md and fixes until green.
- full SDLC spec → plan → implement → review: feature folder holds all 4 artifacts; raw stream confirms executor + checker are really spawned (Agent tool), not run inline.
git diff --name-status origin/next...HEAD — agents swapped, no foreign files.

⚠️ Heads-up

Multi-iteration review loop (N≥2) and a full human-interactive multi-gate walkthrough are not headless-testable; verified by construction, not by run.
02-user-stories-create (the real skill) is untouched; only the leaked partial duplicate was removed.

🔗 Linked issue

Refs #253

✅ I certify

I DO CERTIFY I READ EACH LINE OF THE PULL REQUEST BECAUSE I AM A SOFTWARE ENGINEER, NOT A AI PUPPY.

…elegate SDLC implement Splits 02-implement into three focused actions (prepare → execute → finalize) with a branch guard, and commits each plan/phase status transition immediately so the implementer's clean-tree hygiene cannot revert it. 00-sdlc/03-implement now delegates to the implement skill instead of re-spawning the implementer, deleting the duplicated status-write pattern. Interactive gate 3 fires after the whole implement step. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…t to follow-only Phase done-gate moves from the implementer's self-reported completion_score to the assert skill passing (objective). Drift no longer rewrites the plan in place; any mismatch surfaces 'replan needed' to the caller (replanning is the planner's job). Also brings the actions to skill-authoring conformance: one-word step labels, no semicolons in process steps. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…stic Router drops how-detail and agent mentions (description trimmed to what+when; status lifecycle moved into a transversal rule; commit rule scoped to the skill's own tracking). Execute steps go tool-agnostic: no `Task`, delegate to the implementer by role, assert the phase instead of naming the assert skill. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ents ARCHITECTURE.md gains the keystone: skills are caller-agnostic recipes, agents are leaf executors, a skill may delegate to an agent and an agent may invoke a declared skill list (never read a skill's files), only the conductor spawns. The planner/implementer/reviewer agents drop hardcoded plugin paths for capability names (R25), replace orchestration-sense 'Planner' with 'conductor' (resolving the implementer's return-to-Planner contradiction), and the implementer carries its BLOCKED criteria inline instead of pointing into the implement skill's files. Precious content preserved. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Agents drop to a one-or-two-sentence Role plus Behavior/Guardrails/Skills; project conventions (gitignore hygiene, provider-test style) leave the agent since they are not its to decide. Planner trims to plan, reviewer keeps review and audit. ARCHITECTURE adds the rule: name a same-plugin skill directly, a cross-plugin skill by capability. SDLC 03-implement is rewritten agnostic (no orchestrator wording) in prepare-style with implemented/blocked sub-cases. Em-dashes removed throughout. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… by capability Agents reference a same-plugin skill by its deterministic plugin:folder address (aidd-dev:03-assert) and a cross-plugin skill by capability (commit). Refines the agent-authoring rule and template, and the ARCHITECTURE composition rule, from never-hardcode-a-plugin-name to never-hardcode-a-cross-plugin-name. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ostic rules The five SDLC actions prose-ify their outputs (no frozen YAML), drop the self-referential 'return to the SDLC orchestrator', remove the cross-skill plan-status reference, name same-plugin delegates by address (aidd-dev:01-plan, 02-implement, 05-review) and cross-plugin ones by capability, use one-word process labels, and singular Input/Output. No em-dashes. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…diff-guard Make spawning an orchestration decision, never a skill's. The 02-implement recipe no longer spawns; the SDLC spawns one leaf per step (planner, implementer, reviewer), each running a recipe in its own context, so the write path stays two layers deep. This dissolves the cross-context status-commit bug instead of patching it. - 02-execute drops the per-phase agent spawn; the recipe codes inline and commits one unit per phase (code + status), no separate in-progress marks. - SDLC iterate hands findings as a fix list, no plan edit; the plan stays implemented until a review of the current diff passes. - 05-ship gates on a reviewed-SHA carried by the verdict: only plan-tracking files may differ since review, so post-review code cannot ship unreviewed. - ARCHITECTURE: recipes never spawn; an agent runs only recipes, may spawn a read-only recon helper; delegation cannot cycle. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… planning Define agents by durable stance, not as skill aliases. Two workers replace the three thin ones; the orchestrator owns the plan. - executor (was planner + implementer): a dispatched doer that turns a scoped task into working, validated code. It never holds the plan-authoring skill, so it cannot rewrite the contract it implements. - checker (was reviewer): judges against the validator and the real need, with an embedded behavioral checklist (duplication, incoherence, over-engineering, dead code) extended by the project's own checklist. - 00-sdlc (B1): the orchestrator authors the plan itself (runs 01-plan in its own context) and spawns executor for build, checker for review. - 05-review de-spawns (was context: fork + agent: reviewer): it is a recipe the checker runs, so it cannot double-spawn under the checker. - Sweep every reference (plan-status iterate stays implemented, 10-todo, ARCHITECTURE, GLOSSARY, agent-authoring, READMEs). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The agents array still pointed at the deleted planner/implementer/reviewer files, so no aidd-dev agent registered and the SDLC silently fell back to general-purpose subagents. Point it at the new executor and checker. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ility The SKILL.md had grown to seven business-logic sections with heavy duplication and hardcoded cross-plugin skill names. Bring it back to the router contract: description, action table, short transversal rules. - SKILL.md down to ~30 lines: six one-line transversal rules, no Modes / Mandatory / Tracking / Gates / References sections, no empty Assets block. - Cross-plugin steps delegate by capability discovered at runtime, in both the action table and the spec/ship actions, not by hardcoded skill name. - Intra-plugin recipes stay named by address for deterministic resolution. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Cut the router description to what + when + modes: drop the internal "pure orchestrator / holds no logic" prose. Natural-language routing verified: end-to-end intent triggers 00-sdlc, single-step requests route to their own skill. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…mand A present human typing /sdlc should see the gates; only a deliberate, configured caller (headless, orchestrator, Gardener) wants an unattended run, and that caller passes `auto`. Opting in to autonomy matches intent, and a forgotten `auto` fails loud at gate 1 instead of shipping unwatched. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

A third review axis next to code and functional: does the change belong? Fit to the real need, conformance to the project's declared rules, and no duplication or over-engineering. It absorbs rules-conformance (slimmed out of review-code, which is now clean-code only) and stays read-only: findings hand off downstream per complexity, never an auto-fix. The SDLC checker now runs all three axes. - new action 03-review-relevancy + its report template (fit / conform / rot) - review-code slimmed to clean-code; rules moved to relevancy - 05-review router, 04-review brief, and READMEs updated to three axes Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…axes Bring 05-review up to the skill contract and make it produce a single report. - Run all three axes by default, or one when named (was pick-one routing). - One `review.md` in the feature folder beside plan.md (was three clobbering files); each axis fills its section, unrun axes marked "Not run". - One overall verdict, strictest across axes (was two vocabularies). - Actions rewritten to the anatomy: prose singular Input/Output, no frozen YAML, no stray Rules section. - Shared scales (severity, verdict, categories, lenses) moved to references/review-rubric.md; one assets/review-template.md replaces three. - Drop the orchestration meta from 05-review and 02-implement: a skill does not describe how the SDLC spawns or isolates it; that lives in the SDLC. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Cut both to two lines and drop the SDLC mention from implement: a skill's description states what it does and when, not how an orchestrator calls it. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

A worker skill surfaces findings; it does not decide who fixes them. Drop the handoff naming (07-refactor / 02-implement / 08-debug) from the SKILL router, the actions, and the report template - each finding now just describes its fix, and the caller routes it. Keep only the one 04-audit disambiguation in the description, where a sibling could mis-trigger. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Keep each skill on its own job. 01-plan description cut to two lines (drop the internal gather/explore/wireframe process and a redundant redirect). 02-implement execute drops the leftover spawn-meta and defers replanning to "the caller", not "the orchestrator". The SDLC 03-implement stops restating the spawn/isolation discipline that already lives in its Orchestration rule. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Descriptions name no other skill, not even in the Do-NOT (just "write code", "review a diff"). The plan-status reference describes the lifecycle by step ("the implement step", "the review step") rather than by skill or layer name, and 02-implement's README drops the leftover spawn-history. SKILL routers and actions are now clean of cross-skill mentions, except the SDLC's own delegations. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

review.md follows the plan/spec/brainstorm convention: one file in a feature folder. Spell the path, and handle the no-plan case: reuse the reviewed work's folder when it has one, otherwise resolve one from the change (branch or slug). The review never creates a plan. The functional axis asks for the criteria and marks itself "Not run" when none are available, so an unattended run never hangs. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

review.md is a snapshot of the current diff, not a history: re-running the review overwrites the single file rather than appending or versioning. This is what the iterate loop needs, a later review of the same work replaces the earlier one. Verified: two runs leave one review.md, refreshed. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…iew.md The template carried name/description/argument-hint frontmatter that copied verbatim into the output report. Replace it with an HTML authoring comment so review.md starts at its own H1. Verified: output begins with "# Review:". Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Bring 03-assert to the same standard as review and implement. - Actions rewritten to the anatomy: prose singular Input/Output, no frozen YAML. - Routing harmonized: run every applicable facet by default (coding always, frontend when a UI and URL exist, architecture only when asked), or one when named, no forced question, so the auto path never blocks. - Framed as a gate: it returns a pass or fail verdict, not a stored report; coding and frontend fix until they pass, architecture only reports. - Fix the task-template frontmatter leak (HTML comment, not YAML) and the stray brace, like the review template. - Short description, no sibling-skill names. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The skill was framed around a "feature", too narrow: assert is the gate the executor runs to validate whatever it just built (a phase, a change), not only a feature. Generalize the unit to "the work" across the skill, actions, and README. No behavior change. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The facet required the running frontend URL as a handed-in input, so in an unattended run nobody supplied it and the facet always skipped. Make it resolve the URL itself: use a given one, otherwise discover the running dev server (its start script, the framework default port, a listening port), confirm it responds, and only skip when no frontend is actually running. It then navigates to the screen the expected behavior targets. Never starts a server. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…ol agnostic The URL-discovery step over-specified the how (dev script, default port, listening port); cut it to "use the given URL or find the running frontend, confirm it responds, else skip". The browser tool stays the project's choice (documented in its browsing memory), named generically as "the project's configured browser tool", never a hardcoded one. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Bring aidd-pm:04-spec to the same standard as the dev skills. - Actions rewritten to the anatomy: prose singular Input/Output, no frozen YAML. - Short two-line description; drop the orchestration trigger ("an orchestrator needs a contract before planning") and the sibling/agent chatter. - Output stated once as a transversal rule: one spec.md in the feature folder, beside plan.md, from the template. - Drop the spawn-meta ("the caller spawns a reviewer"): a spec is validated against spec-validator.yml; how is the caller's concern, not the skill's. - "reviewer findings" become "review findings"; remove the empty References and External-data sections. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The full test caught the report file being skipped: the model reported the verdict inline and claimed a path without writing the file, in both the standalone and SDLC paths. Make the output rule imperative: the deliverable is the file, always written to disk. Verified: review.md now written reliably. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The full test drafted a spec full of plan detail: file names, components, hooks, routes, tokens, and eleven granular criteria, violating the spec's own "intent, never implementation" rule. Forbid exploring the codebase or naming any file, component, or API; keep done-when outcome-level and the criteria few. Verified: a build with code present named zero implementation details. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

One bullet was carrying eight things (deliverable, write, path, reuse, resolve, no-plan, sections, not-run). Split into three one-line rules: Output (always write the file), Folder (where, never a plan), Sections (template, one per axis, "Not run"). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…view A review writes review.md; it has no reason to create a plan. Forbidding it defends against a non-event. Removed. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…lines The four were running ~three lines. Cut to the essential what + when + key do-not, under 210 chars each, keeping the trigger words. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…redesign # Conflicts: # README.md # plugins/aidd-dev/agents/reviewer.md

