Adds a new PreToolUse hook that blocks gh issue/PR/comment creation and
gh api .../issues|/pulls calls targeting a public framework repo
(default: me2resh/apexyard + whatever `upstream` resolves to) when the
title or body references any registered private project from
apexyard.projects.yaml (by name, repo slug, owner/repo#N ticket ref, or
workspace path).

The hook is a sibling to check-secrets.sh — both scan outgoing content
for identifiers that should never leave the local environment. Skip
marker `<!-- private-refs: allow -->` in the body lets a deliberate
reference through with a visible warning.

Files touched:
- .claude/hooks/block-private-refs-in-public-repos.sh (new)
- .claude/hooks/tests/test_block_private_refs.sh (new)
- .claude/rules/leak-protection.md (new)
- .claude/settings.json (wire PreToolUse matchers for the 5 gh shapes)
- docs/rule-audit.md (append section 10 + bump counts)

Refs: #110

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

Fires after `git push` to surface review markers that have gone stale
because new commits were pushed past an existing Rex / CEO / design
approval. The merge gate already catches this at `gh pr merge` time,
but only then -- this hook closes the gap by flagging it immediately
at push-time so the author isn't surprised at merge.

- `.claude/hooks/warn-stale-review-markers.sh`
  - PostToolUse, non-blocking (PostToolUse exit 2 would push noise
    into the conversation; this hook is purely informational).
  - Resolves the PR HEAD via `gh pr view --json headRefOid` -- same
    source-of-truth as the merge-gate hooks post-apexyard#47 / #55.
    Falls back to local HEAD with a visible WARN when gh is offline.
  - Silent on: no PR for branch, no markers, fresh markers,
    failed push (detected via `rejected` / `failed to push` /
    `fatal:` / `error:` markers in tool_response.stderr).
  - Modes: `warn` (default) prints one stderr line per stale marker;
    `delete` opts in to auto-removal via
    `.claude/project-config.json` -> `review_markers.on_stale`.
    TODO(apexyard#109): switch to the shared project-config reader
    once it lands.
- `.claude/settings.json`
  - Wires the hook on PostToolUse / Bash / `git push *`.
- `docs/rule-audit.md`
  - Adds a row under section 3 (Code review & PR quality) and
    bumps the mechanized count 26 -> 27 / total 73 -> 74.
- `.claude/hooks/tests/test_warn_stale_review_markers.sh`
  - 8 cases: no PR, no markers, fresh markers, stale rex / ceo /
    design (warn), delete mode, failed push. All pass locally.

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

… check-runner (#121)

* chore(#111): upgrade pre-push-gate from reminder to blocking check-runner

Previously pre-push-gate.sh just printed a checklist of things to run
locally before pushing — it was advisory. The rule it enforces is a
HARD STOP per pr-workflow.md. That asymmetry meant agents routinely
pushed broken work and discovered it only when CI went red.

Replaces the reminder with a blocking runner that reads the list of
shell commands from project config (.pre_push.commands) and executes
them in sequence before a push is allowed through. First non-zero
exit blocks the push with exit 2 and prints the failing command plus
the last 20 lines of its output.

- Config key: .pre_push.commands[] — array of {name, run} objects.
  Shipped default is an empty list (hook stays a no-op on repos that
  haven't configured their checks yet, including the framework repo
  itself until it wires its own CI).
- Emergency bypass: '<!-- pre-push: skip -->' in the HEAD commit
  message. Grep-able on purpose so bypasses stay auditable.
- Fail-fast: once a command fails, the rest don't run. Parallel
  execution is a follow-up polish.
- 7 test cases in .claude/hooks/tests/test_pre_push_gate.sh — all
  pass on the shipped default + a minimal custom config.

Updates docs/rule-audit.md to flip "partial" → "yes" for the
"before git push" rule.

Integrates with the shared config reader landed in #109.

#111

* fix(#111): remove orphaned footnote reference from rule-audit

The previous advisory-mode footnote was superseded by pre-push-111
but its definition was accidentally kept, tripping markdownlint MD053
(unused reference definition). Drop it.

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

Mechanically enforces the ticket body schema when an agent files raw
`gh issue create` calls instead of going through the interactive
/feature, /task, /bug skills. Matches bracketed title prefix
([Feature] / [Chore] / [Bug] / [Docs] / etc.) against
`.ticket.required_sections` in project-config, and blocks (exit 2)
when any required section is missing or empty. Skip marker
`<!-- validate-issue-structure: skip -->` bypasses with a visible
stderr WARN for legitimate off-template tickets (epics, meta-threads).

Changes:

- .claude/hooks/validate-issue-structure.sh — the hook; reads schema
  via the shared _lib-read-config.sh, with inlined defaults for bare
  checkouts predating the config-schema rollout. Handles
  --body / --body-file / -F path.
- .claude/project-config.defaults.json — extends .ticket with
  required_sections (Feature/Chore/Refactor/Testing/CI/Docs/Bug) and
  skip_marker; other .ticket fields untouched.
- .claude/settings.json — new PreToolUse matcher on Bash(gh issue
  create *) alongside the existing suggest-ticket-template.sh and
  block-private-refs-in-public-repos.sh hooks.
- .claude/hooks/tests/test_validate_issue_structure.sh — 15 cases
  covering pass + fail paths per prefix, empty section detection,
  skip marker, unknown prefix, non-gh invocation, --body-file path.
- docs/rule-audit.md — new section 11 row, mechanized count +1.

Upstream ticket: #107

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

Closes the asymmetry noted in .claude/rules/agdr-decisions.md: every
other HARD STOP in the ruleset (merge approval, ticket-first,
migration-first) is mechanically enforced, but the /decide HARD STOP
was prose-only. The commit-time hook require-agdr-for-arch-changes.sh
catches one architectural change at commit; this new PR-time hook
catches the cumulative diff so reviewers always have a pointer to the
decision record.

- New hook at .claude/hooks/require-agdr-for-arch-pr.sh
  - Fires on Bash(gh pr create *)
  - Parses --title/--body/--body-file/-F <path>
  - Resolves base branch from --base, else upstream/dev, origin/dev,
    upstream/main, origin/main, main, master (in that order)
  - Computes `git diff <merge-base>..HEAD --name-only`
  - Triggers on any changed file matching .agdr_trigger_paths[], OR any
    dep-file addition (package.json via jq key-set diff; other
    dep files via a commented +/- line-count heuristic — version
    bumps match +/- counts and do not fire)
  - Blocks (exit 2) with a helpful message naming the triggers and
    pointing at /decide if the body has no `AgDR-\d+-[a-z0-9-]+`
    reference
  - Skip marker `<!-- agdr: not-applicable -->` bypasses with a
    visible WARN on stderr
  - Silent exit 0 on non-gh commands, empty diffs, unresolvable base

- Wired via .claude/settings.json PreToolUse Bash(gh pr create *)

- Adds two new top-level keys to .claude/project-config.defaults.json:
    agdr_trigger_paths      (shell globs — domain/, infrastructure/,
                             migrations/, *.tf, .github/workflows/, etc.)
    agdr_trigger_dep_files  (literal basenames — package.json,
                             pyproject.toml, Cargo.toml, go.mod, Gemfile)
  Hook has inline fallback defaults kept in sync.

- Adds docs/rule-audit.md entry in the AgDR section; bumps mechanized
  count 26 to 27 and total rows 73 to 74.

- Adds .claude/hooks/tests/test_require_agdr_for_arch_pr.sh (7 cases;
  all green): path-triggered without AgDR (block), with AgDR (pass),
  dep-file added (block), version-only bump (no fire), skip marker
  (pass + warn), non-matching diff (pass), non-gh command (no-op).

Closes #112

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

Extends validate-pr-create.sh with a required-sections check that
replaces the hardcoded Glossary-only grep. The list of required H2
headings is project-configurable via `.pr.required_sections[]`.
Shipped default is ["Testing", "Glossary"], matching the canonical PR
description shape in workflows/code-review.md.

- Each entry must appear as `## <Name>` (case-insensitive).
- Empty sections are tolerated at this layer (the issue-structure hook
  #107 does stricter empty-content checks for issue bodies; for PR
  bodies, empty sections are left to the reviewer's judgement).
- Skip marker `<!-- pr-sections: skip -->` bypasses with a visible
  stderr WARN — for trivial PRs (lint-only fixes, version bumps)
  where the full template is overkill.
- Reads from project config via the shared _lib-read-config.sh (#109).
  Inline fallback matches shipped defaults so bare checkouts predating
  #109 keep working.
- 8 test cases cover: all-sections pass, each missing section,
  missing-both (both errors printed), skip marker, case-insensitive
  headings, H3 rejection.

