* feat(#232): adopter handbooks consumed by Rex during code review

Introduces a `handbooks/` directory at the ops fork root for
adopter-authored coding standards that Rex consults during PR review,
layered on top of the framework's generic rules in `.claude/rules/`.
Discovery is by path-convention (no YAML frontmatter); enforcement
defaults to advisory with explicit blocking opt-in via the literal
`ENFORCEMENT: blocking` marker phrase at the top of a file.

Components:

- handbooks/README.md — index + discovery rules + advisory/blocking
  convention + authoring guidance + the four sample-handbook table
- handbooks/architecture/clean-architecture-layers.md — DDD layering
  (advisory, always-load)
- handbooks/architecture/migration-safety.md — schema-migration
  backwards-compat rule (BLOCKING, always-load) — demonstrates the
  enforcement opt-in shape
- handbooks/language/typescript/strict-mode.md — TS strict-mode + no
  bare `any` (advisory, loads on **/*.{ts,tsx} diff match)
- handbooks/general/commit-message-quality.md — commit-message WHY-
  not-WHAT discipline (advisory, always-load)
- .claude/agents/code-reviewer.md — new § 8 "Adopter Handbooks":
  discovery (path-convention), enforcement (advisory vs blocking
  marker), what-to-surface, Handbook Findings section in the review
  output template, new Rules entry #9 codifying the layer
- CLAUDE.md — Quick Reference + framework integration table updated
  to surface handbooks/ alongside .claude/rules/

Decision rationale: docs/agdr/AgDR-0020-adopter-handbooks-for-rex.md
captures all five load-bearing decisions (scope, discovery,
format, sample-scaffold-per-feature, enforcement default) with
options-considered tables.

Closes #232

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

* fix(#232): correct rule-file count in CLAUDE.md (10 → 11)

Rex flagged the count mismatch in his review of #233. There are
actually 11 rule files in .claude/rules/ (the previous list also
omitted leak-protection). One-line fix.

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

* fix(#232): markdownlint MD028/MD032 in handbook samples

Three handbook samples tripped markdownlint:
- migration-safety.md: list-inside-blockquote needs a blank `>` line
  before the list (MD032)
- commit-message-quality.md: two adjacent blockquotes joined with `>`
  instead of a blank line (MD028)
- strict-mode.md: same fix for three adjacent sample-finding blockquotes
  (MD028 x2)

Pure formatting fix. No content change.

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>

* chore: narrow qa.exempt_labels to qa-bypass only

QA-chain v1 (AgDR-0031) defined a 5-label exempt set (chore, docs,
spike, infra, qa-bypass) on the rationale that "infra-class work with
no user-facing AC doesn't need Salim's AC walk". That conflated content
(no user-facing AC) with verification value (the second pair of eyes).
Concrete failure: kt-sdk #9 (LICENSE) and #10 (uv lock) closed via the
chore exemption today without Salim ever touching them. Auto-closed on
merge; qa-gate.yml allowed it; chain held by design but the design was
wrong about what to exempt.

New policy: every ticket flows through Salim. qa-bypass is the sole
class-level exempt label — applied per-ticket, deliberately, on the
rare cases where verification truly isn't valuable (legal-mandated
hotfix, framework self-bootstrap, etc.).

Changes:
  - .claude/project-config.defaults.json — .qa.exempt_labels = ["qa-bypass"]
  - Both hook fallback strings: chore|docs|spike|infra|qa-bypass → qa-bypass
  - golden-paths/pipelines/qa-gate.yml — env default + comment + header
  - .claude/rules/git-conventions.md — exempt-label table narrowed
  - .claude/rules/pr-workflow.md — exempt-list mention updated
  - .claude/rules/workflow-gates.md — diagram + enforcement table
  - roles/engineering/qa-engineer.md — Exempt Tickets table narrowed

Tests:
  - test_block_closes_without_exempt.sh — flipped chore/spike ALLOW to
    BLOCK; added docs / infra BLOCK; kept qa-bypass ALLOW. 10/10 PASS.
  - test_block_issue_close_without_qa.sh — same flip + new multi-num
    test 42=qa-bypass,43=qa-passed → ALLOW. 21/21 PASS.

Decisions captured in docs/agdr/AgDR-0032-narrow-qa-exempt-to-qa-bypass-only.md
(supersedes AgDR-0031 § Vocabulary re: the exempt-set membership; rest
of AgDR-0031 stays in force).

Closes #3

* chore: address Rex nits — narrow exempt set in 4 stale strings

Rex flagged 4 user-facing strings that still described the OLD 5-label
exempt set after AgDR-0032's policy shift. The hook behavior was
correct but the error messages would have told operators to apply
chore / docs / spike / infra — labels the same hook then blocks.

Fixes:
  - block-closes-without-exempt-label.sh
    - line 11: header comment narrowed
    - lines 163-176: error message recommends gh issue edit
      --add-label qa-bypass (was --add-label chore + 5-label list)
  - block-issue-close-without-qa-passed.sh
    - lines 144-165: same narrowing in close-time error message
  - golden-paths/pipelines/qa-gate.yml
    - lines 69, 73: reopen-comment body (this text gets POSTED to
      every reopened GitHub issue — externally visible) narrowed to
      qa-bypass + AgDR-0031 + AgDR-0032 references

Plus non-blocking nits Rex flagged:
  - docs/agdr/AgDR-0031-qa-chain-mechanization.md Vocabulary section:
    forward-pointer note to AgDR-0032 (supersession was one-way
    recorded; now bi-directional)
  - .claude/rules/git-conventions.md: per-project override discipline
    note — widening the exempt set via override should be a
    documented, time-bounded AgDR-backed decision

Hook tests still 10/10 + 21/21 PASS. yaml validates clean.

Refs #3

---------

Co-authored-by: Omar Elaraby <omar.elaraby@allegiancemd.com>

…d gaps) (me2resh#9)

Closes the last sub-task of the apexstack positioning epic (me2resh#100).
The Explore agent audit flagged specific gaps in both skills; these
were deferred through PRs me2resh#4-me2resh#8 because they're independent of the
positioning work. Now fixed.