…the norm Actions rewritten to the anatomy: prose singular Input/Output, no frozen YAML. commit description cut to two lines; "Available actions" -> "Actions"; drop the empty References and External-data sections in both. pull-request keeps its prefix-routing base resolution and triage labelling (from #326), in prose. commit validated: auto mode produced a conventional commit and sha. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The checker scanned the whole tree, including the full repo copies under .claude/worktrees/* and the .specstory transcripts, reporting ~7700 false "broken" links (template placeholders, history) and blocking every commit. Add worktrees and .specstory to the skipped dirs. Real source: 0 broken in 370 files. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The plan kept drifting to aidd_docs/plans/ instead of the feature folder. State the path imperatively where it is authored (01-plan output) and in the SDLC folder rule: every artifact lands in aidd_docs/tasks/<yyyy_mm>/<slug>/, never aidd_docs/plans/ or anywhere else. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Stating the wrong path ("never aidd_docs/plans/") is over-specification. The imperative "always at aidd_docs/tasks/<yyyy_mm>/<slug>/" is the enforcement on its own. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The commit, pull-request, and assert skills now treat project memory as the source of truth and the bundled defaults as the fallback: - commit follows the convention in aidd_docs/memory/vcs.md when set, else the conventional-commit template. - pull-request prefers the project's PR template (vcs.md or the repo PULL_REQUEST_TEMPLATE) over the bundled body template. - assert names aidd_docs/memory/coding-assertions.md as the assertion source when present. A skill ran in a spawned context does not auto-load memory, so each reads the file explicitly and falls back cleanly when it is absent. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Drop the GitHub-specific PULL_REQUEST_TEMPLATE literal from the skill logic; refer to "the repo's own request-template file, wherever the configured tool keeps it" so GitLab, Bitbucket, and others are equal. Also align the description with the skill norm: add the missing "Do NOT use to..." clause, state the tool comes from the project, and fix the "user ask" grammar slip. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The ship step briefed the commit capability to use Conventional Commits, which overrode the commit skill's project-first deference to aidd_docs/memory/vcs.md. Brief only the plan objective and let the commit capability pick the message format per its own convention. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Trim the wording added for project-memory deference across commit, pull-request, assert, and the SDLC ship step. No behavior change: same precedence, fewer words. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

An independent audit against the skill-authoring contract surfaced one major and seven minor violations across the flow skills. Fixes: - 02-implement/02-execute: drop "via the `commit` skill" so aidd-dev no longer names an aidd-vcs skill; it discovers the commit capability at runtime, matching 05-ship and 03-finalize (cross-plugin orthogonality). - 01-plan and 03-assert: collapse multi-word step labels to the one-word bold lead the action anatomy requires. - 02-pull-request: drop the orphan CONTRIBUTING.md and README.md assets (project-scaffolding scope, never loaded here) and regenerate CATALOG. - 02-implement/blocked: blank line after the heading, split the run-on. - 01-commit: rewrite the router subtitle so it stops restating the description and drops the gerund. No behavior change; routing and outputs are unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