#113

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

* chore(#114): enforce single Closes-keyword per PR body

Caps distinct auto-closing references (close/closes/closed, fix/fixes/
fixed, resolve/resolves/resolved + N or owner/repo+N) at one per PR
body. Closes the loophole where the title validator limited the title
to one ticket but multiple Closes lines in the body would still auto-
close all of them on merge.

- Scans stripped of fenced code blocks so closing keywords inside a
  code sample do not count.
- Distinct counting: the same number referenced twice (e.g. via Fixes
  and Closes) counts as one.
- Cross-repo refs (owner/repo+N) count normally.
- Opt-in escape hatch: pr.allow_multiple_closes=true in
  project-config disables the check for teams that deliberately batch
  rollbacks or dependency bumps.
- Per-PR bypass: a multi-close-approved HTML comment in the body
  prints a visible stderr WARN and lets that PR through. Grep-able
  trace so bypasses are auditable.
- 10 test cases cover: one close passes, no-keyword passes, two
  distinct block, three mixed block, same-number-twice passes, code-
  fence-ignored, skip marker, cross-ref without keyword, opt-in
  config, cross-repo close.

Reads configuration via the shared _lib-read-config.sh (apexyard+109).

#114

* fix(#114): strip inline backticks and tilde fences from close-count scan

Rex caught a self-reflexive bug in the initial commit: documentation
mentioning closing keywords inside inline backticks (say a PR body
that explains the new hook with examples) counted as real closes, and
a skip marker inside inline backticks silently bypassed the check.
Future PRs that document the feature would trip the same trap.

Fix the code-region stripper to cover:
- Triple-backtick fences (already handled)
- Tilde fences (new)
- Inline-backtick spans (new)

Also run the skip-marker check against the stripped body, so a marker
used purely as documentation no longer activates a real bypass.

Three new test cases pin the behaviour:
- closing keywords in inline backticks are ignored
- skip marker inside inline backticks does NOT bypass
- tilde fences also get stripped

13/13 tests pass.

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

The fast happy path for filing 5–20 structured tickets in one intent
without dropping to raw `gh issue create` (non-conformant) or running
`/feature` 20 times serially (~100 turns of interview).

- .claude/skills/tickets-batch/SKILL.md — new skill spec. Asks
  shared-context questions (priority, epic, area-labels, repo) ONCE
  for the whole batch, then runs a ≤3-question micro-interview per
  ticket (type, one-line purpose, optional clarification when the
  inference is low-confidence). Confirms the full batch as a table,
  then files each via specific `gh issue create` calls (never a
  bulk JSON dump — the validator runs per-issue). Output conforms
  to `.ticket.required_sections` by construction. Caps at 20
  tickets per invocation.

- CLAUDE.md — added a row for /tickets-batch in the Available
  Skills table; bumped the count references from 33 to 34.

Refs #108

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

- Add `.claude/skills/fan-out/SKILL.md` — spawns N parallel Agent calls
  in a single assistant message, with per-task agent type, worktree
  isolation, and foreground/background mode. Caps at 5 concurrent
  agents. Refuses fan-out when tasks share file write targets or have
  sequential dependencies. Includes pre-spawn active-ticket safety
  check and worktree merge-back flow that pauses on conflict.
- Add `.claude/rules/parallel-work.md` — trigger heuristic for when an
  agent should proactively offer fan-out (>= 2 file-independent,
  context-independent, individually substantial work items). Pairs
  with the skill: rule says when, skill says how.
- Update `CLAUDE.md` — bump rules count to 9, skills count to 34, add
  `/fan-out` row to the skills table.

Refs #117

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

…work only (#126)

* chore(#116): adopt release-cut branch model (dev/main + tags)

Formalises the dev/main split that already exists informally — dev is
the daily-work branch where every PR lands, main is release-only,
tagged with semver on each merge. Framework-only: managed projects
under apexyard governance stay trunk-based.

Added:
- AgDR-0007 — decision record (options table covers full git flow vs
  trunk-only vs gitflow-lite; chose gitflow-lite)
- /release skill — diff dev against main, propose semver bump from
  conventional commits, generate CHANGELOG, open release PR, tag
  after merge
- docs/release-process.md — prose runbook for cutting a release
  (manual fallback for the skill)
- .git.protected_branches in project-config.defaults.json
  (main/master/dev/develop)

Modified:
- block-main-push.sh — now blocks direct pushes/commits to all
  configured protected branches (was: hardcoded main/master). Reads
  .git.protected_branches via the shared config reader (apexyard#109).
- CLAUDE.md — new section under Git Conventions explaining the
  dev/main model + the framework-only scope. Skill table entry for
  /release. Skills count bumped to 34.
- docs/multi-project.md — note that upstream/main is release-only
  and the dev/main split is framework-only.

Non-consequences (per AgDR-0007):
- No release/* or hotfix/* branches. Hotfixes are normal patches
  cut quickly. Revisit if multi-version maintenance becomes a need.
- No automatic on-merge issue closing for dev PRs. The release PR's
  body aggregates all Closes references for the batch and triggers
  auto-close en masse when it merges to main. Manual close in the
  meantime.
- CI workflows trigger on pull_request regardless of base, so
  dev-targeting PRs already get the full check matrix — no
  workflow file edits needed.

#116

* fix(#116): satisfy markdownlint MD032 + MD060 in 116 docs

Auto-fix added blank lines around lists in AgDR-0007 (MD032) and
spaced the table separators in AgDR-0007 + multi-project.md (MD060).
Content unchanged.

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

…129)

* fix(#106): CHANGELOG fallback in drift hook for squash-merged forks

The v1.1.0 tag-reachability check (`git tag --merged main`) misfires on
forks that sync via GitHub's default squash-merge: the squash collapses
the upstream-tag commit into a synthetic SHA, the tag stops being
reachable, and the banner keeps firing forever.

Discovered live on the first real-world `/update` flow: ops fork
squash-merged the v1.1.0 sync PR, banner kept saying "v1.1.0 available"
even though the fork was content-caught-up.

This commit adds a CHANGELOG-content fallback that fires only when the
primary tag check fails. If the fork's main has a heading
`## [X.Y.Z]` matching the upstream tag's version, treat the release as
absorbed and stay silent. Tolerant grep (matches the apexyard CHANGELOG
format from v1.1.0 onward, with leading-`v` stripping for tag→heading
conversion).

The merge-commit and rebase paths are unchanged — primary tag check
still works for them, and the fallback never fires when it shouldn't.

Test coverage (5 cases):
- squash-merge fork caught up to v1.1.0 → silent (the fix)
- merge-commit fork caught up to v1.1.0 → silent (regression check)
- fork stopped at v1.0.0 → banner fires
- fork has its own newer tag → silent
- squash-merge but no CHANGELOG on fork → banner fires (no false silence)

Records the strategy update in docs/agdr/AgDR-0008-…md (extends
AgDR-0005's tag-based-drift design).

#106

* fix(#106): satisfy markdownlint MD032/MD060 + shellcheck SC2164

Rex flagged two CI-blocking issues on the original 106 commit:

- AgDR-0008 had three bulleted sub-lists in the Consequences section
  without surrounding blank lines (MD032). Added blanks and padded the
  one tight-pipe table separator (MD060).
- The 106 test fixture had five subshell `cd "$fk"` calls without
  `|| exit 1` (SC2164). Added the guard to all five.

5/5 tests still pass after the fix. No semantic change to the hook
or the test logic.

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

A 10-minute, 5-question check that sits between /idea and /write-spec.
Designed for solo founders running ApexYard — not a heavyweight
methodology like event storming or Wardley mapping.

Five questions, asked one at a time:
  1. Who is this specifically for?
  2. What do they do today instead?
  3. What's the smallest version that proves the value?
  4. What would prove this is wrong? (kill criteria)
  5. Build, buy, or rent?

Output: a one-page validation doc with a GREEN/YELLOW/RED verdict.
RED auto-updates the IDEA-NNN backlog row to WONTDO.

Integration:
  /idea — adds an optional default-no "Validate now?" step after
    capture (and after the optional GitHub Issue offer).
  /handover — adds a conditional "this looks dormant, validate?"
    step at the end of the integration plan, gated on the dormancy
    heuristic (last commit > 90d AND zero open PRs AND no recent
    issue activity). Healthy projects don't see the prompt.