/handover — 3 gaps closed
-------------------------

1. **Auto-append to the portfolio registry** (was the CRITICAL gap
   from the audit). Previously the skill printed a YAML snippet
   and told the user "add this to apexstack.projects.yaml". Users
   forgot, or miscopied, or broke the YAML indentation.

   Fix: new step 6 prompts the user `Ready to add {name} to
   apexstack.projects.yaml? (y/n)`. On y, the skill:
   - Locates the registry at the ops-repo root (creates from
     .example if missing, with a warning)
   - Appends the entry via `yq` if available, otherwise plain-text
     append with careful indentation
   - Validates the result by parsing the YAML (yq or python yaml)
   - Rolls back on failure (never leaves the registry broken)
   - Confirms the append to the user with the derived roles list

   On n, prints the snippet for manual copy-paste and continues
   without writing.

2. **Dynamic "Next Steps"** derived from the risks found. The
   old template had generic placeholders like `{first concrete
   action — usually "run /audit-deps and triage criticals"}`
   which was embarrassing to ship.

   New behaviour: a mapping table in the skill spec defines the
   derivation. Examples:
   - CVEs detected → `/audit-deps {name} — triage the {severity}
     {package} CVE before any new feature work`
   - Failing tests → `Fix the {N} failing tests in {module}
     before merging new PRs`
   - No observability → `/decide on observability ({top 2 options
     for this stack})`
   - Stale CI → `Re-enable CI — copy in golden-paths/pipelines/ci.yml`
   - Coverage unknown → `Set up test coverage reporting`
   - Backlog > 10 → `Triage the issue backlog with previous owner`
   - Missing README → `Write a minimum-viable README`

3. **Post-handover checklist** added to the assessment template,
   also dynamically tailored to the risks found. Items like "close
   {top risk} before the first feature PR", "add {name} to weekly
   /stakeholder-update rollup", "set up a coverage baseline".

Bonus gap closed: **derive the role list dynamically** from the
tech stack + CI config + security surface, instead of hard-coding
`[tech-lead, backend-engineer]`. Mapping table in the spec:
- backend deps → backend-engineer
- UI code → frontend-engineer
- CI config → platform-engineer
- deployment evidence → sre
- auth/crypto/secrets → security-auditor
- always → tech-lead

A typical handover ends up with 3-5 roles in the registry entry.

/idea — 3 gaps closed + 1 numbering bug fixed
----------------------------------------------

1. **Input validation for category**. The old step 2 asked for
   category but silently accepted anything. A user who typed
   `foo` got `Category: foo` in the backlog.

   Fix: the prompt now explicitly lists 1/2/3/4, accepts either
   the number or the word (case-insensitive), and **re-prompts
   on invalid input** with `Please pick 1-4 or type the category
   name.` Loops until valid.

   Same treatment for the description field — re-prompt if empty.

2. **Dedup check before appending** (new step 3). Prevents
   submitting the same idea twice.

   Implementation: a simple token-overlap heuristic. Normalise
   both titles (lowercase, strip punctuation), compare with a
   threshold (≥ 80% of the words in the shorter title appear in
   the longer one).

   If a match is found:

     ⚠ Similar idea already in the backlog:
       IDEA-025 — {existing title} (status: {status})

     Is this a duplicate? (y = skip, n = add anyway)

   On y: skip the append, suggest `/write-spec IDEA-025` if the
   user wants to work on the existing idea. On n: continue to
   step 4.

3. **Error handling for `gh issue create`** in step 6. Old
   behaviour: if the tracking issue creation failed, the idea
   was half-saved or the whole thing errored out. New behaviour:

     ⚠ Couldn't create the tracking issue: {reason}
       The idea is still saved in projects/ideas-backlog.md
       as IDEA-NNN.

     Try again? (y = retry, n = skip, gh = show the gh error)

   With a failure-modes table: missing auth scope, label not
   found, rate limit, network error, etc. — each has a specific
   recovery path.

   Guiding principle (now rule me2resh#8): **the backlog entry is the
   primary artefact; the tracking issue is a bonus**. Never lose
   the backlog entry because GitHub was flaky.

4. **Section numbering bug** fixed. The old file had steps 1, 2,
   4, 5, 6 — step 3 was missing because an earlier refactor
   deleted it without renumbering. With the new step 3 (dedup
   check) the sequence is now 1, 2, 3, 4, 5, 6 again.

Rules sections updated for both skills
--------------------------------------

- /handover: rule 4 is now "Auto-append to the registry (with
  confirmation)"; rule 5 is "Derive roles from the stack"; rule
  6 is "Derive next steps from the risks"; rule 10 is "Never
  break the registry".
- /idea: rule 6 is "Validate before accepting"; rule 7 is
  "Dedup before appending"; rule 8 is "The backlog is the
  primary artefact".

What's NOT in this commit
-------------------------

- The tabbed terminal animation (me2resh#102) — separate ticket,
  not part of the epic closure
- Any changes to the skill frontmatter
- Any changes to the `/handover` assessment output format
  beyond the Next Steps and Post-Handover Checklist sections
- Any changes to `/idea`'s backlog file format or ID scheme

This closes the epic me2resh#100. After this merges, apexstack is at
its v0.2 shape: positioning rewritten, lifecycle demo in the
hero, 19 roles wired in, multi-project-only + fork-first install,
and now the /handover and /idea skills that actually work
end-to-end.

Closes me2resh/apexscript-org#101
Closes me2resh/apexscript-org#100

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat(#232): adopter handbooks consumed by Rex during code review

Introduces a `handbooks/` directory at the ops fork root for
adopter-authored coding standards that Rex consults during PR review,
layered on top of the framework's generic rules in `.claude/rules/`.
Discovery is by path-convention (no YAML frontmatter); enforcement
defaults to advisory with explicit blocking opt-in via the literal
`ENFORCEMENT: blocking` marker phrase at the top of a file.