A stray plugins/aidd-pm/skills/02-user-stories/SKILL.md (SKILL.md only, no actions) was swept into an unrelated commit from a local stash. It duplicates the complete 02-user-stories-create skill already on next and is broken on its own. Remove it; regenerate the catalog. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(skills): sync argument hints (#296) * refactor(aidd-dev): redesign 01-plan into gather/explore/wireframe/plan (#271) Four-action plan skill (gather/explore/wireframe/plan), unified feature-folder artifact layout (brainstorm/spec/plan/phases/review), implement reads the new layout with per-plan and per-phase status tracking. Closes #292 Closes #265 Closes #276 * docs(framework): add dominance checks to review and rules (#313) * docs(framework): add dominance checks to review and rules Dominance issues were easy for agents to miss when a weaker rule, criterion, or documentation statement was already covered, overridden, or invalidated by a stronger element. This adds the check at the global context, rule-generation, reviewer, and review-template levels. * docs(framework): reduce dominance check duplication Keep the concise dominance rule in global and rule-authoring surfaces while leaving the detailed case breakdown to the reviewer. Validation now states the shared check without reimplementing reviewer classification. * docs(framework): separate global and rule dominance checks Keep the general dominance reminder and reviewer taxonomy separate from rule-generation specifics. Rule authoring and validation now only check for duplicate, weaker, or contradictory rule bullets. * docs(framework): prefer intention-revealing names Naming guidance belongs in the global project context because it applies across artifacts. This keeps the rule concise and avoids duplicating it in domain-specific rule generation flows. * docs(aidd-context): validate rules against shared contracts Rule validation should apply the rule-authoring and tool-path contracts instead of restating individual rule checks. The test now has a single outcome sentence rather than multiple bullet assertions. * docs(aidd-dev): align dominance wording with review role The reviewer should report parallel elements rather than describe adding behavior. This keeps the dominance check phrased for read-only review while preserving the same taxonomy. * ci(framework): trigger Validate on next branch (#317) Required check "lefthook (framework-local checks)" (next.json ruleset) is published only by validate.yml, whose pull_request/push filter was [main]. PRs targeting next never fired it, so the required status stayed "Expected — Waiting for status" forever, deadlocking merges (e.g. #307). Add next to both push and pull_request branch filters. Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * feat(framework): add markdown link checker (#307) * feat(framework): add markdown link checker * fix(scripts): exempt template scaffold placeholder links 01-plan's plan-template.md / phase-template.md link to ./plan.md and ./phase-N.md — files the skill emits next to the generated output at runtime (see actions/04-plan.md), never committed to the repo. The checker flagged them as broken. Extend the existing asset-template branch: a dot-relative target in a *-template.md that resolves nowhere is an intentional placeholder for a generated sibling, so treat it as ignored rather than broken. The assets/templates fallback (used by 02-project-memory) is unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Baptiste LAFOURCADE <baptiste.lafourcade@gmail.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * feat(aidd-ui): scaffold alpha plugin (0.1.0-alpha.0) (#319) * feat(aidd-ui): scaffold alpha plugin (0.1.0-alpha.0) Add aidd-ui as a new plugin with a single smoke-test skill (01-hello). Registered in marketplace.json (recommended: false) and release config; alpha status is conveyed by the 0.1.0-alpha.0 version in plugin.json. Lives on a dedicated branch off next, so it is not public until it graduates to main. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs(aidd-ui): mark alpha / not-ready across surfaces Add prominent alpha warnings to the plugin README, the homepage plugins table, and the marketplace and plugin descriptions so the not-ready status is unmistakable. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * chore(deps): target next instead of main for dependabot (#324) Dependabot had no target-branch, so its PRs defaulted to main. main only takes the weekly promotion PR and the auto-merged Release PR (RELEASE.md), so dep updates landing there cut off-cycle releases (js-yaml bump shipped v5.0.3 mid-week). Point both ecosystems at next so updates batch into the weekly release like all other work. Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * feat(aidd-vcs): add repo-init skill (init + publish a repository) (#269) New skill aidd-vcs:00-repo-init: init a project's repository (git init, default branch, CONTRIBUTING.md, bootstrap commit) and, on request, create the remote and push. Provider- and mechanism-agnostic: host + reach (CLI, MCP, or API) resolved from VCS memory or environment, main as default-branch fallback. No hardcoded provider or fixed mechanism. Conforms to the skill-authoring contract (R1-R13 + action anatomy). Co-authored-by: Baptiste LAFOURCADE <baptiste.lafourcade@gmail.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs(framework): unify change taxonomy into one source of truth (#325) * ci(deps): bump actions/checkout from 6.0.3 to 7.0.0 (#321) Bumps [actions/checkout](https://github.com/actions/checkout) from 6.0.3 to 7.0.0. - [Release notes](https://github.com/actions/checkout/releases) - [Changelog](https://github.com/actions/checkout/blob/main/CHANGELOG.md) - [Commits](actions/checkout@df4cb1c...9c091bb) --- updated-dependencies: - dependency-name: actions/checkout dependency-version: 7.0.0 dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * docs(framework): unify change taxonomy into one source of truth The same change taxonomy (kind -> where it ships) was re-spelled across branch prefix, commit type, label, and the board "Work Type" field, and the four drifted. A maintainer got lost finding routing info. Make aidd_docs/memory/vcs.md the single canonical home: one dense table (kind, prefix, commit type, triage label, next/main) plus a strict routing rule. CONTRIBUTING and RELEASE now link it instead of restating it; the inline label table is gone. Trim labels 11 -> 8 (drop help wanted, npm, github-actions) and stop dependabot re-adding the dropped ecosystem labels. Add a Project 8 board playbook + label-delete note to MAINTAINERS, and fix its stale "Work Type" reference. Routing is derived from the branch prefix, never a label or board field. Spec + plan recorded under aidd_docs/. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(aidd-vcs): derive PR base from branch prefix, auto-apply triage label The pull-request action resolved the base from origin/HEAD against a candidate list (main/master/develop/staging) that omitted `next`, so a `feat/*` PR targeted `main`. Resolve the base from the head branch's prefix via the project-memory routing table first, falling back to the old detection. After opening, apply the prefix's triage label when it exists. Stays generic: reads project memory, never hardcodes `next`. Also fix the stale `aidd_docs/` -> `docs/` prefix in the branch template. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs(framework): make the routing table an actionable guide Reframe the canonical Types table as "I want to… → I do…": add an Issue template column and a one-line board Status flow, so a contributor finds their row and reads off branch, label, and PR target in one glance. Still the single source of truth — no new table, no duplication. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(aidd-vcs): sync pull-request skill contract with prefix routing (#326) The skill's transversal rules still described base detection from repo state only, omitting the prefix-based routing and auto-labelling the action now performs. Bring the SKILL.md contract in line: base resolves from the branch prefix via project memory (fallback to repo state), and the matching triage label is applied after opening. Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(aidd-refine): align skills with skill contract (#327) * refactor(aidd-refine): align 02-05 skills with skill contract Bring challenge, condense, shadow-areas and fact-check skills to the 04-skill-generate contract and concise, plain language. - SKILL.md routers: `## Actions` header, plain-path references/assets, drop empty None. sections - actions: singular Input/Output, no pipeline-variable notation, bold process labels, no frozen data blocks - move challenge report skeleton to a new assets/report-template.md - extract fact-check output-discipline into a cited reference - shorten the four skill descriptions to two lines Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(aidd-refine): tighten tests and trim duplication Follow-up from a headless skill-by-skill audit against the 04-skill-generate contract. - shadow-areas: collapse redundant Test bullets (6/8/6 -> 5/7/4), one-line Outputs, fix 01-detect handoff (route via 03-diff before 02-render-report when a prior report exists) - challenge: fix Test deal-breaker/confidence rule to match the rubric, drop transversal rules that restated the action - condense: drop persistence-duplicate process step, tighten tests, lean the router (no inline examples) - fact-check: shorten 03-report Output, dedupe caching guidance Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * style(aidd-refine): drop punctuation dashes for concise prose Replace em-dash-style " - " punctuation with commas, colons, or parentheses across skill bodies, actions, and references. Keep title headers and the literal "(unverified - no source found)" output marker. Also fix severity-rubric "Definition rule" -> "Decision rule" label. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(aidd-refine): harden condense and fact-check fidelity Found by the behavioral eval harness running the skills for real. - fact-check: add a `refuted` verdict (a source contradicts the claim, distinct from "no source found"); emit the unverified marker verbatim; forbid leaking the verification method (commands, "by inspection"); remove "cascade exhausted ..." from the report template, which leaked a forbidden word the discipline bans - condense: 01-condense must emit the exact `Condense: ON (<level>).` line, since 02-stats and the hook parse it from the transcript * test(aidd-refine): add behavioral skill-eval harness Runs each refine skill for real through a headless `claude -p` in an isolated temp project (the worktree skill installed under a unique name, so the local copy runs, never a stale global plugin). Deterministic checks by default; `--judge` adds an LLM grader for fuzzy outcomes. Local / opt-in only (metered, no `claude` in CI), so not a CI gate. Covers shadow-areas, fact-check, condense, challenge; brainstorm is interactive and out of scope. --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(framework): executor/checker agents + SDLC orchestration redesign (#314) Collapse the SDLC agent trio (planner/implementer/reviewer) into two leaf agents, executor + checker, with the 00-sdlc orchestrator owning planning itself. Enforce recipe-not-spawn: only the orchestrator spawns, recipe skills focus on themselves and never name other skills or the SDLC. - executor runs 02-implement, checker runs 05-review (embedded checklist). - executor never holds 01-plan (structural immutability). - 05-review gains a relevancy axis (fit/conform/rot) into one review.md. - commit, pull-request, and assert defer to project memory (vcs.md, coding-assertions.md) over bundled defaults; ship stops dictating the commit format. - pull-request is fully VCS-tool-agnostic. - all flow skills renormalized to the skill-authoring contract; orphan assets and a leaked user-stories WIP duplicate removed. Net +673/-1269. Refs #253. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * feat(aidd-dev): route deletions to refactor cleanup with orphan sweep (#328) When the user asks to delete or remove code, the refactor skill is now a trigger target and routes straight to the cleanup action. The cleanup action sweeps for orphaned references a deletion leaves behind, so removals also clean up imports, tests, docs, and config that pointed at the deleted symbol. Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs(aidd-context): unify and trim agent instruction template (#329) * docs(aidd-context): unify and trim agent instruction template - Merge best rules from CLAUDE.md and the project-memory AGENTS.md template into one focused set - Cut aspirational/duplicate lines (goal-driven, redundant simplicity/anti-sycophancy/sureness restatements) - Keep high-impact rules: evidence-over-assertion, quote-shortest-line, dedup meta-rule, intention-revealing names - Align root CLAUDE.md: reference parent for shared behavior, keep framework specifics Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): sync CLAUDE.md with agent template, tighten rules - Make CLAUDE.md a true instance of the project-memory AGENTS.md template (same rules and structure; only title, placement bullet, and filled memory block differ) - Push CLAUDE.md-only rules up into the template: surface tradeoffs, solve own issues first, check it is good practice, less-is-more folded into Communication - Add focus rule: stay focused, not scattered (keeps freedom, no scope-creep framing) - Drop the anthropomorphize line; ban em-dashes alongside emoji Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): apply reviewer fixes to agent template rules - Dedup anti-sycophancy (3 bullets into 1) and brevity (2 into 1) - Add output-token rule: no preamble or recap - Operationalize focus: flag unrelated issues, detour only if blocking - Clarify the good-practice check Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): swap dedup meta-rule for a no-guessing rule - Add: don't guess APIs/signatures/flags/behavior, read source or docs first (prevents hallucination, the top dumb-failure mode) - Remove: the dedup meta-rule (off the runtime objectives; belongs in a standards doc) - Keep intention-revealing names and the first-message greeting Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): scan-first response style and clarify-scope rule - Rewrite brevity rule: telegraphic scan-first style for chat (arrows, fragments), normal prose for docs, full prose for security/irreversible/order-dependent - Extend no-preamble rule to ban trailing next-step padding - Add: ask one sharp question to pin down scope before an ambiguous or expensive build Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): add parallel-execution rules to agent template - Batch independent operations in one pass, not one at a time (full test suite, not test by test) - Fan out independent subtasks to parallel subagents Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): fix rule tensions and add next-step output rule - Merge simple-beats-clever into surgical changes - Make exceeding the ask an explicit exception, not the default - Gate subagent fan-out to genuinely large or parallelizable work - Broaden the scan-first clarity exception to any nuanced explanation - Reframe output rule: no suggestion menus, but always state the single next action Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): restore dedup rule, soften fan-out to orchestrators - Restore the single-source-of-truth rule removed earlier (downstream projects keep it) - Soften fan-out: only when you own the overall flow, per docs/ARCHITECTURE.md spawning boundary - Drop the test-suite example from the batch rule Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): bold rule heads and tighten wording - Bold the lead directive of each rule for scannability - Trim wording where it loses no meaning (notably the dedup rule) - Memory block unchanged (tooling-managed, must stay live) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): clarify escalation, knowledge, and tradeoff rules - Solve-your-own-issues: try yourself first, escalate to the human (names the target) - Drop the vague good-practice clause; keep don't-assume-knowledge-is-current - Surface tradeoffs and evaluate their impact (replaces vague confusion) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): allow commit/push inside authorizing flows Agents running inside an authorizing flow (e.g. the SDLC) may commit/push; the default still blocks unprompted commits. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): simplify commit rule exception for allowed agents Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): revert commit rule to simple user-asks default Drop the vague allowed-to escape hatch; authorized flows grant their own agents in their own definitions. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> * refactor: conform remaining skills to the authoring contract (#334) * refactor(dev): bring 04-audit to the skill-authoring contract - SKILL.md: rename "Available actions" to "Actions", fold the routing and output-contract prose into concise transversal rules, drop the redundant action-include list and the `->` arrows, and trim the description so it no longer names sibling skills. - 7 pillar actions: singular prose Input/Output (no frozen YAML), one-word bold step labels, bullet-list Test, and no cross-skill handoff naming (the audit reports, it does not route fixes). - README: drop the dead `impeccable` skill reference. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(dev): bring 06-test to the skill-authoring contract - SKILL.md: add what+when+Do-NOT triggers to the description, rename "Available actions" to "Actions", fold routing into a flow line, drop the redundant include list. - actions: singular prose Input/Output (no frozen YAML), one-word bold step labels, bullet Test. Remove the "spawn a sub-agent" instruction from test-journey (a recipe never spawns). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(dev): bring 07-refactor to the skill-authoring contract - SKILL.md: rename "Available actions" to "Actions", fold routing, audit-handoff, and conventions into transversal rules, drop the redundant include list and the description's em-dash and sibling-skill names. Keep the push-not-pull audit-fed rule intact. - 4 axis actions: singular prose Input/Output (no frozen YAML), one-word bold step labels, bullet Test, and no cross-skill naming (architecture recommends planning without naming the plan skill). - README: drop the dead `impeccable` skill reference. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(dev): bring 08-debug to the skill-authoring contract - SKILL.md: add what+when+Do-NOT triggers, rename "Available actions" to "Actions", fold routing into a flow line, drop the redundant include list, surface the asset and reference. - 3 actions: singular prose Input/Output (no frozen YAML), one-word bold step labels, bullet Test. Fix the same-skill reference slug (reflect_issue to reflect-issue) and the stray prose arrow. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(dev): bring 09-for-sure to the skill-authoring contract - SKILL.md: replace the Context/Environment/Execution-flow prose with an Actions table and transversal rules, add a Do-NOT clause, drop the empty "Environment: None" section and the redundant include block, and label the template as an asset, not a reference. - 3 actions: singular prose Input/Output (no frozen YAML), one-word bold step labels, bullet Test, prose arrows fixed. - Extract action 03's inline worker-prompt to assets/ and its log-entry format to references/ so the action holds pure anatomy. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs(dev): regenerate catalog after skill refactors * refactor(pm): bring ticket-info, user-stories-create, prd to the contract - SKILL.md x3: rename "Available actions" to "Actions", drop the empty References/Assets/External-data placeholder sections (R9). - actions x3: singular prose Input/Output (no frozen YAML), one-word bold step labels, no prose arrows, bullet Test, and correct `@../assets/` citation paths. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(vcs): bring repo-init, release-tag, issue-create to the contract - SKILL.md x3: rename "Available actions" to "Actions", drop empty References/External-data placeholder sections, fix the repo-init description (colon not dash, no sibling-skill names), correct asset citation paths to `@assets/`. - release-tag and issue-create actions: singular prose Input/Output (no frozen YAML), one-word bold step labels, no prose arrows, bullet Test, `@../assets/` citation paths. repo-init actions were already conformant. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(orchestrator): hygiene-norm 00-async-dev Remove every em-dash (headers to colons, prose to commas), rename the plural `## Inputs`/`## Outputs` action headers to singular, and fix the stray prose arrows. The skill's router, three sub-flow tables, and 21 nested actions were already structurally sound, so this is a hygiene pass, not a body rewrite. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(orchestrator): convert remaining mapping arrows to → in async-dev references * refactor(context,ui,refine): norm bootstrap, hello, condense - 01-bootstrap: rename "Available actions" to "Actions", drop the External-data section that named another skill's path, fix the body dash. Rewrite the 5 actions to the anatomy: singular prose Input/Output (no frozen example blocks), one-word bold step labels, bullet Test, no "Depends on" section. The candidate-audit spawn is preserved as-is. - 01-hello: add the missing title and `## Test`, rename the header, drop the empty `## Inputs - none`. - 03-condense: fix the `@references/` citation, drop the non-norm External-data section (the hook is already covered in the stats action), bullet the stats Test. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor: clear residual contract nits across skills - async-dev: drop the non-norm "## Depends on" sections from all 18 nested actions (the SKILL flow already states the sequence). Includes a linter pass normalizing em-dashes to `--`. - 10-todo: rename the action's plural Input/Output headers to singular. - 12-cook: drop the em-dash separators in the recipe template asset. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor: use plain paths in SKILL.md, no @ (per skill-template) The canonical scaffold prescribes plain paths in the SKILL.md References/Assets sections ("Plain paths, no @"); @ includes belong only in action files, which load them. Strip @ from every SKILL.md path listing and citation across all skills, including the flow skills merged earlier. The only @ left are the literal `@claude` webhook triggers in async-dev. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor: rewrite skill descriptions to the skill-generate norm Descriptions now follow the contract shape: a verb-led what-clause, then "Use when the user wants to <natural intents>", then a plain "Not for <overlap>". Drop every reference to another skill and every /command token (slash commands are tool-native), and fix the semantics that were wrong, notably the PRD, which is generated from a need, idea, or brainstorm and is upstream of user stories, not derived from them. Covers the refactored skills plus the gold-standard and merged skills that carried sibling references or slash tokens (onboard, project-memory, brainstorm, shadow-areas, fact-check, todo, commit, pull-request). Also collapse the few multi-word action step labels left from the prior pass to the one-word bold lead the anatomy requires. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(pm): make the PRD template solution-agnostic Headless testing surfaced a contradiction: the bundled prd-template.md shipped 15 sections including Technical Architecture, Tech Stack, and Data Model, which directly violate the skill's "never include technical implementation detail" rule and the action's own 8-section spec. The action's Test also checked for headings the template did not have. Rewrite the template to the action's eight solution-agnostic sections (Overview, Problem Statement, Goals, Non-Goals, User Stories, Acceptance Criteria, Dependencies, Open Questions) with an HTML-comment header so no frontmatter leaks into output, and tighten the action Test to assert those headings and the absence of any solution detail. Re-verified headless: 8 sections present, zero solution detail. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(context): add missing ## Output to cook 02-upsert action * docs(context): add a description size rule to the skill contract (R5) skill-generate produced oversized descriptions every run because R5 told it to "over-list" triggers with only a 1024-char ceiling and no size target. Add the missing rule: two lines max, ~240 chars, straight to the point, length serves recall not completeness. Also fix R5 to match the norm already enforced: lead with a verb, state triggers as "Use when the user wants to ...", never name another skill, never write a /command token (slash commands are tool-native). Update the template placeholder to match. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor: trim 22 skill descriptions to the 2-line size rule Apply the new R5 size rule to every description over the limit: two lines, ~240 chars, straight to the point, same shape (verb-led, "Use when the user wants to ...", plain "Not for ..."). Also drop the last two sibling-skill mentions (onboard/explore) so no description names another skill. 02-user-stories is trimmed in the working tree and left for its own commit. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(context,refine): close 3 defects found by the headless suite Generic and concise, never naming a specific tool or session mode: - explore: detect a tool only by its own mapped surfaces, never a shared parent directory, so a bare CI folder no longer counts as a tool. Add the presence-signal rule to ai-mapping and key survey step 1 to it. - onboard: tighten the silent-read rule so action 01 emits nothing even in a single turn; the first words come from orient. - fact-check: harden the report scrub into a line-by-line final gate and drop the named output-mode list (no "terse/caveman/condensed") in favor of "any active output mode". Re-tested headless: explore and onboard hold; fact-check's wording is clean, though a forced session output mode can still bleed a summary line at runtime, which is a host-mode interaction, not a skill gap. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(vcs,dev): split commit/pull-request into step actions, fix audit folder Model the real workflow as actions instead of one monolithic action: - commit: collect (stage one concern) → message → commit (push when asked). Drop the retry loop: a hook that rejects the commit is not this skill's job, it reports which hook and why and stops, only re-staging files a hook auto-formatted. - pull-request: collect (resolve base, gather commits) → draft (title and body) → create (open draft, label). Conventions live in transversal rules. - audit: one dated folder per run (aidd_docs/tasks/<yyyy_mm>/ <yyyy_mm_dd>_audit/, like a feature folder) defined once in a transversal rule; each pillar action names only its file (architecture.md, security.md, ...); a full run adds report.md. Headless-tested: commit chain produces a conventional commit; a failing pre-commit hook reports and stops with no commit and no loop; audit writes the pillar file into the dated folder. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(vcs): keep the VCS convention in one transversal rule, no baked example Drop the project-specific `feat/ → next` example from the pull-request skill and de-duplicate the convention source: `vcs.md` now appears only in the SKILL transversal rule (where + fallback), and the actions just apply it ("per the project's convention"). The concrete branch examples stay in the bundled fallback asset (branch.md), where examples belong. The skill reads the project's convention from memory instead of assuming one, so behavior is project-driven, not hardcoded. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(dev): audit full run writes every pillar file plus report.md The rule was ambiguous, so a full run wrote only the merged report. Make it explicit: every pillar that runs always writes its own <pillar>.md, and a full run additionally writes report.md. Re-tested headless: a full run produced all 7 pillar files plus report.md. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor: quick conciseness wins across skills Two mechanical passes against the AI-template feel, no behavior change: - Drop the filler "## Default flow" sections: delete the pure single-action ones ("the router dispatches to X whenever a phrase appears"), and inline the real-flow ones as one tight line under the actions table. - Strip obvious embedded git commands from action prose ("the current branch" not the rev-parse incantation, "the remote URL" not the remote-get-url one), keeping the runnable ones in ## Test. The onboard read-project step also loses its hardcoded .github path. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(vcs,dev): de-duplicate what the transversal rule already owns Actions were restating their skill's transversal rule: - repo-init: both Resolve steps repeated the whole "from VCS memory, else the environment, else main" resolution. They now just say "resolve the default branch and provider" / "resolve the host"; the transversal rule owns how. - release-tag: the bump step dropped "only" (the transversal already scopes the bump commit to version-manager files). - assert: dropped the Boundary step that repeated "stop only when the final sweep passes." No behavior change: each convention has one home, the action defers. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(dev): restore assert Boundary step (loop stop belongs in the process) The de-dup pass dropped it as a duplicate of the transversal stop rule, but a fix-loop's termination is part of the process, not just a passive convention. Keep it explicit where the loop decision happens. * refactor: apply four cross-skill principles (args, heavy steps, guards, plain words) - Tool-agnostic input: `$ARGUMENTS` (a Claude-only token other hosts won't read) becomes the plain word "the arguments" across 16 skills. command-generate keeps it, since the token is what it teaches. - One job per step: split plan's overloaded `Write` step into Folder / Fill / Show. - Sequence vs guards: implement's `02-execute` separates the Open / Code / Assert loop from the Blocked and Drift stop conditions, which are no longer numbered steps. - Plain words: implement's `Branch` drops the embedded `git checkout -b` for "create a feature branch". Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor: lighten the remaining heavy steps (one job per step) Split learn's Ask into Show + Decide and ship's Gate into Gate + Freshness; tighten sdlc's Iterate and refactor's architecture Source. No heavy (>320-char) steps remain. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(dev): separate guards from sequence in 06-test - test-journey: the Failure handling is part of the Walk, not a peer step; fold it under Walk as a sub-bullet. - test: drop the Boundary step that restated the transversal rules (functional only, never compromise quality); one-at-a-time stays implicit in the generate-then-next loop. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(dev): make execute stop-conditions a Guard step with sub-bullets The free-floating 'Stop the loop on either condition' block diverged from the action anatomy. Express it as a numbered Guard step whose conditions are sub-bullets (the anatomy's form for a branch), keeping Blocked and Drift as bullets, not peer steps. Left in the action, not the transversal rules: these guard the execute loop only. * fix(dev): review must not assume 'main' as the diff base All three review actions hardcoded `git diff main`. Resolve the diff against the repository's default branch instead, never a blind main, matching the pull-request skill's base handling. * fix(dev,vcs): two blockers the deep audit caught - for-sure: the spawned-worker prompt had no payment/destructive carve-out, and the loop never consumed the worker's stop status, so a worker that halted on a money gate was retried, re-triggering the spend. Add the carve-out to the worker prompt and make the loop surface a gate stop and halt instead of retrying. - release-tag: a tagless repo fell back to a fictitious v1.0.0 and then ran `git log v1.0.0..HEAD`, which aborts. With no prior tag, list every commit on the branch instead. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix: clear semantic defects from the deep audit (batch 1) - prd task-template: drop the stray brace in "## Main step 1}". - command-generate README: correct the stale "# 10" heading to "# 07". - issue-create template: "GitHub issues" to "issues in the configured tracker" (the skill is tool-agnostic). - 01-plan SKILL: drop the dangling reference to a mermaid-conventions.md that does not exist in the skill and no action uses. - commit and pull-request READMEs: replace the stale "single action" description with the three-action chain each skill now has. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix: semantic defects from the deep audit (batch 2) Mostly Output/Test mismatches where the Test under-verified or contradicted the declared Output, plus a few wording bugs: - sync Tests to the full Output (ticket-info six fields, cook README index row, debug flowchart + confidence, test-journey full fields). - audit: rename the template header to "## Top actions" so all seven pillar Tests match it. - spec: assert "every required section" (the validator marks some optional), matching the Process. - fact-check: re-render restyled lines in plain prose instead of deleting them (deleting lost content). - plan phase-template: legend now reads deleted/created/modified, matching the upstream modify/create/delete vocabulary. - todo: drop "using template" (no such asset); learn-sync names the updated files (its Test wants names, not a count); agent-generate Clarify now asks the model (it is a declared output); test marks a declined behavior pending; refactor allows a security-fix regression test. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(pm): replace user-stories-create with the six-action user-stories skill Rename and rebuild: 02-user-stories-create (one monolithic action) is replaced by 02-user-stories, a six-action chain (clarify-scope, split-epic, draft-stories, estimate-impact, prioritize, sync-tracker) with INVEST, a Definition of Ready and a functional Definition of Done. Update the plugin manifest, catalogs, marketplace, READMEs, and the upgrade map. Applies its two audit findings: estimate-impact now lists dependencies in its Output, and clarify-scope no longer restates the transversal lean-clarification rule. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs: memory-deduplication diagnosis and dedup plan Add the memory-duplication diagnosis and the memory-dedup feature folder (plan, phases, empirical results), plus the two one-line doc touches that came with that work. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(context): repair the garbled scaffold-docs action label The action-table cell had become 'Create the `folder structure' with an unclosed backtick and the lost aidd_docs/ reference. Restore it to 'Create the `aidd_docs/` folder structure'. * fix(skills): resolve semantic contradictions from deep audit Align action contracts with their stated outputs, tests, and SKILL rules: - onboard: orient stops at offer+capture; 03-act owns acting on every choice - sdlc ship: implement the promised on_default_branch gate (commit nothing) - skill-generate: contract is R1-R13; draft output covers plugin-source mode - agent-generate: capture picks one name, output/test follow - bootstrap: pick a candidate by name; fold cost check into the rationale - hook-generate: scope tool-paths to commonly supported moments, no gap promise - challenge: drop orphan report section; make the 100% confidence tier exclusive - shadow-areas: symmetric dispatch; template renders only categories with gaps - assert-frontend: output schema matches the candidate-causes step - pull-request README: list only the assets that ship Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * refactor(skills): drop host tokens, make two descriptions verb-led and plain - review, mermaid: verb-led descriptions, plain "Not for", under the limit - onboard, commit, condense, implement: drop host-specific slash and $ARGUMENTS tokens from instruction prose for plain capability language Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * refactor(skills): drop passive restatements the transversal already owns Remove trailing reminder clauses that echo a SKILL transversal rule while the step's primary action stays intact: explore (skip-surface, no-advice, no-assume), debug (no drive-by refactors), mermaid (no unconfirmed node), learn (preserve edits), review (never patch, x2), assert-architecture (never fix), spec (solution-agnostic). Load-bearing guards, loop terminators, and point-of-action safety constraints are kept. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * refactor(spec): keep the behavioral half of the build-step guard The transversal owns the artifact property (solution-agnostic), but not the agent action: restore "Do not explore the codebase." so the step still forbids reading the code, dropping only the redundant property restatement. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * refactor(onboard): cut meta-commentary from orient step 7 The step instructs; explaining that 03-act delegates here is noise. The handoff is already documented from 03-act's side. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(skills): repair structural anatomy bugs from deep audit - cook: asset ref needs the parent path (@../assets) to resolve from actions/ - agent-generate README: header number 06, not 09 - command-generate tool-paths: drop the duplicated table header row - debug task-template: remove stray brace in "## Main step 1" - todo: dangling colon ends the Report step as a sentence - for-sure tracking template: drop skill-metadata frontmatter that leaked name/description/argument-hint into every generated plan file Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(rule-generate): correct sibling @-path in references @-paths are file-relative (actions use @../references/...), so from inside references/ the sibling is @rule-authoring.md, not @references/... which resolved to references/references/. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(skills): repair four defects surfaced by the headless behavioral test - agent-generate: Codex TOML rule mandated single-quote literals, which cannot hold an apostrophe; use a double-quoted basic string with escaping - async-dev: actions 03/04/05 read assets/<t> but the templates live at assets/setup/<t>; restore the setup/ segment so the Read resolves - shadow-areas: empty source stopped with an uncategorizable blocker gap that broke the locked-category rule; stop with a plain warning, no gap - todo: the mandated per-executor refine resolved to an interactive-only skill that blocks in autonomous subagents; fall back to inline refine Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(skills): wire onboard hand-off, drop unused issue type field - onboard: orient now offers "run it in a fresh session instead", so 03-act's Hand off outcome is reachable instead of a dead branch - issue-create: drop the type field from Input/Output/Validate; Create never applied it and gh issue create has no native type flag Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(docs): repair broken markdown links failing CI Validate - regenerate plugin CATALOGs: drop the deleted 05-pull-request-list section (aidd-vcs) and the stale 01-plan mermaid-conventions row (aidd-dev) - task scratch docs: the @-build illustration used literal [x](../x) link syntax the checker resolved as a path; reword to prose Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(aidd-context): dedupe project-memory templates and lighten actions (#347) Memory generation restated the same fact across files (architecture vs codebase-map, integration internal flow, stack libs) because templates overlapped and the fill ran blind in parallel. Reframe ownership so each fact has one home and others reference it. Templates - architecture: drop Structure (codebase-map owns areas + entry point); Stack lists cross-cutting libs only, a capability's lib lives in its file - integration: reframe to external/third-party services (drop Internal, which duplicated architecture's flow) - api/navigation: scope routes to server vs client Actions / rules - 03: fill is no longer blind in parallel; each template knows what earlier ones captured - SKILL: split rules into Memory rules (content) and Action rules (process); add "one fact, one file" with the define-here / reference-elsewhere nuance - 04: semantic dedup review step (no fact's definition duplicated) - new references/memory-block.md; drop capability-signals' redundant section Validated headless on three stacks (cli, kairos web/mobile, kairos existing): definition-level duplication gone, files stay self-contained. Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> * docs(onboard): brainstorm + validated plan for the onboard refactor Brainstorm (guide that enforces a skippable foundation, then a project- adapted menu) and a 4-phase plan, hardened across three fixture-matrix dry-runs: flow-walk owns the default, signal map ranks secondary tools, read-once with a session ledger that tracks done/skipped steps. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * feat(onboard): phase 1 — read-once loop with a session ledger Replace re-read-every-loop with a single reusable snapshot plus a session ledger of run/skipped steps. 01 produces the snapshot (cheap signals + a bounded richer read) and the ledger; 03 writes it (run → done, different step → skipped); 02 consumes it (a stage is met by a disk fact OR the ledger), so a just-run or declined step is never re-suggested even when it left no file behind. journey.md points to 02's met-definition, no restatement. Behaviorally verified: empty-repo re-nag, off-disk Track completion, and the skip path all advance instead of looping. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * feat(onboard): phase 2 — foundation gate, skippable 02-orient marks a foundation stage as the loud pre-selected default with an explicit branch — architect a stack on an empty repo, set up the memory bank on a content repo with weak memory — rendered as the first choice tagged (recommended), always skippable. SKILL keeps suggest-never-force with the skippable + skip-remembered note. 01 pins that a docs/README-only repo counts as empty. Behaviorally verified: empty→architect, weak-memory→memory, skip excludes the default from the next loop (no re-nag), and the gate never blocks. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * feat(onboard): phases 3-4 — capability map + project-adapted menu Push detail into references/assets so the actions stay one line: - references/signals.md: the signal catalogue (cheap + a bounded richer read: code-quality sample, bug-marker scan) and the capability map (signal → stage-gated secondary tool) - assets/menu.md: the menu shape — default, ranked secondary tools (≤~5), a never-dropped footer, need-stage and end-stage (review-never-skipped) rules - 02-orient: place → rank → resolve → brief → assemble, one line each - journey.md: cumulative-stage placement made explicit (a plan implies the need is clarified/tracked); open-PR row reviews before ship; tool ranking deferred to signals.md, no duplication Behaviorally verified: mid-build → Build alone + add-tests/audit/debug; open-PR → Review then Ship; rough idea → no loud default, Clarify first; empty/code-no-memory foundation + ledger advance regress clean. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(onboard): refresh README for the snapshot+ledger+menu flow Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * refactor(onboard): slim 03-act, drop colon from the description - move the per-choice outcome table into references/outcomes.md; 03-act becomes a thin act + ledger dispatch (722 -> 171 words) - description no longer uses a colon Verified live (headless claude -p): the menu and the "explain the step" outcome render correctly from the new reference. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * refactor(skills): forbid colon and em dash in descriptions, codify in R5 Rewrite the four descriptions that used a colon (cook, assert, async-dev, repo-init) into two sentences, and add the rule to R5 in skill-authoring.md so it holds going forward. assert also drops "Do NOT use to" for plain "Not for". Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * chore: back-merge main into next (v5.0.3 drift) (#359) * ci(deps): bump actions/checkout from 6.0.3 to 7.0.0 (#321) Bumps [actions/checkout](https://github.com/actions/checkout) from 6.0.3 to 7.0.0. - [Release notes](https://github.com/actions/checkout/releases) - [Changelog](https://github.com/actions/checkout/blob/main/CHANGELOG.md) - [Commits](actions/checkout@df4cb1c...9c091bb) --- updated-dependencies: - dependency-name: actions/checkout dependency-version: 7.0.0 dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * chore(deps-dev): bump js-yaml from 4.2.0 to 5.0.0 (#322) * chore(deps-dev): bump js-yaml from 4.2.0 to 5.0.0 Bumps [js-yaml](https://github.com/nodeca/js-yaml) from 4.2.0 to 5.0.0. - [Changelog](https://github.com/nodeca/js-yaml/blob/master/CHANGELOG.md) - [Commits](nodeca/js-yaml@4.2.0...5.0.0) --- updated-dependencies: - dependency-name: js-yaml dependency-version: 5.0.0 dependency-type: direct:development update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <support@github.com> * fix(framework): use js-yaml v5 named export in validate-yaml js-yaml 5.0.0 is pure ESM and drops the default export. Switch to the named { load } import so the yaml-validity hook stops throwing the missing default export error. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Baptiste LAFOURCADE <baptiste.lafourcade@gmail.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * chore: release main (#323) Co-authored-by: aidd-bot[bot] <290648294+aidd-bot[bot]@users.noreply.github.com> --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: aidd-bot[bot] <290648294+aidd-bot[bot]@users.noreply.github.com> --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: Alex <8973343+alexsoyes@users.noreply.github.com> Co-authored-by: aidd-bot[bot] <aidd-bot[bot]@users.noreply.github.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: aidd-bot[bot] <290648294+aidd-bot[bot]@users.noreply.github.com>