CLAUDE.md skills count bumped to 35; new skills row added.

#130

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

…prompts-on-pause) (#135)

Stop hook that speaks the assistant's question aloud (Jarvis-style)
when it pauses for user input. Initial phase is macOS-only via `say`,
no voice input — user replies via keyboard.

Default OFF. Adopters opt in by overriding `voice_prompts.enabled` to
true in `.claude/project-config.json`.

Files:
- .claude/hooks/voice-prompt-on-pause.sh — Stop hook with config gate,
  trigger heuristic (questions-only by default), markdown stripping,
  sentence-boundary truncation, fire-and-forget say invocation
- .claude/hooks/tests/test_voice_prompt_on_pause.sh — 9 cases covering
  disabled-default, enabled+question, enabled+statement, approved-pattern,
  abc-menu, malformed-transcript, no-say-on-PATH, trigger-always,
  markdown-stripping
- .claude/project-config.defaults.json — voice_prompts schema block
  added (enabled, voice, max_chars, rate_wpm, trigger), default OFF
- .claude/settings.json — new Stop hook entry wired with the standard
  ops-root resolver wrapper
- docs/agdr/AgDR-0009-voice-prompts-on-pause.md — design rationale,
  options matrix (status quo / macOS say / cloud TTS / ML detection),
  consequences, future phases
- docs/project-config.md — new "Voice prompts" section with override
  examples and privacy notes

Test mode: hook respects VOICE_PROMPTS_SYNC=1 to run say synchronously
(test runners need this so assertions don't race against orphaned
background processes). Production invocations always run async.

Future phases (out of scope here, AgDR §"Future phases"):
- Phase 2: cross-platform TTS (Linux espeak, Windows SpeechSynthesizer)
- Phase 3: cloud TTS providers (OpenAI, ElevenLabs) — privacy AgDR-worthy
- Phase 4: voice input via Whisper-based STT
- Phase 5: per-message overrides

Refs: #134

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

…#142)

* feat(#141): add /debug skill — structured hypothesis-driven debugging

Adds a methodology skill that enforces five disciplines:

  1. Capture the symptom precisely (exact URL, exact response, exact step)
  2. Read the architecture before guessing (map every layer the request
     touches, file by file)
  3. Form a hypothesis ladder (3–5 candidates, each with an explicit
     evidence test that confirms or refutes it)
  4. Gather evidence first, fix second
  5. Verify the fix against the original symptom evidence (re-run the
     same `curl` / browser repro you used in step 4 — unit tests
     verify code, not feature, correctness)

Stack appendices (Web, Desktop) carry stack-specific surface-evidence
requirements (step 1), architecture-surface maps (step 2), and
evidence-tests cookbooks (step 4). The methodology body stays portable
across stacks; appendices are where stack-specific knowledge accrues
over time.

Web appendix covers browser routing, framework configs (Next/Nuxt/Vite),
SPA-fallback layers, CDN, origin, the shared API client, backend
handlers, and auth providers. Desktop appendix covers Electron / Tauri
/ native-shell concerns: app entry points, IPC bridges, native modules,
auto-updater, sandbox / entitlements, code signing, crash reports.

Includes "When NOT to use" guidance so the methodology overhead doesn't
sandbag simple bugs (typos, off-by-ones, greenfield exploration).

Motivated by a real OAuth debug session in a managed project where
three sequential fixes chased adjacent symptoms because each was
hypothesis-then-fix without evidence in between. The skill is the
"never do that again" guardrail.

Closes #141

* fix(#141): scrub private project issue numbers from anti-pattern table

Rex review on PR #142 caught that line 155 of the skill's anti-pattern
table still named the originating PRs (#375, #377, #380) from the
private project where the methodology was first exercised. The PR body
and commit message were correctly abstracted earlier, but this in-file
reference slipped through — the leak-protection hook only scans gh
issue/pr writes, not staged file content, so mechanical enforcement
didn't catch it.

Replaced with "Three sequential PRs chasing the same symptom because
each was based on a different guess (no evidence test in between
cycles)" — same pedagogical value, zero attribution.

Refs #141

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

…144)

Adopters on GitHub Free with any private project hit a silent privacy bug
following today's docs: forking apexyard makes a public fork that you cannot
later flip to private (GitHub policy), and committing `apexyard.projects.yaml`
+ `projects/<name>/` to the fork publishes private project names + handover
findings on a public GitHub repo.

This PR documents the supported workaround — split-portfolio mode — and
adds an upfront privacy gate to the /setup skill so new adopters never
hit the trip-wire silently.

docs/multi-project.md:

- New "Two setup modes — pick the one that matches your privacy needs"
  section before TL;DR, with a side-by-side table and the explicit
  trip-wire callout.
- Existing TL;DR retitled "TL;DR — single-fork mode (default)" with a
  one-line pointer to the split-portfolio section.
- New "Split-portfolio mode — public framework + private portfolio"
  section between the existing setup steps and the directory-layout
  section. Includes:
  - The two-repo layout (~/ops/apexyard public + ~/ops/portfolio private)
  - 7-step setup walkthrough with copy-pasteable commands
  - Daily workflow + upstream sync notes (both unchanged)
  - Trade-offs (two repos to maintain, two clones per machine, one
    upstream-sync conflict path on `projects/README.md`)
  - "Migrating from single-fork to split-portfolio" recovery flow with
    the explicit warning that GitHub Issue / PR edit history survives a
    force-push and must be redacted separately

.claude/skills/setup/SKILL.md:

- New Step 2a: privacy gate — asks "are any projects private?" before
  proposing the config. Branches on the answer:
  - All public                                → single-fork mode
  - GitHub Pro / Team / Enterprise            → single-fork mode (private
                                                  forks of public repos
                                                  are supported on those
                                                  plans)
  - Any private + GitHub Free                 → split-portfolio mode
- New Step 2b: walks through the split-portfolio setup interactively
  (private repo create, sibling clone, gitignore + symlink) when the
  privacy gate triggers.
- Detection: `test -L apexyard.projects.yaml` short-circuits Step 2b
  for adopters already in split mode.
- Explicit "do NOT auto-migrate" rule for adopters already in single-fork
  mode with private names already pushed — that path is destructive
  (force-push history rewrite + redact issue/PR bodies + delete backup
  branch) and warrants a deliberate, eyes-open run, not a /setup side
  effect.

Out of scope for this PR (tracked separately on #143):

- `portfolio:` config block in `onboarding.yaml` schema
- Skill audit + refactor to honour configured `registry` / `projects_dir`
  / `ideas_backlog` paths instead of hardcoded fork-relative paths
- `/split-portfolio` migration helper skill that automates the recovery
  flow currently documented manually

This is the docs-and-setup-question minimum-viable starter — the
framework code refactor is mechanical and lands as a follow-up.

Refs #143

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

…#147)

* feat(#145): portfolio config block + self-healing + /split-portfolio helper

Closes the framework primitive deferred from #144. Adds
first-class config-driven path resolution for the portfolio registry,
projects dir, and ideas backlog, with self-healing surfacing of broken
config at session start, plus a new /split-portfolio skill that automates
the destructive recovery flow.