Components:

- handbooks/README.md — index + discovery rules + advisory/blocking
  convention + authoring guidance + the four sample-handbook table
- handbooks/architecture/clean-architecture-layers.md — DDD layering
  (advisory, always-load)
- handbooks/architecture/migration-safety.md — schema-migration
  backwards-compat rule (BLOCKING, always-load) — demonstrates the
  enforcement opt-in shape
- handbooks/language/typescript/strict-mode.md — TS strict-mode + no
  bare `any` (advisory, loads on **/*.{ts,tsx} diff match)
- handbooks/general/commit-message-quality.md — commit-message WHY-
  not-WHAT discipline (advisory, always-load)
- .claude/agents/code-reviewer.md — new § 8 "Adopter Handbooks":
  discovery (path-convention), enforcement (advisory vs blocking
  marker), what-to-surface, Handbook Findings section in the review
  output template, new Rules entry #9 codifying the layer
- CLAUDE.md — Quick Reference + framework integration table updated
  to surface handbooks/ alongside .claude/rules/

Decision rationale: docs/agdr/AgDR-0020-adopter-handbooks-for-rex.md
captures all five load-bearing decisions (scope, discovery,
format, sample-scaffold-per-feature, enforcement default) with
options-considered tables.

Closes #232

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

* fix(#232): correct rule-file count in CLAUDE.md (10 → 11)

Rex flagged the count mismatch in his review of #233. There are
actually 11 rule files in .claude/rules/ (the previous list also
omitted leak-protection). One-line fix.

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

* fix(#232): markdownlint MD028/MD032 in handbook samples

Three handbook samples tripped markdownlint:
- migration-safety.md: list-inside-blockquote needs a blank `>` line
  before the list (MD032)
- commit-message-quality.md: two adjacent blockquotes joined with `>`
  instead of a blank line (MD028)
- strict-mode.md: same fix for three adjacent sample-finding blockquotes
  (MD028 x2)

Pure formatting fix. No content change.

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>

…d gaps) (me2resh#9)

Closes the last sub-task of the apexstack positioning epic (me2resh#100).
The Explore agent audit flagged specific gaps in both skills; these
were deferred through PRs me2resh#4-me2resh#8 because they're independent of the
positioning work. Now fixed.

/handover — 3 gaps closed
-------------------------

1. **Auto-append to the portfolio registry** (was the CRITICAL gap
   from the audit). Previously the skill printed a YAML snippet
   and told the user "add this to apexstack.projects.yaml". Users
   forgot, or miscopied, or broke the YAML indentation.

   Fix: new step 6 prompts the user `Ready to add {name} to
   apexstack.projects.yaml? (y/n)`. On y, the skill:
   - Locates the registry at the ops-repo root (creates from
     .example if missing, with a warning)
   - Appends the entry via `yq` if available, otherwise plain-text
     append with careful indentation
   - Validates the result by parsing the YAML (yq or python yaml)
   - Rolls back on failure (never leaves the registry broken)
   - Confirms the append to the user with the derived roles list

   On n, prints the snippet for manual copy-paste and continues
   without writing.

2. **Dynamic "Next Steps"** derived from the risks found. The
   old template had generic placeholders like `{first concrete
   action — usually "run /audit-deps and triage criticals"}`
   which was embarrassing to ship.

   New behaviour: a mapping table in the skill spec defines the
   derivation. Examples:
   - CVEs detected → `/audit-deps {name} — triage the {severity}
     {package} CVE before any new feature work`
   - Failing tests → `Fix the {N} failing tests in {module}
     before merging new PRs`
   - No observability → `/decide on observability ({top 2 options
     for this stack})`
   - Stale CI → `Re-enable CI — copy in golden-paths/pipelines/ci.yml`
   - Coverage unknown → `Set up test coverage reporting`
   - Backlog > 10 → `Triage the issue backlog with previous owner`
   - Missing README → `Write a minimum-viable README`

3. **Post-handover checklist** added to the assessment template,
   also dynamically tailored to the risks found. Items like "close
   {top risk} before the first feature PR", "add {name} to weekly
   /stakeholder-update rollup", "set up a coverage baseline".

Bonus gap closed: **derive the role list dynamically** from the
tech stack + CI config + security surface, instead of hard-coding
`[tech-lead, backend-engineer]`. Mapping table in the spec:
- backend deps → backend-engineer
- UI code → frontend-engineer
- CI config → platform-engineer
- deployment evidence → sre
- auth/crypto/secrets → security-auditor
- always → tech-lead

A typical handover ends up with 3-5 roles in the registry entry.

/idea — 3 gaps closed + 1 numbering bug fixed
----------------------------------------------

1. **Input validation for category**. The old step 2 asked for
   category but silently accepted anything. A user who typed
   `foo` got `Category: foo` in the backlog.

   Fix: the prompt now explicitly lists 1/2/3/4, accepts either
   the number or the word (case-insensitive), and **re-prompts
   on invalid input** with `Please pick 1-4 or type the category
   name.` Loops until valid.

   Same treatment for the description field — re-prompt if empty.

2. **Dedup check before appending** (new step 3). Prevents
   submitting the same idea twice.

   Implementation: a simple token-overlap heuristic. Normalise
   both titles (lowercase, strip punctuation), compare with a
   threshold (≥ 80% of the words in the shorter title appear in
   the longer one).

   If a match is found:

     ⚠ Similar idea already in the backlog:
       IDEA-025 — {existing title} (status: {status})

     Is this a duplicate? (y = skip, n = add anyway)

   On y: skip the append, suggest `/write-spec IDEA-025` if the
   user wants to work on the existing idea. On n: continue to
   step 4.

3. **Error handling for `gh issue create`** in step 6. Old
   behaviour: if the tracking issue creation failed, the idea
   was half-saved or the whole thing errored out. New behaviour:

     ⚠ Couldn't create the tracking issue: {reason}
       The idea is still saved in projects/ideas-backlog.md
       as IDEA-NNN.

     Try again? (y = retry, n = skip, gh = show the gh error)

   With a failure-modes table: missing auth scope, label not
   found, rate limit, network error, etc. — each has a specific
   recovery path.

   Guiding principle (now rule me2resh#8): **the backlog entry is the
   primary artefact; the tracking issue is a bonus**. Never lose
   the backlog entry because GitHub was flaky.

4. **Section numbering bug** fixed. The old file had steps 1, 2,
   4, 5, 6 — step 3 was missing because an earlier refactor
   deleted it without renumbering. With the new step 3 (dedup
   check) the sequence is now 1, 2, 3, 4, 5, 6 again.