…esign (#314) Collapse the SDLC agent trio (planner/implementer/reviewer) into two leaf agents, executor + checker, with the 00-sdlc orchestrator owning planning itself. Enforce recipe-not-spawn: only the orchestrator spawns, recipe skills focus on themselves and never name other skills or the SDLC. - executor runs 02-implement, checker runs 05-review (embedded checklist). - executor never holds 01-plan (structural immutability). - 05-review gains a relevancy axis (fit/conform/rot) into one review.md. - commit, pull-request, and assert defer to project memory (vcs.md, coding-assertions.md) over bundled defaults; ship stops dictating the commit format. - pull-request is fully VCS-tool-agnostic. - all flow skills renormalized to the skill-authoring contract; orphan assets and a leaked user-stories WIP duplicate removed. Net +673/-1269. Refs #253. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* ci(deps): bump actions/checkout from 6.0.3 to 7.0.0 (#321) Bumps [actions/checkout](https://github.com/actions/checkout) from 6.0.3 to 7.0.0. - [Release notes](https://github.com/actions/checkout/releases) - [Changelog](https://github.com/actions/checkout/blob/main/CHANGELOG.md) - [Commits](actions/checkout@df4cb1c...9c091bb) --- updated-dependencies: - dependency-name: actions/checkout dependency-version: 7.0.0 dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * chore(deps-dev): bump js-yaml from 4.2.0 to 5.0.0 (#322) * chore(deps-dev): bump js-yaml from 4.2.0 to 5.0.0 Bumps [js-yaml](https://github.com/nodeca/js-yaml) from 4.2.0 to 5.0.0. - [Changelog](https://github.com/nodeca/js-yaml/blob/master/CHANGELOG.md) - [Commits](nodeca/js-yaml@4.2.0...5.0.0) --- updated-dependencies: - dependency-name: js-yaml dependency-version: 5.0.0 dependency-type: direct:development update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <support@github.com> * fix(framework): use js-yaml v5 named export in validate-yaml js-yaml 5.0.0 is pure ESM and drops the default export. Switch to the named { load } import so the yaml-validity hook stops throwing the missing default export error. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Baptiste LAFOURCADE <baptiste.lafourcade@gmail.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * chore: release main (#323) Co-authored-by: aidd-bot[bot] <290648294+aidd-bot[bot]@users.noreply.github.com> * feat(skills): sync argument hints (#296) * refactor(aidd-dev): redesign 01-plan into gather/explore/wireframe/plan (#271) Four-action plan skill (gather/explore/wireframe/plan), unified feature-folder artifact layout (brainstorm/spec/plan/phases/review), implement reads the new layout with per-plan and per-phase status tracking. Closes #292 Closes #265 Closes #276 * docs(framework): add dominance checks to review and rules (#313) * docs(framework): add dominance checks to review and rules Dominance issues were easy for agents to miss when a weaker rule, criterion, or documentation statement was already covered, overridden, or invalidated by a stronger element. This adds the check at the global context, rule-generation, reviewer, and review-template levels. * docs(framework): reduce dominance check duplication Keep the concise dominance rule in global and rule-authoring surfaces while leaving the detailed case breakdown to the reviewer. Validation now states the shared check without reimplementing reviewer classification. * docs(framework): separate global and rule dominance checks Keep the general dominance reminder and reviewer taxonomy separate from rule-generation specifics. Rule authoring and validation now only check for duplicate, weaker, or contradictory rule bullets. * docs(framework): prefer intention-revealing names Naming guidance belongs in the global project context because it applies across artifacts. This keeps the rule concise and avoids duplicating it in domain-specific rule generation flows. * docs(aidd-context): validate rules against shared contracts Rule validation should apply the rule-authoring and tool-path contracts instead of restating individual rule checks. The test now has a single outcome sentence rather than multiple bullet assertions. * docs(aidd-dev): align dominance wording with review role The reviewer should report parallel elements rather than describe adding behavior. This keeps the dominance check phrased for read-only review while preserving the same taxonomy. * ci(framework): trigger Validate on next branch (#317) Required check "lefthook (framework-local checks)" (next.json ruleset) is published only by validate.yml, whose pull_request/push filter was [main]. PRs targeting next never fired it, so the required status stayed "Expected — Waiting for status" forever, deadlocking merges (e.g. #307). Add next to both push and pull_request branch filters. Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * feat(framework): add markdown link checker (#307) * feat(framework): add markdown link checker * fix(scripts): exempt template scaffold placeholder links 01-plan's plan-template.md / phase-template.md link to ./plan.md and ./phase-N.md — files the skill emits next to the generated output at runtime (see actions/04-plan.md), never committed to the repo. The checker flagged them as broken. Extend the existing asset-template branch: a dot-relative target in a *-template.md that resolves nowhere is an intentional placeholder for a generated sibling, so treat it as ignored rather than broken. The assets/templates fallback (used by 02-project-memory) is unchanged. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Baptiste LAFOURCADE <baptiste.lafourcade@gmail.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * feat(aidd-ui): scaffold alpha plugin (0.1.0-alpha.0) (#319) * feat(aidd-ui): scaffold alpha plugin (0.1.0-alpha.0) Add aidd-ui as a new plugin with a single smoke-test skill (01-hello). Registered in marketplace.json (recommended: false) and release config; alpha status is conveyed by the 0.1.0-alpha.0 version in plugin.json. Lives on a dedicated branch off next, so it is not public until it graduates to main. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs(aidd-ui): mark alpha / not-ready across surfaces Add prominent alpha warnings to the plugin README, the homepage plugins table, and the marketplace and plugin descriptions so the not-ready status is unmistakable. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * chore(deps): target next instead of main for dependabot (#324) Dependabot had no target-branch, so its PRs defaulted to main. main only takes the weekly promotion PR and the auto-merged Release PR (RELEASE.md), so dep updates landing there cut off-cycle releases (js-yaml bump shipped v5.0.3 mid-week). Point both ecosystems at next so updates batch into the weekly release like all other work. Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * feat(aidd-vcs): add repo-init skill (init + publish a repository) (#269) New skill aidd-vcs:00-repo-init: init a project's repository (git init, default branch, CONTRIBUTING.md, bootstrap commit) and, on request, create the remote and push. Provider- and mechanism-agnostic: host + reach (CLI, MCP, or API) resolved from VCS memory or environment, main as default-branch fallback. No hardcoded provider or fixed mechanism. Conforms to the skill-authoring contract (R1-R13 + action anatomy). Co-authored-by: Baptiste LAFOURCADE <baptiste.lafourcade@gmail.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs(framework): unify change taxonomy into one source of truth (#325) * ci(deps): bump actions/checkout from 6.0.3 to 7.0.0 (#321) Bumps [actions/checkout](https://github.com/actions/checkout) from 6.0.3 to 7.0.0. - [Release notes](https://github.com/actions/checkout/releases) - [Changelog](https://github.com/actions/checkout/blob/main/CHANGELOG.md) - [Commits](actions/checkout@df4cb1c...9c091bb) --- updated-dependencies: - dependency-name: actions/checkout dependency-version: 7.0.0 dependency-type: direct:production update-type: version-update:semver-major ... Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> * docs(framework): unify change taxonomy into one source of truth The same change taxonomy (kind -> where it ships) was re-spelled across branch prefix, commit type, label, and the board "Work Type" field, and the four drifted. A maintainer got lost finding routing info. Make aidd_docs/memory/vcs.md the single canonical home: one dense table (kind, prefix, commit type, triage label, next/main) plus a strict routing rule. CONTRIBUTING and RELEASE now link it instead of restating it; the inline label table is gone. Trim labels 11 -> 8 (drop help wanted, npm, github-actions) and stop dependabot re-adding the dropped ecosystem labels. Add a Project 8 board playbook + label-delete note to MAINTAINERS, and fix its stale "Work Type" reference. Routing is derived from the branch prefix, never a label or board field. Spec + plan recorded under aidd_docs/. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(aidd-vcs): derive PR base from branch prefix, auto-apply triage label The pull-request action resolved the base from origin/HEAD against a candidate list (main/master/develop/staging) that omitted `next`, so a `feat/*` PR targeted `main`. Resolve the base from the head branch's prefix via the project-memory routing table first, falling back to the old detection. After opening, apply the prefix's triage label when it exists. Stays generic: reads project memory, never hardcodes `next`. Also fix the stale `aidd_docs/` -> `docs/` prefix in the branch template. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs(framework): make the routing table an actionable guide Reframe the canonical Types table as "I want to… → I do…": add an Issue template column and a one-line board Status flow, so a contributor finds their row and reads off branch, label, and PR target in one glance. Still the single source of truth — no new table, no duplication. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(aidd-vcs): sync pull-request skill contract with prefix routing (#326) The skill's transversal rules still described base detection from repo state only, omitting the prefix-based routing and auto-labelling the action now performs. Bring the SKILL.md contract in line: base resolves from the branch prefix via project memory (fallback to repo state), and the matching triage label is applied after opening. Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(aidd-refine): align skills with skill contract (#327) * refactor(aidd-refine): align 02-05 skills with skill contract Bring challenge, condense, shadow-areas and fact-check skills to the 04-skill-generate contract and concise, plain language. - SKILL.md routers: `## Actions` header, plain-path references/assets, drop empty None. sections - actions: singular Input/Output, no pipeline-variable notation, bold process labels, no frozen data blocks - move challenge report skeleton to a new assets/report-template.md - extract fact-check output-discipline into a cited reference - shorten the four skill descriptions to two lines Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(aidd-refine): tighten tests and trim duplication Follow-up from a headless skill-by-skill audit against the 04-skill-generate contract. - shadow-areas: collapse redundant Test bullets (6/8/6 -> 5/7/4), one-line Outputs, fix 01-detect handoff (route via 03-diff before 02-render-report when a prior report exists) - challenge: fix Test deal-breaker/confidence rule to match the rubric, drop transversal rules that restated the action - condense: drop persistence-duplicate process step, tighten tests, lean the router (no inline examples) - fact-check: shorten 03-report Output, dedupe caching guidance Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * style(aidd-refine): drop punctuation dashes for concise prose Replace em-dash-style " - " punctuation with commas, colons, or parentheses across skill bodies, actions, and references. Keep title headers and the literal "(unverified - no source found)" output marker. Also fix severity-rubric "Definition rule" -> "Decision rule" label. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(aidd-refine): harden condense and fact-check fidelity Found by the behavioral eval harness running the skills for real. - fact-check: add a `refuted` verdict (a source contradicts the claim, distinct from "no source found"); emit the unverified marker verbatim; forbid leaking the verification method (commands, "by inspection"); remove "cascade exhausted ..." from the report template, which leaked a forbidden word the discipline bans - condense: 01-condense must emit the exact `Condense: ON (<level>).` line, since 02-stats and the hook parse it from the transcript * test(aidd-refine): add behavioral skill-eval harness Runs each refine skill for real through a headless `claude -p` in an isolated temp project (the worktree skill installed under a unique name, so the local copy runs, never a stale global plugin). Deterministic checks by default; `--judge` adds an LLM grader for fuzzy outcomes. Local / opt-in only (metered, no `claude` in CI), so not a CI gate. Covers shadow-areas, fact-check, condense, challenge; brainstorm is interactive and out of scope. --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(framework): executor/checker agents + SDLC orchestration redesign (#314) Collapse the SDLC agent trio (planner/implementer/reviewer) into two leaf agents, executor + checker, with the 00-sdlc orchestrator owning planning itself. Enforce recipe-not-spawn: only the orchestrator spawns, recipe skills focus on themselves and never name other skills or the SDLC. - executor runs 02-implement, checker runs 05-review (embedded checklist). - executor never holds 01-plan (structural immutability). - 05-review gains a relevancy axis (fit/conform/rot) into one review.md. - commit, pull-request, and assert defer to project memory (vcs.md, coding-assertions.md) over bundled defaults; ship stops dictating the commit format. - pull-request is fully VCS-tool-agnostic. - all flow skills renormalized to the skill-authoring contract; orphan assets and a leaked user-stories WIP duplicate removed. Net +673/-1269. Refs #253. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * feat(aidd-dev): route deletions to refactor cleanup with orphan sweep (#328) When the user asks to delete or remove code, the refactor skill is now a trigger target and routes straight to the cleanup action. The cleanup action sweeps for orphaned references a deletion leaves behind, so removals also clean up imports, tests, docs, and config that pointed at the deleted symbol. Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs(aidd-context): unify and trim agent instruction template (#329) * docs(aidd-context): unify and trim agent instruction template - Merge best rules from CLAUDE.md and the project-memory AGENTS.md template into one focused set - Cut aspirational/duplicate lines (goal-driven, redundant simplicity/anti-sycophancy/sureness restatements) - Keep high-impact rules: evidence-over-assertion, quote-shortest-line, dedup meta-rule, intention-revealing names - Align root CLAUDE.md: reference parent for shared behavior, keep framework specifics Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): sync CLAUDE.md with agent template, tighten rules - Make CLAUDE.md a true instance of the project-memory AGENTS.md template (same rules and structure; only title, placement bullet, and filled memory block differ) - Push CLAUDE.md-only rules up into the template: surface tradeoffs, solve own issues first, check it is good practice, less-is-more folded into Communication - Add focus rule: stay focused, not scattered (keeps freedom, no scope-creep framing) - Drop the anthropomorphize line; ban em-dashes alongside emoji Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): apply reviewer fixes to agent template rules - Dedup anti-sycophancy (3 bullets into 1) and brevity (2 into 1) - Add output-token rule: no preamble or recap - Operationalize focus: flag unrelated issues, detour only if blocking - Clarify the good-practice check Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): swap dedup meta-rule for a no-guessing rule - Add: don't guess APIs/signatures/flags/behavior, read source or docs first (prevents hallucination, the top dumb-failure mode) - Remove: the dedup meta-rule (off the runtime objectives; belongs in a standards doc) - Keep intention-revealing names and the first-message greeting Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): scan-first response style and clarify-scope rule - Rewrite brevity rule: telegraphic scan-first style for chat (arrows, fragments), normal prose for docs, full prose for security/irreversible/order-dependent - Extend no-preamble rule to ban trailing next-step padding - Add: ask one sharp question to pin down scope before an ambiguous or expensive build Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): add parallel-execution rules to agent template - Batch independent operations in one pass, not one at a time (full test suite, not test by test) - Fan out independent subtasks to parallel subagents Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): fix rule tensions and add next-step output rule - Merge simple-beats-clever into surgical changes - Make exceeding the ask an explicit exception, not the default - Gate subagent fan-out to genuinely large or parallelizable work - Broaden the scan-first clarity exception to any nuanced explanation - Reframe output rule: no suggestion menus, but always state the single next action Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): restore dedup rule, soften fan-out to orchestrators - Restore the single-source-of-truth rule removed earlier (downstream projects keep it) - Soften fan-out: only when you own the overall flow, per docs/ARCHITECTURE.md spawning boundary - Drop the test-suite example from the batch rule Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): bold rule heads and tighten wording - Bold the lead directive of each rule for scannability - Trim wording where it loses no meaning (notably the dedup rule) - Memory block unchanged (tooling-managed, must stay live) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): clarify escalation, knowledge, and tradeoff rules - Solve-your-own-issues: try yourself first, escalate to the human (names the target) - Drop the vague good-practice clause; keep don't-assume-knowledge-is-current - Surface tradeoffs and evaluate their impact (replaces vague confusion) Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): allow commit/push inside authorizing flows Agents running inside an authorizing flow (e.g. the SDLC) may commit/push; the default still blocks unprompted commits. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): simplify commit rule exception for allowed agents Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(aidd-context): revert commit rule to simple user-asks default Drop the vague allowed-to escape hatch; authorized flows grant their own agents in their own definitions. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> * refactor: conform remaining skills to the authoring contract (#334) * refactor(dev): bring 04-audit to the skill-authoring contract - SKILL.md: rename "Available actions" to "Actions", fold the routing and output-contract prose into concise transversal rules, drop the redundant action-include list and the `->` arrows, and trim the description so it no longer names sibling skills. - 7 pillar actions: singular prose Input/Output (no frozen YAML), one-word bold step labels, bullet-list Test, and no cross-skill handoff naming (the audit reports, it does not route fixes). - README: drop the dead `impeccable` skill reference. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(dev): bring 06-test to the skill-authoring contract - SKILL.md: add what+when+Do-NOT triggers to the description, rename "Available actions" to "Actions", fold routing into a flow line, drop the redundant include list. - actions: singular prose Input/Output (no frozen YAML), one-word bold step labels, bullet Test. Remove the "spawn a sub-agent" instruction from test-journey (a recipe never spawns). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(dev): bring 07-refactor to the skill-authoring contract - SKILL.md: rename "Available actions" to "Actions", fold routing, audit-handoff, and conventions into transversal rules, drop the redundant include list and the description's em-dash and sibling-skill names. Keep the push-not-pull audit-fed rule intact. - 4 axis actions: singular prose Input/Output (no frozen YAML), one-word bold step labels, bullet Test, and no cross-skill naming (architecture recommends planning without naming the plan skill). - README: drop the dead `impeccable` skill reference. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(dev): bring 08-debug to the skill-authoring contract - SKILL.md: add what+when+Do-NOT triggers, rename "Available actions" to "Actions", fold routing into a flow line, drop the redundant include list, surface the asset and reference. - 3 actions: singular prose Input/Output (no frozen YAML), one-word bold step labels, bullet Test. Fix the same-skill reference slug (reflect_issue to reflect-issue) and the stray prose arrow. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(dev): bring 09-for-sure to the skill-authoring contract - SKILL.md: replace the Context/Environment/Execution-flow prose with an Actions table and transversal rules, add a Do-NOT clause, drop the empty "Environment: None" section and the redundant include block, and label the template as an asset, not a reference. - 3 actions: singular prose Input/Output (no frozen YAML), one-word bold step labels, bullet Test, prose arrows fixed. - Extract action 03's inline worker-prompt to assets/ and its log-entry format to references/ so the action holds pure anatomy. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs(dev): regenerate catalog after skill refactors * refactor(pm): bring ticket-info, user-stories-create, prd to the contract - SKILL.md x3: rename "Available actions" to "Actions", drop the empty References/Assets/External-data placeholder sections (R9). - actions x3: singular prose Input/Output (no frozen YAML), one-word bold step labels, no prose arrows, bullet Test, and correct `@../assets/` citation paths. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(vcs): bring repo-init, release-tag, issue-create to the contract - SKILL.md x3: rename "Available actions" to "Actions", drop empty References/External-data placeholder sections, fix the repo-init description (colon not dash, no sibling-skill names), correct asset citation paths to `@assets/`. - release-tag and issue-create actions: singular prose Input/Output (no frozen YAML), one-word bold step labels, no prose arrows, bullet Test, `@../assets/` citation paths. repo-init actions were already conformant. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(orchestrator): hygiene-norm 00-async-dev Remove every em-dash (headers to colons, prose to commas), rename the plural `## Inputs`/`## Outputs` action headers to singular, and fix the stray prose arrows. The skill's router, three sub-flow tables, and 21 nested actions were already structurally sound, so this is a hygiene pass, not a body rewrite. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(orchestrator): convert remaining mapping arrows to → in async-dev references * refactor(context,ui,refine): norm bootstrap, hello, condense - 01-bootstrap: rename "Available actions" to "Actions", drop the External-data section that named another skill's path, fix the body dash. Rewrite the 5 actions to the anatomy: singular prose Input/Output (no frozen example blocks), one-word bold step labels, bullet Test, no "Depends on" section. The candidate-audit spawn is preserved as-is. - 01-hello: add the missing title and `## Test`, rename the header, drop the empty `## Inputs - none`. - 03-condense: fix the `@references/` citation, drop the non-norm External-data section (the hook is already covered in the stats action), bullet the stats Test. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor: clear residual contract nits across skills - async-dev: drop the non-norm "## Depends on" sections from all 18 nested actions (the SKILL flow already states the sequence). Includes a linter pass normalizing em-dashes to `--`. - 10-todo: rename the action's plural Input/Output headers to singular. - 12-cook: drop the em-dash separators in the recipe template asset. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor: use plain paths in SKILL.md, no @ (per skill-template) The canonical scaffold prescribes plain paths in the SKILL.md References/Assets sections ("Plain paths, no @"); @ includes belong only in action files, which load them. Strip @ from every SKILL.md path listing and citation across all skills, including the flow skills merged earlier. The only @ left are the literal `@claude` webhook triggers in async-dev. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor: rewrite skill descriptions to the skill-generate norm Descriptions now follow the contract shape: a verb-led what-clause, then "Use when the user wants to <natural intents>", then a plain "Not for <overlap>". Drop every reference to another skill and every /command token (slash commands are tool-native), and fix the semantics that were wrong, notably the PRD, which is generated from a need, idea, or brainstorm and is upstream of user stories, not derived from them. Covers the refactored skills plus the gold-standard and merged skills that carried sibling references or slash tokens (onboard, project-memory, brainstorm, shadow-areas, fact-check, todo, commit, pull-request). Also collapse the few multi-word action step labels left from the prior pass to the one-word bold lead the anatomy requires. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(pm): make the PRD template solution-agnostic Headless testing surfaced a contradiction: the bundled prd-template.md shipped 15 sections including Technical Architecture, Tech Stack, and Data Model, which directly violate the skill's "never include technical implementation detail" rule and the action's own 8-section spec. The action's Test also checked for headings the template did not have. Rewrite the template to the action's eight solution-agnostic sections (Overview, Problem Statement, Goals, Non-Goals, User Stories, Acceptance Criteria, Dependencies, Open Questions) with an HTML-comment header so no frontmatter leaks into output, and tighten the action Test to assert those headings and the absence of any solution detail. Re-verified headless: 8 sections present, zero solution detail. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(context): add missing ## Output to cook 02-upsert action * docs(context): add a description size rule to the skill contract (R5) skill-generate produced oversized descriptions every run because R5 told it to "over-list" triggers with only a 1024-char ceiling and no size target. Add the missing rule: two lines max, ~240 chars, straight to the point, length serves recall not completeness. Also fix R5 to match the norm already enforced: lead with a verb, state triggers as "Use when the user wants to ...", never name another skill, never write a /command token (slash commands are tool-native). Update the template placeholder to match. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor: trim 22 skill descriptions to the 2-line size rule Apply the new R5 size rule to every description over the limit: two lines, ~240 chars, straight to the point, same shape (verb-led, "Use when the user wants to ...", plain "Not for ..."). Also drop the last two sibling-skill mentions (onboard/explore) so no description names another skill. 02-user-stories is trimmed in the working tree and left for its own commit. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(context,refine): close 3 defects found by the headless suite Generic and concise, never naming a specific tool or session mode: - explore: detect a tool only by its own mapped surfaces, never a shared parent directory, so a bare CI folder no longer counts as a tool. Add the presence-signal rule to ai-mapping and key survey step 1 to it. - onboard: tighten the silent-read rule so action 01 emits nothing even in a single turn; the first words come from orient. - fact-check: harden the report scrub into a line-by-line final gate and drop the named output-mode list (no "terse/caveman/condensed") in favor of "any active output mode". Re-tested headless: explore and onboard hold; fact-check's wording is clean, though a forced session output mode can still bleed a summary line at runtime, which is a host-mode interaction, not a skill gap. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(vcs,dev): split commit/pull-request into step actions, fix audit folder Model the real workflow as actions instead of one monolithic action: - commit: collect (stage one concern) → message → commit (push when asked). Drop the retry loop: a hook that rejects the commit is not this skill's job, it reports which hook and why and stops, only re-staging files a hook auto-formatted. - pull-request: collect (resolve base, gather commits) → draft (title and body) → create (open draft, label). Conventions live in transversal rules. - audit: one dated folder per run (aidd_docs/tasks/<yyyy_mm>/ <yyyy_mm_dd>_audit/, like a feature folder) defined once in a transversal rule; each pillar action names only its file (architecture.md, security.md, ...); a full run adds report.md. Headless-tested: commit chain produces a conventional commit; a failing pre-commit hook reports and stops with no commit and no loop; audit writes the pillar file into the dated folder. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(vcs): keep the VCS convention in one transversal rule, no baked example Drop the project-specific `feat/ → next` example from the pull-request skill and de-duplicate the convention source: `vcs.md` now appears only in the SKILL transversal rule (where + fallback), and the actions just apply it ("per the project's convention"). The concrete branch examples stay in the bundled fallback asset (branch.md), where examples belong. The skill reads the project's convention from memory instead of assuming one, so behavior is project-driven, not hardcoded. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(dev): audit full run writes every pillar file plus report.md The rule was ambiguous, so a full run wrote only the merged report. Make it explicit: every pillar that runs always writes its own <pillar>.md, and a full run additionally writes report.md. Re-tested headless: a full run produced all 7 pillar files plus report.md. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor: quick conciseness wins across skills Two mechanical passes against the AI-template feel, no behavior change: - Drop the filler "## Default flow" sections: delete the pure single-action ones ("the router dispatches to X whenever a phrase appears"), and inline the real-flow ones as one tight line under the actions table. - Strip obvious embedded git commands from action prose ("the current branch" not the rev-parse incantation, "the remote URL" not the remote-get-url one), keeping the runnable ones in ## Test. The onboard read-project step also loses its hardcoded .github path. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(vcs,dev): de-duplicate what the transversal rule already owns Actions were restating their skill's transversal rule: - repo-init: both Resolve steps repeated the whole "from VCS memory, else the environment, else main" resolution. They now just say "resolve the default branch and provider" / "resolve the host"; the transversal rule owns how. - release-tag: the bump step dropped "only" (the transversal already scopes the bump commit to version-manager files). - assert: dropped the Boundary step that repeated "stop only when the final sweep passes." No behavior change: each convention has one home, the action defers. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(dev): restore assert Boundary step (loop stop belongs in the process) The de-dup pass dropped it as a duplicate of the transversal stop rule, but a fix-loop's termination is part of the process, not just a passive convention. Keep it explicit where the loop decision happens. * refactor: apply four cross-skill principles (args, heavy steps, guards, plain words) - Tool-agnostic input: `$ARGUMENTS` (a Claude-only token other hosts won't read) becomes the plain word "the arguments" across 16 skills. command-generate keeps it, since the token is what it teaches. - One job per step: split plan's overloaded `Write` step into Folder / Fill / Show. - Sequence vs guards: implement's `02-execute` separates the Open / Code / Assert loop from the Blocked and Drift stop conditions, which are no longer numbered steps. - Plain words: implement's `Branch` drops the embedded `git checkout -b` for "create a feature branch". Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor: lighten the remaining heavy steps (one job per step) Split learn's Ask into Show + Decide and ship's Gate into Gate + Freshness; tighten sdlc's Iterate and refactor's architecture Source. No heavy (>320-char) steps remain. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(dev): separate guards from sequence in 06-test - test-journey: the Failure handling is part of the Walk, not a peer step; fold it under Walk as a sub-bullet. - test: drop the Boundary step that restated the transversal rules (functional only, never compromise quality); one-at-a-time stays implicit in the generate-then-next loop. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(dev): make execute stop-conditions a Guard step with sub-bullets The free-floating 'Stop the loop on either condition' block diverged from the action anatomy. Express it as a numbered Guard step whose conditions are sub-bullets (the anatomy's form for a branch), keeping Blocked and Drift as bullets, not peer steps. Left in the action, not the transversal rules: these guard the execute loop only. * fix(dev): review must not assume 'main' as the diff base All three review actions hardcoded `git diff main`. Resolve the diff against the repository's default branch instead, never a blind main, matching the pull-request skill's base handling. * fix(dev,vcs): two blockers the deep audit caught - for-sure: the spawned-worker prompt had no payment/destructive carve-out, and the loop never consumed the worker's stop status, so a worker that halted on a money gate was retried, re-triggering the spend. Add the carve-out to the worker prompt and make the loop surface a gate stop and halt instead of retrying. - release-tag: a tagless repo fell back to a fictitious v1.0.0 and then ran `git log v1.0.0..HEAD`, which aborts. With no prior tag, list every commit on the branch instead. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix: clear semantic defects from the deep audit (batch 1) - prd task-template: drop the stray brace in "## Main step 1}". - command-generate README: correct the stale "# 10" heading to "# 07". - issue-create template: "GitHub issues" to "issues in the configured tracker" (the skill is tool-agnostic). - 01-plan SKILL: drop the dangling reference to a mermaid-conventions.md that does not exist in the skill and no action uses. - commit and pull-request READMEs: replace the stale "single action" description with the three-action chain each skill now has. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix: semantic defects from the deep audit (batch 2) Mostly Output/Test mismatches where the Test under-verified or contradicted the declared Output, plus a few wording bugs: - sync Tests to the full Output (ticket-info six fields, cook README index row, debug flowchart + confidence, test-journey full fields). - audit: rename the template header to "## Top actions" so all seven pillar Tests match it. - spec: assert "every required section" (the validator marks some optional), matching the Process. - fact-check: re-render restyled lines in plain prose instead of deleting them (deleting lost content). - plan phase-template: legend now reads deleted/created/modified, matching the upstream modify/create/delete vocabulary. - todo: drop "using template" (no such asset); learn-sync names the updated files (its Test wants names, not a count); agent-generate Clarify now asks the model (it is a declared output); test marks a declined behavior pending; refactor allows a security-fix regression test. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(pm): replace user-stories-create with the six-action user-stories skill Rename and rebuild: 02-user-stories-create (one monolithic action) is replaced by 02-user-stories, a six-action chain (clarify-scope, split-epic, draft-stories, estimate-impact, prioritize, sync-tracker) with INVEST, a Definition of Ready and a functional Definition of Done. Update the plugin manifest, catalogs, marketplace, READMEs, and the upgrade map. Applies its two audit findings: estimate-impact now lists dependencies in its Output, and clarify-scope no longer restates the transversal lean-clarification rule. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * docs: memory-deduplication diagnosis and dedup plan Add the memory-duplication diagnosis and the memory-dedup feature folder (plan, phases, empirical results), plus the two one-line doc touches that came with that work. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * fix(context): repair the garbled scaffold-docs action label The action-table cell had become 'Create the `folder structure' with an unclosed backtick and the lost aidd_docs/ reference. Restore it to 'Create the `aidd_docs/` folder structure'. * fix(skills): resolve semantic contradictions from deep audit Align action contracts with their stated outputs, tests, and SKILL rules: - onboard: orient stops at offer+capture; 03-act owns acting on every choice - sdlc ship: implement the promised on_default_branch gate (commit nothing) - skill-generate: contract is R1-R13; draft output covers plugin-source mode - agent-generate: capture picks one name, output/test follow - bootstrap: pick a candidate by name; fold cost check into the rationale - hook-generate: scope tool-paths to commonly supported moments, no gap promise - challenge: drop orphan report section; make the 100% confidence tier exclusive - shadow-areas: symmetric dispatch; template renders only categories with gaps - assert-frontend: output schema matches the candidate-causes step - pull-request README: list only the assets that ship Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * refactor(skills): drop host tokens, make two descriptions verb-led and plain - review, mermaid: verb-led descriptions, plain "Not for", under the limit - onboard, commit, condense, implement: drop host-specific slash and $ARGUMENTS tokens from instruction prose for plain capability language Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * refactor(skills): drop passive restatements the transversal already owns Remove trailing reminder clauses that echo a SKILL transversal rule while the step's primary action stays intact: explore (skip-surface, no-advice, no-assume), debug (no drive-by refactors), mermaid (no unconfirmed node), learn (preserve edits), review (never patch, x2), assert-architecture (never fix), spec (solution-agnostic). Load-bearing guards, loop terminators, and point-of-action safety constraints are kept. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * refactor(spec): keep the behavioral half of the build-step guard The transversal owns the artifact property (solution-agnostic), but not the agent action: restore "Do not explore the codebase." so the step still forbids reading the code, dropping only the redundant property restatement. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * refactor(onboard): cut meta-commentary from orient step 7 The step instructs; explaining that 03-act delegates here is noise. The handoff is already documented from 03-act's side. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(skills): repair structural anatomy bugs from deep audit - cook: asset ref needs the parent path (@../assets) to resolve from actions/ - agent-generate README: header number 06, not 09 - command-generate tool-paths: drop the duplicated table header row - debug task-template: remove stray brace in "## Main step 1" - todo: dangling colon ends the Report step as a sentence - for-sure tracking template: drop skill-metadata frontmatter that leaked name/description/argument-hint into every generated plan file Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(rule-generate): correct sibling @-path in references @-paths are file-relative (actions use @../references/...), so from inside references/ the sibling is @rule-authoring.md, not @references/... which resolved to references/references/. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(skills): repair four defects surfaced by the headless behavioral test - agent-generate: Codex TOML rule mandated single-quote literals, which cannot hold an apostrophe; use a double-quoted basic string with escaping - async-dev: actions 03/04/05 read assets/<t> but the templates live at assets/setup/<t>; restore the setup/ segment so the Read resolves - shadow-areas: empty source stopped with an uncategorizable blocker gap that broke the locked-category rule; stop with a plain warning, no gap - todo: the mandated per-executor refine resolved to an interactive-only skill that blocks in autonomous subagents; fall back to inline refine Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(skills): wire onboard hand-off, drop unused issue type field - onboard: orient now offers "run it in a fresh session instead", so 03-act's Hand off outcome is reachable instead of a dead branch - issue-create: drop the type field from Input/Output/Validate; Create never applied it and gh issue create has no native type flag Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * fix(docs): repair broken markdown links failing CI Validate - regenerate plugin CATALOGs: drop the deleted 05-pull-request-list section (aidd-vcs) and the stale 01-plan mermaid-conventions row (aidd-dev) - task scratch docs: the @-build illustration used literal [x](../x) link syntax the checker resolved as a path; reword to prose Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> * refactor(aidd-context): dedupe project-memory templates and lighten actions (#347) Memory generation restated the same fact across files (architecture vs codebase-map, integration internal flow, stack libs) because templates overlapped and the fill ran blind in parallel. Reframe ownership so each fact has one home and others reference it. Templates - architecture: drop Structure (codebase-map owns areas + entry point); Stack lists cross-cutting libs only, a capability's lib lives in its file - integration: reframe to external/third-party services (drop Internal, which duplicated architecture's flow) - api/navigation: scope routes to server vs client Actions / rules - 03: fill is no longer blind in parallel; each template knows what earlier ones captured - SKILL: split rules into Memory rules (content) and Action rules (process); add "one fact, one file" with the define-here / reference-elsewhere nuance - 04: semantic dedup review step (no fact's definition duplicated) - new references/memory-block.md; drop capability-signals' redundant section Validated headless on three stacks (cli, kairos web/mobile, kairos existing): definition-level duplication gone, files stay self-contained. Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com> * docs(onboard): brainstorm + validated plan for the onboard refactor Brainstorm (guide that enforces a skippable foundation, then a project- adapted menu) and a 4-phase plan, hardened across three fixture-matrix dry-runs: flow-walk owns the default, signal map ranks secondary tools, read-once with a session ledger that tracks done/skipped steps. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * feat(onboard): phase 1 — read-once loop with a session ledger Replace re-read-every-loop with a single reusable snapshot plus a session ledger of run/skipped steps. 01 produces the snapshot (cheap signals + a bounded richer read) and the ledger; 03 writes it (run → done, different step → skipped); 02 consumes it (a stage is met by a disk fact OR the ledger), so a just-run or declined step is never re-suggested even when it left no file behind. journey.md points to 02's met-definition, no restatement. Behaviorally verified: empty-repo re-nag, off-disk Track completion, and the skip path all advance instead of looping. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * feat(onboard): phase 2 — foundation gate, skippable 02-orient marks a foundation stage as the loud pre-selected default with an explicit branch — architect a stack on an empty repo, set up the memory bank on a content repo with weak memory — rendered as the first choice tagged (recommended), always skippable. SKILL keeps suggest-never-force with the skippable + skip-remembered note. 01 pins that a docs/README-only repo counts as empty. Behaviorally verified: empty→architect, weak-memory→memory, skip excludes the default from the next loop (no re-nag), and the gate never blocks. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * feat(onboard): phases 3-4 — capability map + project-adapted menu Push detail into references/assets so the actions stay one line: - references/signals.md: the signal catalogue (cheap + a bounded richer read: code-quality sample, bug-marker scan) and the capability map (signal → stage-gated secondary tool) - assets/menu.md: the menu shape — default, ranked secondary tools (≤~5), a never-dropped footer, need-stage and end-stage (review-never-skipped) rules - 02-orient: place → rank → resolve → brief → assemble, one line each - journey.md: cumulative-stage placement made explicit (a plan implies the need is clarified/tracked); open-PR row reviews before ship; tool ranking deferred to signals.md, no duplication Behaviorally verified: mid-build → Build alone + add-tests/audit/debug; open-PR → Review then Ship; rough idea → no loud default, Clarify first; empty/code-no-memory foundation + ledger advance regress clean. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * docs(onboard): refresh README for the snapshot+ledger+menu flow Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * refactor(onboard): slim 03-act, drop colon from the description - move the per-choice outcome table into references/outcomes.md; 03-act becomes a thin act + ledger dispatch (722 -> 171 words) - description no longer uses a colon Verified live (headless claude -p): the menu and the "explain the step" outcome render correctly from the new reference. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * refactor(skills): forbid colon and em dash in descriptions, codify in R5 Rewrite the four descriptions that used a colon (cook, assert, async-dev, repo-init) into two sentences, and add the rule to R5 in skill-authoring.md so it holds going forward. assert also drops "Do NOT use to" for plain "Not for". Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com> * chore: release main (#364) Co-authored-by: aidd-bot[bot] <290648294+aidd-bot[bot]@users.noreply.github.com> --------- Signed-off-by: dependabot[bot] <support@github.com> Co-authored-by: dependabot[bot] <49699333+dependabot[bot]@users.noreply.github.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com> Co-authored-by: aidd-bot[bot] <290648294+aidd-bot[bot]@users.noreply.github.com> Co-authored-by: Alex <8973343+alexsoyes@users.noreply.github.com>