Schema, helper, and hook:
- .claude/project-config.defaults.json: new portfolio: block
  (registry, projects_dir, ideas_backlog) with defaults matching
  today's single-fork layout
- .claude/hooks/_lib-portfolio-paths.sh: new sourceable helper exposing
  portfolio_registry, portfolio_projects_dir, portfolio_ideas_backlog,
  portfolio_validate, portfolio_clear_cache. Resolves relative paths
  against the ops-fork root.
- .claude/hooks/check-portfolio-config.sh: new SessionStart hook —
  silent on OK, one-line banner on broken config, never blocks session
- .claude/hooks/tests/test_portfolio_paths.sh: 13 cases covering
  defaults, absolute/relative overrides, validate states, cache clear

Skill audit (18 SKILL.md files):
- Adds Path resolution callout pointing at the helper
- handover bash blocks now source helper and use $(portfolio_registry)
  instead of literal apexyard.projects.yaml
- setup Step 2b now writes the portfolio: config block (recommended)
  and validates via portfolio_validate before declaring success;
  symlink approach kept as legacy fallback

New skill (.claude/skills/split-portfolio/SKILL.md):
- 10-step migration with explicit operator-confirmation gates at each
  destructive step (force-push, body redaction, branch deletion)
- --verify mode: read-only state report (mode, paths, validate, drift)
- --dry-run mode: prints commands without executing
- Pre-flight refusals: already-private fork, paid GitHub plan, dirty
  working tree, already-migrated state
- Step 9 writes the portfolio: config block (not symlinks) — symlink
  fallback documented for adopters on older framework versions
- Step 9 surfaces the GitHub timeline-API survival caveat verbatim
- Idempotent re-runs: detects partial-migration state and resumes

Docs (docs/multi-project.md):
- Layout section describes both modes (config-block recommended,
  symlink legacy) with self-healing notes
- Setup steps split into config-block mode and legacy symlink mode
- Migration section now points at /split-portfolio skill; manual
  recipe preserved as fallback

AgDR (docs/agdr/AgDR-0010-portfolio-config-and-self-healing.md):
- Full Y-statement, options, decision, consequences, future phases
- Schema decision rationale: project-config.json over onboarding.yaml
  because runtime path resolution belongs in project-config

Closes #145
Refs #146 (delivered same PR; closed manually post-merge
per the single-Closes-keyword rule in validate-pr-create.sh)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(#145): markdownlint MD031 — blank lines around fences

CI's markdownlint-cli2 (v0.34.0) flagged the JSON + bash fenced code
blocks I added in setup/SKILL.md Step 2b without surrounding blank
lines. Added the required blanks. No content change.

Refs #147

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…ork auto-publish (#149)

The privacy-gate wording introduced in PR #144 (and unchanged in PR #147)
attributed the publication to the framework rather than the adopter:

  "the standard fork-and-commit setup will silently publish your private
   project names on a public GitHub repo"

That's factually wrong. ApexYard never pushes anything without explicit
operator approval — the publication only happens when the adopter
themselves runs git push. The "silently publish" framing read as if the
framework auto-publishes, which is misleading and undermines trust in
the rest of the framework's safety claims.

Two prose-only edits, no code, no behavior change:

- .claude/skills/setup/SKILL.md Step 2a — replaced "will silently
  publish ..." with adopter-action language ("you might accidentally
  publish ... a stray git push after registering them — I won't push
  without your approval, but the risk is on the adopter once the data
  is committed locally")

- docs/multi-project.md trip-wire callout — replaced "silently publish
  their portfolio names the moment they push" with "risk accidentally
  publishing their portfolio names with a stray push (the framework
  itself never pushes without operator approval, but once the registry
  is committed locally the next push exposes it)"

Verified: `grep -r "silently publish" .claude/skills/ docs/` returns no
hits.

Closes #148

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Closes the legitimate-bypass case (#150) and the
illegitimate-bypass case (#151) together so the
ticket-first gate is coherent. Shipping either alone would leave a
window where the framework is internally inconsistent — see AgDR-0011
for the full rationale.

Bootstrap exemption (#150):
  - .claude/session/active-bootstrap marker, written by /setup,
    /handover, /update, /split-portfolio on entry; cleared on exit
  - SessionStart sweep (clear-bootstrap-marker.sh) for stale markers
    from interrupted sessions
  - require-active-ticket.sh reads the marker and exempts skills on
    the configured ticket.bootstrap_skills list
  - bootstrap_skills list lives in .claude/project-config.defaults.json
    (extendable per fork via .claude/project-config.json)

Bash-write coverage (#151):
  - new _lib-detect-bash-write.sh — heuristic detector for output
    redirection, tee, sed -i, awk -i inplace, python/node/ruby
    embedded interpreters
  - require-active-ticket.sh + require-migration-ticket.sh now fire
    on Bash in addition to Edit|Write|MultiEdit
  - design choice: false-negatives preferred over false-positives
    (the matcher errs toward "let through" rather than block legit
    read-only commands)

Tests: 32 unit cases on the lib + 12 integration cases on the hook,
including the exact #151 bypass repro from the issue
body.

Closes #150

Will manually close #151 post-merge per the
single-Closes-per-PR rule (precedent: AgDR-0010 / PR #147).

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…155)

Closes #153.

Extends `_lib-detect-bash-write.sh` (introduced for #151
in PR #152) with the matcher families flagged by Rex's review of #152.
AgDR-0011 already frames the matcher as a living list extended on
observation; this commit just walks the list.

New matcher families:
- File-moving builtins: `cp`, `mv`, `rm`, `dd`, `install` (anchored at
  command-start; `--help`/`--version` and `git rm`/`git mv` excluded)
- Archive / network writes: `tar -x` / `tar --extract`, `curl -o` /
  `--output`, `wget -O` / `--output-document`
- Additional embedded interpreters: `perl -e`, `php -r` (keyword-gated
  like python/node/ruby); `go run`, `deno run`/`deno script.ts`,
  `bun run`/`bun script.ts` (categorical script runners)
- Python helpers: `pathlib.Path().touch()`, `shutil.copy*`,
  `shutil.move`, `os.rename` added to the `python -c` and python
  heredoc keyword list
- Heredoc variants for `ruby` and `node` (previously only python
  heredoc was covered)

Extractor extensions:
- `cp` / `mv`: last positional arg
- `curl -o` / `--output`: file argument
- `wget -O` / `--output-document`: file argument
- `tar -x`, `go run`, `deno`, `bun`, `perl -e`, `php -r`: return empty
  (caller applies gate categorically per AgDR-0011)

Test count rose from 32 to 86. Negative-class counterexamples cover
the trickiest false-positive surfaces: `tar -t` listing, `cp --help`,
`rm --version`, `git rm`, `curl` bare URL fetch, `wget` bare URL fetch,
`deno fmt`, `deno test`, `go build`. Existing
`test_require_active_ticket_bash.sh` regression suite still passes 12/12.

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…cy (#156)

- add _lib-mock-gh.sh helper that installs a fake `gh` on the sandbox PATH;
  intercepts `gh issue view <N> ... --json ...` and returns synthetic
  `{"number":N,"state":"OPEN"}` (overridable per-num via mock_gh_set_state)
- wire the shim into test_single_closes_per_pr.sh and
  test_validate_pr_required_sections.sh so the validator's CLOSED-issue
  refusal no longer breaks the suite when upstream issues are closed
- both files previously failed every case (0/13 and 0/8) because their PR
  titles reference #114 / #113, which are now CLOSED upstream
- post-fix: 13/13 and 8/8; full suite remains green

Closes #154

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>

…#158)

* feat(#132): structured CEO marker + same-turn merge in /approve-merge

Closes #132 (drop the "stop before merge" rule) and
#48 (harden CEO marker against self-approval bypass)
together. The two threads compose — see AgDR-0012 for the full
rationale.

Streamline (#132):
  - /approve-merge now runs `gh pr merge --squash --delete-branch`
    in the same turn as the marker write, by default
  - --no-merge opt-out preserves the deferred-merge case
  - The discrete approval moment is the SKILL INVOCATION, not a
    follow-up "now do the merge" message

Harden (#48):
  - CEO marker is now a structured key/value file with required fields:
      sha=<HEAD>
      approved_by=user
      skill_version=2
    Validated by block-unreviewed-merge.sh; bare-SHA legacy markers
    rejected with a clear "stale format" error pointing at /approve-merge
  - The model's bare `echo SHA > <pr>-ceo.approved` bypass is now
    mechanically rejected. Forging the structured fields requires a
    deliberate, visible rule violation rather than a one-line accident
  - Optional audit fields (approved_at, approval_summary) capture the
    "what did the user say when they approved" trail
  - Rex marker stays bare-SHA — different threat model (automated
    reviewer, not human authorization moment)

pr-workflow.md reframed: "the load-bearing rule is explicit per-PR
approval, not two user messages." The merge is a deterministic
consequence of the approval invocation.

Tests: 12 cases on the hardened hook covering the new format end-to-end
(valid v2, missing rex/ceo, bare-SHA legacy rejected, missing
approved_by, wrong approved_by, skill_version=1, sha mismatch,
non-merge no-op, gh-api shape gated). Full suite: 205/205 across 13
test files.

Will manually close #48 post-merge per the
single-Closes-per-PR rule (precedent: PR #152 / AgDR-0011).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* chore(#132): redact private project reference from AgDR-0012

Abstracted two references to a registered private project that named
the project's owner/repo. The leak-protection hook caught one in the
PR body; this fixup removes the matching references from the
AgDR-0012 file content (which would otherwise have shipped the names
to me2resh/apexyard public repo via the merge).