Rules sections updated for both skills
--------------------------------------

- /handover: rule 4 is now "Auto-append to the registry (with
  confirmation)"; rule 5 is "Derive roles from the stack"; rule
  6 is "Derive next steps from the risks"; rule 10 is "Never
  break the registry".
- /idea: rule 6 is "Validate before accepting"; rule 7 is
  "Dedup before appending"; rule 8 is "The backlog is the
  primary artefact".

What's NOT in this commit
-------------------------

- The tabbed terminal animation (me2resh#102) — separate ticket,
  not part of the epic closure
- Any changes to the skill frontmatter
- Any changes to the `/handover` assessment output format
  beyond the Next Steps and Post-Handover Checklist sections
- Any changes to `/idea`'s backlog file format or ID scheme

This closes the epic me2resh#100. After this merges, apexstack is at
its v0.2 shape: positioning rewritten, lifecycle demo in the
hero, 19 roles wired in, multi-project-only + fork-first install,
and now the /handover and /idea skills that actually work
end-to-end.

Closes me2resh/apexscript-org#101
Closes me2resh/apexscript-org#100

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix(#99): align default-mode framing, roles count, file count, SENTINEL nit (#3)

Post-merge cleanup after the Rex re-review of PR #1. Four blockers
and one nit — all small, all correctness fixes, no scope creep.

N1 — Default-mode contradiction (CRITICAL)
Four files claimed single-project was the default, contradicting
CLAUDE.md, README, onboarding.yaml, docs/multi-project.md, the site,
and 7 of 8 new skills (which all say multi-project is the default).
The commit message for ba892bc literally said "multi-project mode
(default)". These four were the stragglers; they now align:

- apexstack.projects.yaml.example: header comment rewritten to say
  this file is the registry for the default multi-project mode,
  single-project is the opt-in
- workspace/README.md: section order swapped (multi-project first as
  the default, single-project second as opt-in), intro paragraph
  reworded, mode comments in the onboarding.yaml snippet flipped
- projects/README.md: opening paragraph flipped so multi-project is
  described as the default
- .claude/skills/projects/SKILL.md: mode detection table row order
  flipped (multi-project "or missing" is now default), sections
  swapped so multi-project comes first as the default, single-project
  moved below as opt-in, the "how to flip back" prompt reworded

N2 — Roles count
- README.md:166 — "20 software development roles" → "19"
  (actual count: 7 Engineering + 3 Product + 3 Design + 3 Security
  + 3 Data = 19)

N3 — Site file count
- site/index.html:956 — dropped the "79 files · " prefix from the
  tree header line to kill the drift-prone number entirely. The
  descriptive tagline "the runnable + portfolio layers added in v0.1"
  stays. Any time a file is added to the repo, the header no longer
  lies.

Stale SENTINEL comment
- golden-paths/pipelines/ci.yml:93 — the security section comment
  said "SENTINEL — Security Scanning" but the job itself is named
  "Shield: Security" (the Sentinel→Shield rename from PR #1). Fixed
  the comment to match.

What's NOT in this commit
- N4 (PR body "5 skills" + Sentinel/Scout in glossary) — GitHub PR
  bodies are immutable after merge, so this is recorded in the
  tracking issue as NOT FIXABLE.
- Retroactive AgDRs — intentionally skipped per the CEO's call
  (v0.1 doesn't need them backfilled).

Verification
  grep -n "20 software development roles" README.md     → no hits
  grep -n "79 files" site/index.html                    → no hits
  grep -n "SENTINEL" golden-paths/pipelines/ci.yml      → no hits
  grep for "single-project is the default" in the 4 N1 files → no hits

Closes me2resh/apexscript-org#99

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs(#100): positioning rewrite — "where projects get forged" (#4)

* docs(#100): positioning rewrite — "where projects get forged"

Reframes apexstack from "an opinionated stack for Claude Code" to
"a multi-project ops repo where projects get forged." Targets the
solo founder / technical CEO running 2-5 products and drowning in
context-switching.

Hero (site/index.html)
- Eyebrow flips from "plain markdown · zero dependencies" to
  "for founders forging multiple products"
- H1 changes from the brand name ("apexstack") to the value
  prop ("Where projects get forged.")
- New tagline: "The multi-project ops repo where your projects
  reference each other, learn from shared experience, and ship
  production-ready under a strict SDLC."
- New subhead leans into the forge framing: "You don't add
  apexstack to a project — projects get forged inside it. One
  ops repo. Every product. Shared memory. Strict gates.
  Production-ready MVPs." Addresses the Claude-Code-lock-in
  objection head-on: "Claude Code is the default driver, but
  the rules, hooks, and templates are plain markdown and shell.
  Swap the AI. Keep the forge."
- NEW hero CTA block with star-first priority:
  [★ Star on GitHub] · [or install it now ↓]
- Install command fixed from `~/.apexstack` (home dir, broken
  per the audit) to `.apexstack` (per-ops-repo, canonical)
- Caveat rewritten to match and point at both install paths

CSS
- Added .hero__cta, .cta, .cta--primary, .cta--ghost, .cta__sep
  to the existing terminal-brutalism style block. Sharp corners,
  hairline borders, no gradients. Primary CTA goes from ink → accent
  on hover. Ghost CTA is text-only with accent on hover.

Install section (site/index.html SECTION 05)
- Kept the six-step per-ops-repo walkthrough as the canonical
  path
- Kept the single-project opt-out at the bottom
- NEW "alternative — global install" step for users running
  several ops repos (founders with multiple orgs, consultants
  with multiple clients). Documents the tradeoff: symlinks
  use absolute paths, breaks if ~/.apexstack/ moves.
- Hero caveat links to #install-global for the alternative

README.md
- Hero rewritten end-to-end to match the site
- Added "Global install (alternative)" section after the
  single-project opt-out, mirroring the new site step
- gstacks.org attribution reworded to say "for teams running
  more than one product at a time"

CLAUDE.md
- Intro title changed from "AI-Native Development Stack" to
  "A Multi-Project Forge for Claude Code"
- First paragraph expanded with the forge framing and the
  explicit "projects get forged inside it, not added to it"
  phrasing, while keeping the operational instructions
  untouched so the rest of the file still drives Claude
  correctly.