alexsoyes assigned blafourcade Jun 20, 2026

blafourcade and others added 2 commits June 20, 2026 21:37

blafourcade force-pushed the refactor/implement-redesign branch from dce36ef to e448922 Compare June 22, 2026 04:20

blafourcade force-pushed the refactor/implement-redesign branch from e448922 to dec5acb Compare June 22, 2026 04:30

blafourcade and others added 6 commits June 22, 2026 07:28

blafourcade changed the title ~~refactor(framework): split implement into prepare/execute/finalize, delegate SDLC implement~~ refactor(framework): executor/checker agents + SDLC orchestration redesign (recipe-not-spawn) Jun 23, 2026

blafourcade and others added 16 commits June 23, 2026 09:42

docs(framework): shorten review and implement descriptions

170d0b1

Cut both to two lines and drop the SDLC mention from implement: a skill's description states what it does and when, not how an orchestrator calls it. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

blafourcade and others added 17 commits June 24, 2026 12:14

refactor(framework): drop the redundant 'never create a plan' from re…

d784e70

…view A review writes review.md; it has no reason to create a plan. Forbidding it defends against a non-event. Removed. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

docs(framework): trim sdlc, assert, review, spec descriptions to two …

2ffe975

…lines The four were running ~three lines. Cut to the essential what + when + key do-not, under 210 chars each, keeping the trigger words. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Merge remote-tracking branch 'origin/next' into refactor/sdlc-agents-…

dcec76d

…redesign # Conflicts: # README.md # plugins/aidd-dev/agents/reviewer.md

refactor(vcs,dev): tighten the project-memory rules

e890822

Trim the wording added for project-memory deference across commit, pull-request, assert, and the SDLC ship step. No behavior change: same precedence, fewer words. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

blafourcade merged commit 79ebfdd into next Jun 24, 2026
7 checks passed

blafourcade deleted the refactor/implement-redesign branch June 24, 2026 19:53

aidd-bot Bot mentioned this pull request Jun 30, 2026

chore: release main #364

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

refactor(framework): executor/checker agents + SDLC orchestration redesign (recipe-not-spawn)#314

refactor(framework): executor/checker agents + SDLC orchestration redesign (recipe-not-spawn)#314
blafourcade merged 43 commits into
nextfrom
refactor/implement-redesign

blafourcade commented Jun 20, 2026 •

edited

Loading

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

blafourcade commented Jun 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

🎯 What & why

🛠️ How it works

🧪 How to verify

⚠️ Heads-up

🔗 Linked issue

✅ I certify

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

blafourcade commented Jun 20, 2026 •

edited

Loading