The pre-existing reference at .claude/rules/pr-workflow.md:130
(documenting #47) is untouched — it predates this PR and is already
on the public repo's history.

Refs #132.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* chore(#132): markdownlint blanks-around-fences + typo fix

Two small fixups against red CI / Rex feedback:

- AgDR-0012 line 63: fenced code block now has a blank line before it
  (MD031). markdownlint-cli2 0.13.0 was rejecting the indented fence
  inside the bullet because the fence's preceding line was the bullet
  text (no blank).
- approve-merge SKILL.md line 172: typo `deferes` → `defers` (Rex flag
  on PR #158).

Refs #132.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…#161)

* chore(#157): remove voice-prompts-on-pause feature + correct hook/skill counts

Closes #157 (sunset the voice-prompts feature) and
#77 (hook count off-by-one in CHANGELOG / CLAUDE.md)
in one bundled PR. AgDR-0013 supersedes AgDR-0009 with the full
rationale; both AgDRs are preserved (decision records are append-only
history).

Removed (#157):
  - .claude/hooks/voice-prompt-on-pause.sh
  - .claude/hooks/tests/test_voice_prompt_on_pause.sh
  - Stop matcher block in .claude/settings.json (became empty after
    voice removal)
  - voice_prompts block in .claude/project-config.defaults.json
  - "## Voice prompts" section in docs/project-config.md
  - voice_prompts mention in AgDR-0010 line 32 (replaced with
    leak_protection / ticket as still-current example config blocks)

Preserved:
  - docs/agdr/AgDR-0009-voice-prompts-on-pause.md — historical record;
    new "Superseded by: AgDR-0013" header at the top
  - AgDR-0010 line 115 reference to AgDR-0009 — still accurate as a
    historical pattern reference

Counts corrected (#77):
  - CHANGELOG.md v0.3.0 stats: "17 hooks" → "18 hooks" (historical
    fix — at v0.3.0 there were actually 18 hooks)
  - CLAUDE.md table line: "18 shell scripts" → "24 shell scripts"
    (current count after this removal)
  - CLAUDE.md table line: "35 slash commands" → "39 slash commands"
  - CLAUDE.md "Available skills (34)" → "Available skills (39)"
  - CLAUDE.md quick-reference "Skills (35 slash commands)" →
    "(39 slash commands)"

Why bundled: #77's correct count depends on whether voice is still in
the framework. Shipping #77 before #157 would write a number that's
wrong by one again the moment #157 lands. Same shape as previous
bundles (PR #152 / AgDR-0011, PR #158 / AgDR-0012).

Why no adopter-facing changelog mention of the voice removal: the
feature never reached a tagged release on main. v1.1.0 didn't have
it; v1.2.0 won't have it. From the adopter's perspective there's
nothing to retire. AgDR-0013 captures the framework's internal record
for future contributors. See AgDR-0013 § "No adopter-facing changelog
mention".

Tests: full hook test suite green (196 cases across 12 files —
test_voice_prompt_on_pause.sh removed). No regressions.

Will manually close #77 post-merge per the
single-Closes-per-PR rule (precedent: PRs #152, #158).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* chore(#157): unwire voice from settings + configs + docs + AgDR-0013

Continuation of d09d7b5 (the file deletions). Squash-merge will
collapse both into one PR commit. This commit captures:

- .claude/settings.json — Stop matcher block removed (was the only
  hook in it; entire matcher gone)
- .claude/project-config.defaults.json — voice_prompts block + its
  _comment removed
- docs/project-config.md — "## Voice prompts" section removed
- docs/agdr/AgDR-0009-voice-prompts-on-pause.md — "Superseded by"
  header added at the top, content otherwise preserved as history
- docs/agdr/AgDR-0010-portfolio-config-and-self-healing.md — line 32
  example reference swapped from voice_prompts to
  leak_protection / ticket (still-current config blocks)
- docs/agdr/AgDR-0013-sunset-voice-prompts.md — new supersession AgDR
- CLAUDE.md — hook count 18 → 24, skill count 35 → 39 (three
  occurrences each, all aligned to current reality)
- CHANGELOG.md v0.3.0 stats — "17 hooks" → "18 hooks" historical fix
  (#77 acceptance criterion 1)

Refs #157 + #77.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* chore(#157): markdownlint MD028 + resolve stale AgDR-numbering ref

Three small fixups against Rex CHANGES-REQUESTED on PR #161:

- AgDR-0009 line 3: promote "Superseded by:" header out of a
  blockquote. The original "I decided ..." canonical blockquote at
  line 5 was being merged with the new supersession blockquote
  (markdownlint MD028 — "no blanks inside blockquote").
- AgDR-0013 line 3: same shape — "Supersedes:" header now a plain
  bold paragraph, canonical "I decided ..." blockquote untouched.
- approve-merge SKILL.md line ~187: stale conditional reference
  "AgDR-0012 (or 0013 — depends on whether voice-removal lands
  first)" — order is now resolved (12 = approve-merge bundle, 13 =
  voice removal). Drop the parenthetical.

Refs #157 + #77.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Landing-site `site/index.html` terminal demo previously played one
canonical flow ("one ticket, start to finish") on autoplay. With the
v1.2.0 skill surface expansion (4 new skills, 8+ new hooks), one flow
no longer represents the framework's breadth.

Adds tabs to the terminal chrome — four flows visitors can either let
auto-cycle or click directly:

  1. one ticket — existing flow, unchanged content
  2. /handover  — adopt an external repo into the portfolio
  3. /setup     — first-run framework bootstrap on a fresh fork
  4. /fan-out   — spawn 3 parallel agents on independent tickets

Auto-advance: on completion of the active tab's script, the demo
pauses ~1.8s then advances to the next tab. Loops at the end. User
can interrupt by clicking any tab or hitting Replay.

Implementation:
- HTML chrome — single title span replaced with a tablist of 4 button
  tabs. ARIA `role="tablist"` / `role="tab"` / `aria-selected` so
  keyboard + screen-reader users get the same semantics as sighted
  ones.
- CSS — new `.shell-demo__tabs` + `.shell-demo__tab` with an accent
  underline on the active tab. Tabs scroll horizontally on narrow
  viewports (mobile responsive).
- JS — refactored the existing IIFE from one `script` array to an
  array of four. Added `setActiveTab()` for ARIA state, `play(idx)`
  takes a tab index, end-of-script auto-advances to `(idx + 1) % N`
  unless the user clicked away during the pause. Adds a new `cmd`
  type alongside `you` for slash-command invocations (renders with
  the same `>` prompt prefix). prefers-reduced-motion still bails
  early and leaves the static seed visible.
- Static seed (the non-JS fallback) still shows tab 0's content, so
  reduced-motion / no-JS visitors see the one-ticket flow as before.

Hero metrics also corrected to current reality (#77 / PR #161 covers
CLAUDE.md and CHANGELOG.md; this PR catches the same numbers in the
landing site):

  Skills    32 → 39
  Hooks     18 → 24

The tabs ship in v1.2.0 alongside the framework changes that make
the new flows worth showcasing.

Refs #160 (release v1.2.0 + landing-site refresh).

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…portfolio (#164)

Closes #163.

The split-portfolio mode docs and skills previously suggested
`your-org/ops` as the default name for the private sibling repo, and
`portfolio/` as the local clone directory. Both too generic — adopters
running multiple ops setups end up with `your-org/ops` collisions, and
a bare `portfolio/` dir gives no signal about which framework it
belongs to when it sits next to other unrelated `portfolio/` dirs.

`<fork>-portfolio` is now the default — keeps the relationship to the
public fork explicit on disk and on GitHub. If the fork is named
`your-org/apexyard`, the portfolio defaults to
`your-org/apexyard-portfolio`. If the fork was renamed (e.g. `cos`),
the portfolio defaults to `cos-portfolio`. Adopters with custom names
keep working — the `portfolio:` config block resolves whatever path
they configured.