Audience discovery (recorded in the epic issue)
- Primary persona: solo founder / technical CEO juggling
  2-5 products
- Tone: punchy and confident, short sentences, strong verbs
- Forge metaphor: literal
- Conversion: star first, clone second
- Objections addressed: Claude-Code lock-in (explicit),
  opinionated (leaned into — "strict gates" is the point)

NOT in this PR (tracked in the epic)
- Commands sub-page (PR 2)
- Hero shell animation (PR 3, depends on PR 2)
- Roles integration (PR 4 — the audit flagged all 19 roles
  are completely disconnected from workflows/skills/CLAUDE.md
  imports, will fix there)
- /idea and /handover skill polish (PR 5)
- AgDR backfill — explicitly skipped
- Anvil / metalwork SVG iconography — holding off on visual
  changes until the copy is locked

Refs me2resh/apexscript-org#100

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs(#100): fix page title, meta description, CTA separator, caveat copy

Address Rex's review on PR #4 commit a044bd2:

B1 (blocker) — Page title and meta description still said "an
opinionated stack for Claude Code" and "one drop-in stack for
Claude Code". These are the highest-leverage strings in the file
(browser tabs, search results, social cards) and they directly
contradicted the rest of the PR. Replaced with the forge framing:

  <title>ApexStack — where projects get forged</title>
  <meta description="... a multi-project ops repo where your
  projects reference each other, learn from shared experience,
  and ship production-ready under a strict SDLC. Built for
  founders running 2–5 products at once. Open source. Claude
  Code native, but plain markdown underneath.">

S2 (non-blocking) — On narrow screens where .hero__cta wraps,
the `·` separator could orphan onto its own line. Added
@media (max-width: 520px) { .cta__sep { display: none; } }
so the two CTAs stack cleanly on mobile.

S3 (non-blocking) — Hero caveat said "Five more steps in install
below" but the install section now has 6 per-ops-repo steps plus
the single-project opt-out and the new global-install alternative.
Rephrased to "Full walkthrough in install below" so the copy is
count-agnostic.

S1 (skipped) — Rex noted CLAUDE.md doesn't carry "Swap the AI.
Keep the forge." but also explicitly said the spec only required
the line in site + README, so leaving CLAUDE.md's operational
tone intact.

Refs me2resh/apexscript-org#100

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs(#100): revert H1 to "apexstack_" with flicker, promote forge to tagline

CEO feedback on the live site at a044bd2:
- Missed the flickering "_" cursor on the old H1 — the three-word
  H1 ("Where projects get forged.") also overflowed the hero
  width because .hero__title has white-space: nowrap; overflow:
  hidden; and was clipping at ~"projects" on a standard laptop
- Forge statement belongs in the subheading, not the H1

Changes:
- H1 reverts to "apexstack" (was "Where projects get forged.").
  The .hero__title::after pseudo-element already appends a "_"
  and animates it via the blink keyframe, so the flicker comes
  back for free.
- Tagline becomes "Where projects get forged." — the forge
  statement now sits in the visually-prominent subheading slot
  right under the H1.
- .hero__tagline font-size bumped from clamp(1rem, 1.6vw, 1.25rem)
  to clamp(1.5rem, 3vw, 2.5rem) so a 4-word punchy statement
  carries visual weight. Also font-weight 500 → 600, line-height
  1.5 → 1.25, and a small negative letter-spacing for the
  display-style treatment.
- Sub paragraph absorbs the previous tagline content merged
  with the existing subhead — one longer descriptive paragraph
  covering the multi-project promise, the forge framing, and
  the Claude-Code lock-in objection.

Typography hierarchy is now:
- H1 (huge, monospace, with blinking cursor) — brand mark
- Tagline (display-size, punchy) — the one-line promise
- Sub (body-size, descriptive) — the multi-paragraph pitch

Refs me2resh/apexscript-org#100

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat(#100): animated /idea shell demo in the hero (#5)

PR 3 of the apexstack positioning epic. Adds an animated shell
window showing a real /idea invocation end-to-end, placed in the
hero between the CTA block and the install command so readers see
proof of what the stack feels like before they commit to cloning.

The user picked /idea (over /handover or the full ticket flow)
because it's relatable, lightweight, and demonstrates the multi-
project story: the demo creates IDEA-025, stores it in the shared
projects/ideas-backlog.md, and files a tracking issue against
your-org/example-app — all three surfaces the copy above promises.

Demo script (~11s, typewriter timing)
  $ /idea Usage-based pricing for creators
  Category? (1) New Product (2) Feature (3) Internal Tool (4) Process
  > 2
  Description? (one line)
  > Let creators tier paid access by usage instead of seat count
  Create GitHub Issue in your-org/example-app? (y/n)
  > y
  ✓ Captured: IDEA-025 — Usage-based pricing for creators
    Backlog:  projects/ideas-backlog.md
    Tracking: your-org/example-app#31
    Next:     triage, then /write-spec

Multi-project neutrality: no real project names leak — placeholder
is `your-org/example-app`, matching the `apexstack.projects.yaml.example`
registry convention the rest of the repo already uses.

Implementation
- New `.shell-demo` block in the hero (after `.hero__cta`, before
  `.install`), with a chrome bar (three brutalist dots, title,
  replay button) and a body that the JS drives
- Pre-rendered static HTML inside the body, so the demo is readable
  even with JS disabled or reduced motion
- CSS for the shell window, line types (cmd/sys/you/ok/muted),
  prompt coloring (accent for $, ink-faint for >), and replay
  button (hidden by default, revealed by JS)
- ~90 lines of JS (IIFE) in the existing end-of-body script tag:
  - Typewriter effect with variable delays per script entry
  - Per-char type speed + ±15ms jitter so it feels human
  - Clears and replays on replay-button click
  - Respects `prefers-reduced-motion`: if set, keeps the static
    pre-rendered content, hides the replay button, and exits
    without animating (no forced 10s wait for reduced-motion users)

No new dependencies. No bundler. Inline JS at the bottom of body,
runs after the DOM is parsed.