Files updated:

  docs/multi-project.md
    - Layout diagrams use `apexyard-portfolio/` as the sibling
    - Setup walkthrough Step 2 + Step 3 use `your-org/apexyard-portfolio`
      and explain the `<fork>-portfolio` pattern
    - Config-block + symlink path examples updated to
      `../apexyard-portfolio/...`
    - Daily workflow + cross-machine clone commands updated
    - The two existing `your-org/ops` references that remain are
      fork-rename examples (lines 52, 64) — kept as-is, since renaming
      the fork to `ops` is still valid (the portfolio would then
      default to `ops-portfolio`)

  .claude/skills/setup/SKILL.md
    - Step 2b's "default suggestion" for the private repo name is now
      `your-org/<fork>-portfolio`, computed dynamically from the
      fork's repo name via `gh repo view --json name -q .name` so the
      suggestion is correct even when the fork was renamed
    - Clone command no longer needs a second arg — the repo name IS
      the directory name
    - Config-block paths updated to `../apexyard-portfolio/...`

  .claude/skills/split-portfolio/SKILL.md
    - Step 3's suggested-name template is now `<account>/<fork>-portfolio`
      with the same dynamic-fork-name resolution

Mechanism unchanged. The `portfolio:` config block in
`.claude/project-config.json` still takes any path; this PR is purely
default-suggestion + example prose.

No tests required (skills are markdown instructions; no automated
coverage today).

Refs `apexyard.projects.yaml.example` uses "ops repo" as a generic
term meaning "the operational management fork" — not a name — kept
as-is.

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

#167)

Closes #165.

Adds a public, browseable index of every apexyard slash command
alongside a one-click changelog link from the homepage nav.

site/skills.html (new):
  - Lists all 39 skills currently shipping in .claude/skills/
  - Each entry: slash command, argument hint, description (taken
    verbatim from the SKILL.md frontmatter so the page matches the
    runtime exactly)
  - 10 categories: Setup & onboarding, Daily ops, Tickets & ideas,
    Specs & decisions, Code review & merge, Architecture & dev tools,
    Production-readiness audits, Workflow primitives, Communications,
    Deprecated
  - Same brutalist-terminal design tokens as the homepage — JetBrains
    Mono, paper-cream background, single warning-red accent, sharp
    corners. Inlined CSS to keep the static-only no-build-step
    convention; design vars duplicated rather than extracted to a
    shared file (~18 vars; cheap to keep in sync).
  - Mobile responsive — skill grid collapses to single-column under
    720px; titlebar nav hides non-CTA items on narrow viewports.
  - Reduced-motion friendly (no animation in the first place).
  - Internal anchor TOC at the top so the page scans in seconds.

site/index.html (nav addition):
  - Added two nav links to the titlebar between "what's in the box"
    and the github CTA:
      • skills      → ./skills.html
      • changelog → https://github.com/me2resh/apexyard/releases
  - The changelog link points at the GitHub releases page (not the
    raw CHANGELOG.md file) so it auto-resolves to the latest tagged
    release on each visit. v1.2.0 lands and the link is already there.

No new dependencies, no build step, no JS for the skills page. The
existing site convention (one-html-file-per-route, inlined CSS,
optional progressive-enhancement JS) is preserved.

Refs #160 (release v1.2.0 + landing-site refresh) —
this is the second site-side deliverable for that ticket; the release
tag itself follows once #159 (testing) closes.

Follow-up worth a separate ticket: a small generator script that
walks .claude/skills/*/SKILL.md, parses YAML frontmatter, and emits
the skills.html sections automatically. Out of scope for v1.2.0 —
first version is hand-curated and will need maintenance until that
generator lands.

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…169)

Closes #168.

The /release skill prescribed `release/vA.B.C` as the source-branch
name and `release: vA.B.C` as the PR title for the dev → main release
PR (per AgDR-0007). Both were rejected by the framework's own
validators:

  validate-branch-name.sh required {type}/{TICKET-ID}-{description};
  release/v1.2.0 has no ticket-id portion.

  validate-pr-create.sh required type(SCOPE): form with `release` not
  in pr.title_type_whitelist.

The contradiction surfaced cutting v1.2.0 — the first release under
the dev/main model.

Three small changes:

1. .claude/hooks/validate-branch-name.sh — added an early-out branch
   that accepts ^release/vN.N.N(-rcN)?$ as a valid name. Narrow,
   intentional exception for the framework's release-cut convention;
   release branches don't carry a ticket-id because the release itself
   IS the ticket.

2. .claude/project-config.defaults.json — added "release" to
   pr.title_type_whitelist so a title like `release(#160): v1.2.0`
   passes validate-pr-create.sh's existing regex unchanged.

3. .claude/skills/release/SKILL.md step 4 — corrected the prescribed
   PR title to `release(#<release-ticket>): vA.B.C` so future /release
   invocations produce a title that satisfies the validators by
   construction.

Tested:

  bash .claude/hooks/validate-branch-name.sh against:
    release/v1.2.0           → 0   (allowed, release-special-case)
    release/v1.2.0-rc1       → 0   (allowed, RC variant)
    release/v9.9.9           → 0   (allowed)
    release/foo              → 2   (correctly blocked)
    release/v1               → 2   (correctly blocked)
    chore/GH-168-fix         → 0   (allowed, standard pattern)
    feature/GH-1-x           → 0   (allowed, standard pattern)

  Full hook test suite: 196/196 cases green across 12 test files.

Refs: surfaced 2026-05-04 cutting the first release under AgDR-0007.

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…d check (#171)

Closes #170.

Completes the work started in #169 (closing #168). #169 added a
release-pattern early-out to validate-branch-name.sh so release/vN.N.N
branches pass the branch-name validator. But validate-pr-create.sh has
its own independent branch-id check at line 273 that #169 didn't
touch — and it still rejects release/v1.2.0 because that name doesn't
contain a ticket-id substring.

This is the same class of contradiction #168 fixed; the fix is the
same shape. Add the same release-pattern early-out to the branch-id
check in validate-pr-create.sh:

  - if the branch matches ^release/vN.N.N(-rcN)?$ → exempt (release
    branches don't carry ticket-ids; the release itself is the ticket)
  - otherwise → require a ticket-id substring as before

#168's acceptance criterion 3 ("validate-pr-create.sh accepts a PR
title `release(#160): v1.2.0` against the `release/v1.2.0` branch")
was checked off based on the title regex alone but didn't catch the
secondary branch-id check living in the same file. Surfaced trying
to open the v1.2.0 release PR.

Tested:

  bash .claude/hooks/validate-pr-create.sh against:
    release/v1.2.0           → 0   (allowed, exempt)
    release/v1.2.0-rc1       → 0   (allowed, RC variant)
    chore/GH-1-fix           → 0   (allowed, has ticket-id)
    release/foo              → 2   (correctly blocked)
    chore/no-ticket          → 2   (correctly blocked)

Refs #168 (the parent bug) + #169 (the partial fix).

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Closes #173.

The release-cut model (AgDR-0007 / #116) squash-merges
release/vN.N.N → main, which means the v1.2.0 CHANGELOG section that
landed on main via PR #172 was never propagated back to dev. This
commit copies main's CHANGELOG.md verbatim onto dev so the v1.2.0
section is now present on both branches.

The diff is exactly the v1.2.0 entry being prepended; no other lines
change.

Without this sync, the next release PR cut from dev would build a
v1.3.0 section on top of v1.1.0, silently dropping v1.2.0 from dev's
running history. The corollary skill-level fix (option B in the
ticket) — updating /release to source the previous CHANGELOG from
upstream/main — is filed as a follow-up and out of scope for this PR.

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Adds /agdr — a portfolio-wide index for Agent Decision Records.
Walks apexyard.projects.yaml, reads each project's docs/agdr/*.md
(local clone if available, else gh api fallback), parses the optional
YAML frontmatter for category + projects, and answers four queries:

- /agdr browse        list across the portfolio, grouped by category
- /agdr search <term> full-text grep across all bodies, returns
                      <project>/AgDR-NNNN paths + matching paragraph
- /agdr show <id>     print a specific record, disambiguates duplicates
- /agdr stats         counts per category (the marketing-slide tile,
                      now backed by real data)

Six-category taxonomy: architecture | tech-stack | security | patterns
| integrations | other. Legacy AgDRs without frontmatter remain
first-class — they bucket as `other` and are flagged in browse so
operators can migrate at their own pace.

Backwards-compatible template change: templates/agdr.md gains an
optional `category:` (and optional `projects:`) line in the existing
frontmatter block. Omitting the line keeps every existing AgDR valid;
the skill defaults to `other` when missing.

Doc note in workflows/sdlc.md § Phase 2 points at /agdr search for
"have we decided this before?" lookups before drafting a design.