Placement rationale
- C1 = /idea (CEO picked in this thread)
- C2 = CSS + small JS (from the epic UX discovery)
- C3 = hero (from the epic UX discovery)
- Inserted between .hero__cta and .install so the funnel reads:
  pitch → CTA → proof → install → caveat → metrics

Accessibility
- `aria-label` on the shell-demo wrapper describes what it is
- `aria-hidden` on the decorative chrome dots
- Replay button has `aria-label`
- `prefers-reduced-motion` fully respected (no animation, no
  forced delay, no replay loop)
- Static pre-rendered content is a functional fallback — the
  demo is readable even with JS disabled

What's NOT in this PR (tracked in the epic #100)
- Commands sub-page (intentionally skipped per CEO call — we
  picked the animation target without needing the reference
  page first)
- Roles integration (PR 4 in the original plan)
- /handover and /idea skill polish (PR 5)

Refs me2resh/apexscript-org#100

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat(#100): lifecycle demo + catchier /idea + Swift proof line (#6)

* feat(#100): lifecycle demo + catchier /idea content + Swift proof line

Bundles three related landing-page upgrades that the CEO asked for
in one review thread:

1. /idea demo content rewrite
   - Swapped the placeholder "Usage-based pricing for creators"
     idea for "Swift menu-bar app for photo library dedup" — a
     concrete, catchier idea that also surfaces the fact that
     apexstack has been used to ship native Swift macOS apps.
   - Category now 1 (New Product) instead of 2 (Feature), since
     it's a new app.
   - Tracking target changed from `your-org/example-app#31` to
     `your-org/ops#17` because a new-product idea doesn't belong
     in an existing app's issue tracker — it lives in the ops repo.
   - Both the pre-rendered static HTML and the JS script array
     are updated line-for-line.

2. New SECTION 00 / LIFECYCLE — full ticket lifecycle demo
   A second animated shell window placed directly below the hero,
   covering all 14 phases of the apexstack workflow end-to-end:

     /inbox → pick up #58 → Tech Lead planning → plan approval
     → Backend Engineer implementation → AgDR via /decide
     → local lint/types/tests/build → commit/push/PR → /code-review
     → Rex verdict → CEO 👍 → merge → PR→QA state transition
     → pipeline on main → QA Engineer verification on staging
     → Done

   Playback: scroll-triggered via IntersectionObserver (0.3
   threshold), plays ONCE when the section enters the viewport,
   never blocks page load. Replay button at top-right of the
   chrome. Respects prefers-reduced-motion by exiting early and
   keeping the static content in place.

   Fallback: if IntersectionObserver is unavailable, auto-plays
   after a 3s delay.

   Pre-rendered static HTML mirrors the animation line-for-line
   so the demo is readable with JS disabled.

   ASPIRATIONAL role lines flagged: "Tech Lead role active",
   "Backend Engineer role active", "QA Engineer role active" —
   roles are currently PASSIVE per the audit. PR 7 (renumbered
   roles integration) will wire the activation for real. The demo
   previews that future state.

3. Proven-stacks line + README bullet
   New small `<p class="hero__proof">` caption below the hero
   metrics:
     Proven shipping · TypeScript + AWS Lambda backends · Next.js
     web apps · Chrome extensions · native Swift macOS apps

   Uses the existing .kw accent-coloured span for the stack
   keywords so they pop without adding new CSS.

   README.md hero gains a matching bullet in the second paragraph
   calling out the four proven stack types — directly answers the
   CEO's "do we have this somewhere?" question. We did not. Now
   we do.

Also
- Titlebar nav: replaced #stack (dead anchor since PR 1 reshuffle)
  with #lifecycle pointing at the new section.
- .shell-demo--tall variant: reserves 52rem min-height on the body
  so the demo doesn't collapse while the JS rebuilds it line by
  line. Smaller font-size (12px vs 12.5px) so 50+ lines fit.
- .hero__proof: small caption style matching the existing
  ink-faint + .kw accent pattern.

The lifecycle demo's JS IIFE shares ~80 lines of helper code with
the /idea IIFE (typeInto, play loop, timeouts cleanup). Tolerated
for now as copy-paste because extracting a shared helper is a
separate refactor; both IIFEs remain self-contained and easy to
remove individually.

Multi-project neutrality verified
- grep -iE "flat-?mate|curios|sharp ?pick|movetwo|yumyum" site/index.html
  returns 0 hits. The Swift reference is via the generic phrase
  "Swift menu-bar app for photo library dedup" — no SharpPick name
  leak. Repo placeholders use your-org/{ops,billing-api} per the
  apexstack.projects.yaml.example convention.

Refs me2resh/apexscript-org#100

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat(#100): lifecycle demo scrolls like a terminal, not a full box

CEO feedback on the new animation in PR #6: "i like the new animation,
but yes too much on the page, doesn't have to be a full box displaying
everything, can be smaller like 10 lines and scrolling with new data
like terminal".

Changes
- .shell-demo--tall .shell-demo__body — fixed height 17rem (~10 lines
  at 1.55 line-height), overflow-y: auto, hidden scrollbars on both
  Firefox (scrollbar-width: none) and WebKit (::-webkit-scrollbar).
  Replaces the old min-height: 52rem that reserved the full 54-line
  height on the page.

- Added a soft top fade gradient via .shell-demo--tall::before that
  sits just under the chrome bar. Implies depth — content scrolled
  above the viewport fades out rather than hard-clipping. Uses
  --paper-2 → transparent so it flips correctly in dark mode.

- Mobile breakpoint adjusted: 18rem height + 11px font at ≤720px.

- Lifecycle IIFE play() now calls a tiny scrollToBottom() helper
  after every line append and after every typeInto callback, so the
  viewport snaps to the latest line like a real terminal. scrollTop
  resets to 0 at the start of play() so replay works from the top.

- scroll-behavior: auto (not smooth) — instant snap matches the
  terminal feel and avoids "scroll is still happening while the
  next line is typing" jank.

Behaviour
- JS-enabled: terminal-like viewport, new lines stream in at the
  bottom, old lines scroll out of view
- No-JS: same 17rem viewport with the full static content;
  overflow-y: auto means the user can scroll the box manually to
  see all 54 lines
- Reduced motion: static pre-rendered content, no animation, user
  scrolls manually (same as no-JS)

Page footprint drops from ~52rem to 17rem — the lifecycle section
now takes roughly the same vertical space as the /idea hero demo.

Refs me2resh/apexscript-org#100

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat(#100): replace /idea demo with lifecycle in the hero

CEO feedback on the live preview: "i like this more than the idea,
lets replace it". Dropping the /idea hero demo entirely and moving
the lifecycle demo into its slot.

Changes
- Hero: /idea shell-demo block deleted, replaced in the same slot
  with the lifecycle shell-demo (now carrying `rise rise-4` classes
  to match the rest of the hero fade-in sequence).
- Section 00 wrapper (`#lifecycle` section with its header + lede)
  deleted entirely. The demo is now the hero's visual focal point
  instead of a separate section below the fold.
- `/idea` JS IIFE deleted (~100 lines).
- Lifecycle JS IIFE comment updated from "(section 00)" to "(hero)"
  explaining the new placement. Code itself untouched — still uses
  IntersectionObserver with the `hasPlayed` gate + 500ms delay, so
  it auto-plays once on load (since the hero is visible immediately)
  and never re-plays on scroll.
- Titlebar nav: removed the `<a href="#lifecycle">lifecycle</a>`
  link. The demo is in the hero now, so a jump-to link is redundant.
  Nav drops from 6 items to 5 (tree · layers · examples · install
  · github →).

Behaviour
- JS on: lifecycle demo auto-plays once on load in the hero, new
  lines stream into the 17rem viewport, old lines scroll up and
  out like a terminal.
- No-JS: 17rem scrollable viewport with full static content,
  replay button hidden.
- Reduced motion: static content, no animation, replay button
  hidden.

Page structure now:
  hero → tree → layers → examples → install → scope → footer
  (SECTION 00 is gone; tree becomes the first post-hero section)

The footprint below the hero is ~52rem lighter than before the
revert — the single in-hero demo is the only shell demo on the
page.

Refs me2resh/apexscript-org#100

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat(#100): lifecycle animation — autonomous by default, 3 human touchpoints

CEO feedback: "for the animation, it looks like we run every command
in the cycle, it should be autonomous unless plan & push & waiting
for approval". The old script prefixed every step with '$ ' prompts
as if a human was typing each command. Reality: apexstack runs
almost everything autonomously once a ticket is picked up. The human
only intervenes at THREE well-defined touchpoints.

Three human touchpoints now explicit
  1. Plan approval   → "Approve plan? (y/n)"     → > y
  2. Push approval   → "Ready to push PR #84?"   → > y
  3. CEO approval    → "Waiting for CEO 👍"       → CEO 👍

Everything between those is narrated autonomously — no $ prompts,
no implied human typing. The demo now reads as a log stream of what
Atlas did between the three decision points.

Script rewrite (both static HTML and JS array)
- Removed every "$ /command" and "$ git ..." line. They become
  sys-type narrations: "Inbox:", "Picking up #58...", "Plan for
  #58:", "Detected decision: ...", "Local checks:", "Ready to
  push PR #84?", "Pushed · Opened PR #84", "Rex reviewing a3f9c21...",
  "Pipeline on main:", etc.
- Added a new "Ready to push PR #84? → > y" touchpoint between
  the local-checks block and the PR creation. Makes the "push"
  human intervention explicit.
- Role lines softened: "Tech Lead role active — planning" →
  "Tech Lead role: planning" (less shouty; still aspirational —
  roles become real in PR 7).
- Dropped the square-bracket theatrics on the CEO wait line:
  "[waiting for CEO 👍]" → "Waiting for CEO 👍" (cleaner, and
  the long 1500ms delay already conveys "we're waiting").

JS next() simplification
- Removed the `cmd` branch special-case in next(). The `$ ` prompt
  emission is gone. Only `you` type still prefixes with a `prompt`
  span (the `> ` marker). Every other type is rendered as plain
  textContent, with colour driven by the CSS class.
- The `.cmd` CSS class is still in the stylesheet for possible
  future use but nothing in the current script emits it.

Runtime impact
- ~55 script entries (was 52) — +3 for the new push-approval
  touchpoint (sys + you + blank)
- Total runtime drops from ~40-50s to ~18-22s because most lines
  are now instant (speed=0) instead of typewriter. The only
  typewriter action left is the two "y" responses. Much snappier.

Behaviour
- JS on: narrated log stream, only "> y" lines are typed, demo
  auto-plays once on load, scrolls like a terminal
- No JS / reduced motion: same static content, manual scroll
- Multi-project neutrality: 0 hits for flat-mate/curios/sharppick

Refs me2resh/apexscript-org#100

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat(#100): wire 19 role files into workflows, skills, and CLAUDE.md (#7)

Closes the biggest gap the PR #1 Rex audit flagged: all 19 role
files exist and are well-specified, but nothing in the running
system references them. CLAUDE.md listed them in a table but
didn't @import anything. No workflow named a role file. No skill
invoked a specific role. Users had to manually say "please read
roles/engineering/qa-engineer.md and act as the QA Engineer" for
anything to happen. Roles were passive reference docs.

This PR makes them first-class participants.

New file — .claude/rules/role-triggers.md (~107 lines)
- Activation table for all 19 roles (department, file path, when
  to activate with 2-4 concrete conditions per role)
- Activation protocol (read file → adopt identity → follow
  handoff rules → stay in role until task completes)
- Auto-activation signal table (ticket label, PR path match,
  incident, PRD draft, etc. → which role activates)
- Prompted activation examples ("act as the QA Engineer for #42")
- Role boundary enforcement (CAN/CANNOT from each role file is
  strict — hand off when you hit a CANNOT)
- Handoff artefact table (PRD, tech design, testable build,
  security findings, AC sign-off, etc.) — the contracts between
  roles
- Explicit "aspirational → real" closing note explaining what
  this file makes concrete