Smoke test in .claude/hooks/tests/test_agdr_skill.sh covers the
parser the spec specifies — frontmatter extraction, category
bucketing (including the legacy default-to-other path), id reading,
stats aggregation, and search match counts. 18 assertions, all green;
12 pre-existing test suites also green (no regressions).

Closes #181

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

…session cwd (#580)

* fix(#549): block-main-push checks the operation's target branch, not session cwd

The hook previously resolved the current branch via `git branch
--show-current` in the hook's cwd (the harness primary checkout). When
the primary checkout sat on a protected branch (e.g. `dev`) while the
actual operation targeted a feature-branch worktree, the hook
false-blocked legitimate commits and pushes.

Push path: reuse `_lib-extract-push-ref.sh` (the same lib
`validate-branch-name.sh` already uses for the same worktree-cwd
problem — #194, #547) to read the DESTINATION branch directly from the
push command. Tag pushes remain no-ops. Falls back to local HEAD only
when no explicit ref is present in the command (bare `git push`
relying on upstream tracking).

Commit path: detect the `cd <path> && git commit` compound shell
pattern and resolve the branch of the TARGET worktree via
`git -C <path> branch --show-current` instead of the session cwd.
Plain `git commit` (no `cd` prefix) retains the original behaviour.

Tests: 26 cases covering (a) push to protected still blocks,
(b) push to feature passes when cwd=dev, (c) tag push passes,
(d) cd wt && commit on feature-branch wt passes while cwd=dev,
(e) plain commit on protected still blocks.

Refs #549

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(#549): strip cd-path quotes + resolve last cd target (review fix)

Addresses the #580 Rex review:
- block-main-push.sh now strips surrounding quotes from the `cd <path>`
  target before `git -C`. Without this, `cd "path" && git commit` passed
  quotes to git -C → error → empty branch → the protected-branch check was
  skipped and a commit into a protected worktree slipped through (a
  false-negative in the dangerous direction).
- Resolves the LAST `cd` in a chain (cd a && cd b && commit → b), fixing the
  secondary over-block Rex noted.
- 5 new tests: quoted (double/single) cd into protected → blocks; quoted cd
  into feature → passes; chained cd resolves last target both directions.
  Suite now 31/31.

Refs #549

* test(#549): drop dead WT_PROTECTED scaffolding (shellcheck SC2034)

The original test left abandoned WT_PROTECTED + wt-main worktree setup
from a first approach; the active assertion uses a secondary sandbox
(SB2). shellcheck flagged WT_PROTECTED as unused (SC2034), failing CI.
Removed the dead lines.

Refs #549

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

…rgs (#581)

- Strip fd-to-fd (`2>&1`), combined (`&>`), append (`>>`), and overwrite
  (`>`) redirections from the `gh pr merge` span before scanning for the
  PR number — a digit that is part of a redirect operator can no longer
  be mistaken for a PR number.
- Require the first post-`merge` token to be a bare integer; if it is an
  unexpanded shell variable (`$pr`, `${PR_NUMBER}`) return empty so the
  existing `gh pr view` fallback resolves the PR, rather than latching
  onto a stray digit (the `2` in `2>&1`).
- Use `grep -oE` for word-boundary matching after `merge` instead of
  BSD-sed `\b` which is not portable to macOS.
- Add `test_extract_pr.sh` — 14 focused unit tests for `extract_pr_number`
  covering the bug repro, all redirect forms, unexpanded variables, the
  `gh api` URL path shape, and edge cases (no number, non-merge command).

Live repro (issue #568): `gh pr merge $pr --squash 2>&1 | tail` with
$pr unexpanded at hook-eval time caused `block-unreviewed-merge.sh` to
block with "PR #2 has no recorded code-reviewer approval" — the `2` came
from the stderr redirect, not from a PR.

Refs #568

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

…e ticket gate (#582)

* fix(#569): exempt .claude/, /tmp, rm, and \$VAR targets from bash-write ticket gate

require-active-ticket.sh applied the ticket gate categorically to all
Bash writes detected by _lib-detect-bash-write.sh, without applying the
same target-aware path exemptions already used for Edit/Write/MultiEdit.
This over-blocked four common workflow commands:

  - `cat > "\$CEO"` (variable target resolving to .claude/ path)
  - `cat > /tmp/commit-msg.txt` (absolute path outside the repo tree)
  - `cp .claude/session/... .claude/session/...` (already exempt dest)
  - `rm -f workspace/proj/.git/tmpfile` (deletion, adds no content)

Fix: three targeted exemptions added to the Bash path of
require-active-ticket.sh, reusing the existing exempt-path logic:

  1. Deletion-only commands (`rm` with no content-writing sibling) exit 0
     immediately via the new `bash_command_is_deletion_only` helper added
     to _lib-detect-bash-write.sh — deleting files never adds tracked
     content so a ticket requirement is pure friction.

  2. Variable targets (`\$VAR`, `\${VAR}`) are unresolvable statically;
     treating them as non-blocking avoids false positives when the variable
     holds an exempt path (the canonical .claude/ marker-write case).

  3. Absolute paths not under REPO_ROOT (e.g. /tmp/, /var/tmp/) are
     outside the tracked source tree and cannot affect apexyard-governed
     content — exempted after the REL_PATH strip fails.

Writes into non-exempt tracked source paths (outside .claude/, docs/,
*.md) are still blocked when no ticket is set — the gate is tightened,
not loosened for the cases that matter.

Tests: 10 new cases in test_require_active_ticket_bash.sh (the four #569
repros + 6 regressions); 11 new cases in test_detect_bash_write.sh for
bash_command_is_deletion_only positive and negative classes.
Both test suites: 97/97 and 26/26 passing.

Refs #569

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(#569): close the variable-target bypass (review fix)

Addresses the #582 Rex review: the blanket `$*` exemption let a write
into tracked source slip past the ticket gate whenever the target began
with `$` (e.g. `echo x > $PWD/src/app.ts`, `> $D/app.ts`) — fail-open on
a security gate.

- Replace the broad `$*) exit 0` with: exempt only a temp-dir var
  (`$TMPDIR`/`$TMP`) and a BARE whole-target variable (`$CEO`,
  `${marker}` — no path tail). A variable WITH a path tail is neither
  bare nor absolute, so it falls through to the gate and blocks.
- Do NOT expand $PWD/$HOME (adds nothing for blocking; tripped a
  /var↔/private/var symlink mismatch).
- 3 new regression tests: `> $PWD/src/app.ts` and `> $D/app.ts` BLOCK;
  bare `${marker}` exempt. Suite 29/29.

Refs #569

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

#586)

* feat(#585): add interactive LLM game to the site with social score-sharing

Ships the 'LLM Field Guide' interactive game (self-contained, no build)
at /game, with an end-of-game 'Share your score' block.

- site/game.html: the game + SEO/OG/Twitter meta (twitter:site=@me2resh,
  canonical, summary_large_image). renderOutro now offers Post on X
  (via=me2resh), LinkedIn, WhatsApp, and Copy link — each carrying the
  score, a link back to the game, and 'via @me2resh' (X uses its native
  via param). Plus a quiet back-to-ApexYard link.
- Discovery: 'play the game' added to the primary nav + footer across
  index / how-it-works / architecture / skills, the 404 quick-links, and
  a homepage CTA (button + line in the final section).
- /game clean URL via _redirects; /game added to sitemap.xml.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(#585): rebrand the game to 'You vs. the LLM' + ApexYard branding

Make the game catchier and clearly ApexYard-branded:
- Header now leads with the AY ApexYard mark + 'ApexYard · You vs. the
  LLM' (was an unbranded 'Field Guide · LLMs & Agents'); logo links home.
- Intro hero: 'You vs. the LLM' + a challenge hook ('Think you understand
  how AI actually works? Prove it.'); start button 'Take the challenge'.
- End screen + page title + OG/Twitter + share copy all renamed to
  'You vs. the LLM — an ApexYard game'; share line is now a beat-my-score
  hook. Homepage CTA/teaser + 404 updated to match.
- Folded in the two review a11y nits: share-label var(--faint) to
  var(--dim) for AA contrast; underline the back-to-ApexYard link.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(#585): polish homepage CTA label (drop awkward colon)

Nour's editorial nit on PR #586: 'play: You vs. the LLM' reads awkward;
use an en-dash. 'play — You vs. the LLM'.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…t dev (#589)

Dependabot bump PRs (and release-sync merges) carry no ticket ID by
design, so the 'Verify Ticket ID' CI failed on every one (#539-#543),
blocking clean dependency patches. The local hooks already exempt sync/
(#545); give the CI check the same exemption.

- pr-title-check.yml: early-return success when the PR head ref matches
  dependabot/ or sync/. Normal feature PRs still require a ticket.
- dependabot.yml: target-branch dev, so future bumps follow the dev-first
  flow instead of targeting main.

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Bumps [gitleaks/gitleaks-action](https://github.com/gitleaks/gitleaks-action) from 2 to 3.
- [Release notes](https://github.com/gitleaks/gitleaks-action/releases)
- [Commits](gitleaks/gitleaks-action@v2...v3)

---
updated-dependencies:
- dependency-name: gitleaks/gitleaks-action
  dependency-version: '3'
  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>

Bumps [actions/github-script](https://github.com/actions/github-script) from 7 to 9.
- [Release notes](https://github.com/actions/github-script/releases)
- [Commits](actions/github-script@v7...v9)

---
updated-dependencies:
- dependency-name: actions/github-script
  dependency-version: '9'
  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>

Bumps [DavidAnson/markdownlint-cli2-action](https://github.com/davidanson/markdownlint-cli2-action) from 16 to 23.
- [Release notes](https://github.com/davidanson/markdownlint-cli2-action/releases)
- [Commits](DavidAnson/markdownlint-cli2-action@v16...v23)

---
updated-dependencies:
- dependency-name: DavidAnson/markdownlint-cli2-action
  dependency-version: '23'
  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>

Manual application of dependabot #539 (upload-artifact) and #541 (codeql)
against dev — their branches were cut from main and conflict with dev's
diverged workflows, and dependabot can't target dev until the config
(#588) reaches main via a release. Supersedes #539/#541.

- upload-artifact@v4 -> v7: extract-subpacks-on-release, security-scan, scorecard
- codeql-action @V3 -> v4: codeql (init+analyze), scorecard (upload-sarif)

Closes #590

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* feat(#594): governed looping — loop-mode trigger rule + AgDR

Adopt looping as a first-class, governed apexyard pattern (not a new
engine): a trigger-heuristic rule that proactively recommends a closed
loop when work fits, bound to apexyard's existing gates as the eval.

- .claude/rules/loop-mode.md: sibling of parallel-work.md / plan-mode.md.
  When to OFFER (repetitive over a set · machine-verifiable · bounded),
  when not to, which primitive (/loop · /fan-out · Workflow), and the
  guardrails — halt at the per-PR CEO merge gate (never self-approve),
  verify = build + tests + Rex (not just build), budget/iteration ceiling,
  no-progress detection, call skills not re-derived prompts.
- AgDR-0068: records the decision (options, the ReAct→orchestration
  lineage, the in-session lesson that an unverified loop ships broken
  tests). Auto-loads via the @.claude/rules/*.md import.
- CLAUDE.md + README.md: rule count 11→12, list + a code-standards bullet.

Dogfooded: the discovery for this feature was itself run as a 4-agent
discovery loop (Workflow).

Closes #594

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* docs(#594): clarify the verify-stage guardrail is advisory (Tariq review)

Tariq's architecture-review note: distinguish the mechanically-enforced
halt (CEO merge gate) from the verify=build+tests+Rex clause, which is
recommended loop discipline caught at the merge boundary (red-CI + Rex),
not blocked mid-iteration.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Squash-merge divergence from the v3.1.0 release PR creates phantom divergence
between main and dev. This merge makes the v3.1.0 squash commit an ancestor
of dev so future dev->main release PRs only see genuinely-new commits.

Strategy: -X ours (dev wins on conflicts) — correct because dev already has the
un-squashed equivalents of everything in the squash commit.

Refs #403

The -X ours merge kept dev's site/index.html, which lacks the version-string
bumps written on main during the v3.1.0 release flow (hero pill apexyard v3.1,
releases-shipped range v0.1 → v3.1). dev ended up with softwareVersion=3.1.0
(non-conflicting, merged from main) but pill=v3.0 (dev won the conflict region),
which is internally inconsistent and fails test_site_counts.sh.

This commit restores main's site/index.html so dev tracks the full release
version-state forward. Same shape as the CHANGELOG carry-forward (step 5b /
apexyard#448) — site/index.html is a second "main-leads" file now that /release
bumps it (#493/#562). main is a superset of dev's site here (the release branch
was cut from dev + the version bumps), so nothing is lost.

Refs #448

sync(#403): main→dev after v3.1.0 release

The floating @v2 major tag no longer resolves ("unable to find version v2"),
failing the Scorecard workflow on every push to main (seen on the v3.1.0
release merge, run 27188194993). Pin to the latest published release v2.4.3.

Closes #600

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…eering level (#603)

- New capstone level "Engineer the loop" (levelLoop): toggle the guardrails that
  make a loop safe to run unattended (verify w/ tests+review, halt at human gate,
  budget+iteration ceiling, stop-on-no-progress, call named skills) and avoid the
  traps (self-approve merges, build-only verify, run unbounded). Reuses the cost-*
  toggle/meter UI; pass = all guardrails on, no trap enabled. 11 levels now.
- Universal share message: "via @me2resh" -> "via #apexyard" across X / WhatsApp /
  Copy (dropped the X-only via= param; hashtag works as a brand tag everywhere).
  Card meta (twitter:site/creator) left as-is.
- Mobile responsiveness: extended the @media(<=640px) block with the widget-level
  fixes the game lacked (RAG paper sheet, prompt-builder slots stack, knowledge-
  cutoff cards full-width, hallucination/injection docs, cost+loop toggles) plus a
  global overflow-x guard and a tighter <=380px pass. Touch was already handled
  (pointer events). Every round now fits a ~360-390px phone.
- Bumped OG/Twitter "10 rounds" -> "11 rounds".

Closes #602

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Adds a persistent SKIP button in the footer (next to HINT), shown during every
level's play phase and hidden on teaching slides / intro / outro / after submit.
Tap-to-confirm guard (first tap arms + warns, second tap within 4s skips) so a
stuck player can't zero a level by accident. On confirm: records 0 for the level
and calls advance() to the next concept (or outro on the last level). Reuses the
.hint-btn styling and the universal advance() flow — no per-level edits, no new
behaviour in the 11 render functions.

Closes #606

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Make the v3.1.1 squash commit an ancestor of dev so future dev->main release
PRs only see genuinely-new commits. -X ours: dev wins on conflicts (it already
has the un-squashed equivalents).

Refs #403

sync(#403): main→dev after v3.1.1 release

…p (mobile) (#610)

The Embeddings level used free-form drag of a word from a bottom tray into a tall
2D plane — unplayable on a phone (invisible cluster zones, tray far below the drop
area, 2D drag doesn't work with touch+scroll). Replaced the mechanic with a
tap-the-meaning-group picker: one card per word + Animals/Fruits/Vehicles buttons,
mirroring the mobile-proven levelCutoff pattern and reusing its .cut-* styling.
Same lesson (similar meanings group together), no dragging; score = correct
group per word. Removed the now-dead .emb-*/cluster-* CSS (incl. the @media rules).
JS node --check clean; LEVELS still 11.

Closes #609

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Make the v3.1.2 squash commit an ancestor of dev. -X ours: dev wins on conflicts.

Refs #403

sync(#403): main→dev after v3.1.2 release

)

state.score could exceed the 1100 max (observed share text "1188/1100") and
didn't match the per-level breakdown. Two levels called addScore() per round
(summing into state.score) but recorded only the average in levelScores:

- level1 (Token splitter): drop the per-round addScore(rs); call addScore(avg)
  once at finish, after levelScores is set — so the level contributes its
  recorded 0-100 exactly once (matching every other level / level6's pattern).
- levelTemp (Temperature dial): same per-round fix, plus fix a double-division
  (pts = rs/3 then avg = totalScore/SCEN.length capped the level at ~33 even on
  a perfect play). Now accumulate totalScore += rs so avg is the true average
  and a perfect run scores 100.

Result: every level adds exactly its 0-100 to state.score once; total maxes at
1100 and matches the breakdown. node --check clean; LEVELS still 11.

Closes #614

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

- Add the v3.1.3 CHANGELOG entry — patch: game scoring fix (#614).
- Bump site version strings: softwareVersion 3.1.2 -> 3.1.3, version link
  -> v3.1.3, releases-shipped 17 -> 18 (pill + range stay v3.1).

Closes #616

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>