CLAUDE.md
- ROLES section gains an "Activation" subsection explaining the
  model in three short paragraphs
- New @.claude/rules/role-triggers.md import at the end of the
  subsection — the trigger table is loaded into every session so
  Claude always knows which role to activate on which signal

workflows/sdlc.md — each phase gets a "Primary role" header
- Phase 1 Planning → Tech Lead (+ Product Manager, Head of Eng
  on escalation)
- Phase 2 Tech Design → Tech Lead (+ Head of Eng, UX/UI Designer)
- Phase 3 Build → Backend / Frontend Engineer (+ Tech Lead
  coordinator)
- Phase 4 Code Review → Tech Lead + Rex (+ Security Auditor on
  auth/crypto diff, UI Designer on UI diff)
- Phase 5 QA → QA Engineer (MANDATORY — merged code is never Done
  without QA sign-off; expands the existing "QA gate" warning)
- Phase 6 Deploy → Platform Engineer (+ SRE for runbook/rollback)
- Phase 7 Monitor → SRE (+ Head of Eng escalation)
- "Roles Summary" table at the bottom rewritten with every role
  as a markdown link to its file

workflows/code-review.md — Roles table expanded
- Author column now names the actual Backend/Frontend Engineer
  roles
- Automated reviewer named explicitly (Rex agent)
- Human approval gate = Tech Lead with a link to the role file
- Conditional Security Auditor and UI Designer rows added with
  their trigger conditions
- QA Engineer called out as "not a reviewer" to kill the common
  misconception that QA approves merges

workflows/deployment.md — new Roles table at the top
- Maps each deployment stage (CI/CD maintenance, staging, prod
  gate, incident response, post-deploy monitoring, security gate)
  to its activating role and trigger condition
- Head of Eng sign-off noted for risky production promotions

Skills — 5 files get an "Activated role" section between the
intro and the process
- /decide → Tech Lead (+ Head of Eng for arch-review threshold,
  Security Auditor if decision touches auth/secrets)
- /write-spec → Product Manager (+ Head of Product escalation,
  UX/UI Designer for design-heavy features; hands off to Tech
  Lead at Tech Design phase)
- /code-review → Rex + Tech Lead (+ conditional Security Auditor,
  conditional UI Designer) — the richest of the five because
  code review has the most conditional branches
- /security-review → Shield + Security Auditor (+ Head of Security
  for strategic calls, Penetration Tester for active testing)
- /roadmap → Head of Product (+ Product Analyst for data-driven
  reprioritisation)

Not in scope for this PR
- The 19 role files themselves are untouched — their content was
  already good per the audit
- Skills /inbox, /status, /tasks, /projects, /idea, /handover,
  /audit-deps, /stakeholder-update don't get role activation
  sections because they're portfolio / CEO-facing or the role
  is implicit from the task context
- No new roles added
- The .claude/rules/role-triggers.md file is NOT imported from
  every workflow/skill — only CLAUDE.md imports it once, and
  individual files reference it relatively when they need to
  cite the activation protocol. Keeps the context footprint
  lean.

Token budget
- role-triggers.md is ~107 lines (~1.3k tokens). Loaded once per
  session via the CLAUDE.md @import. The 19 full role files are
  NOT pre-loaded — they're read on demand when their trigger
  fires (saves ~22k tokens that would otherwise sit idle).

Multi-project neutrality
- grep for flat-mate/curios/sharppick/movetwo/yumyum across all
  10 changed files → 0 hits
- All file paths use `roles/{department}/{role}.md` and
  `.claude/rules/role-triggers.md` — no org-specific references

Refs me2resh/apexscript-org#100

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* refactor(#100): multi-project only + fork-first install (#8)

* refactor(#100): multi-project only + fork-first install

Two strategic direction changes from the CEO rolled into one PR
because they're deeply coupled — removing single-project mode
changes the install flow, and the fork-first install changes
every doc that talked about cloning into .apexstack/.

1. Drop single-project mode entirely
---------------------------------------
Single-project mode was a dual-mode compromise that added
complexity without matching how apexstack is actually used.
The audit found every skill had a mode branch that was never
exercised in practice. Users who thought they only had one repo
almost always ended up with two within a few months.

- onboarding.yaml: removed the apexstack.mode field + the
  multi-paragraph mode explainer
- apexstack.projects.yaml.example: rewritten header comment to
  lead with "no single-project fallback mode; register your one
  repo and the skills work the same"
- CLAUDE.md: OPERATING MODE section replaced with PORTFOLIO MODEL
  section explaining the fork-as-ops-repo mental model. SETUP
  section dropped the "read apexstack.mode" step.
- docs/multi-project.md: full rewrite as the canonical setup
  guide. Removed the TL;DR comparison table, the "when to switch
  to single-project" section, the "migrating from single → multi"
  section, the "going back: multi → single" section, and every
  trade-off framed against single-project.
- 8 skill files (handover, idea, inbox, projects, roadmap,
  stakeholder-update, status, tasks): dropped every mode detection
  block (grep onboarding.yaml for mode:), dropped every
  mode-comparison table, simplified scope statements to "iterates
  the registry" or equivalent. Dead code removed.
- workspace/README.md, projects/README.md: removed the "two
  modes" sections, simplified to describe the single portfolio
  pattern.
- README.md: removed the Operating modes table + the
  Single-project install (opt-out) section + the Global install
  (alternative) section (see item 2 for why the global install
  is gone).
- site/index.html: removed the "Operating modes (single &
  multi-project)" hero metric, updated the layers section lede,
  removed every single-project mention in Layer 05, removed the
  "opt-out — single-project mode" install step, removed the
  "alternative — global install" install step.

The only remaining "single-project" strings in the codebase are
TWO intentional negation disclaimers (in docs/multi-project.md
and site/index.html Layer 05) that tell anyone searching for the
concept: "there is no single-project fallback". That's a feature,
not a bug.

2. Fork-first install (replaces clone-into-.apexstack/)
--------------------------------------------------------
The old install model cloned apexstack into a hidden .apexstack/
subdirectory of a separate ops repo, then symlinked .claude/ up
one level. Three problems:

  a. Brand invisibility — .apexstack/ is a dotfile, hidden from
     ls and GitHub views. Nobody knew you were using apexstack.
  b. Two repos to maintain — your ops repo plus the nested clone.
  c. Symlink fragility — the .claude/ symlink broke on dotfile
     sync tools and Windows setups.

New model: your ops repo IS a fork of apexstack. One repo, no
nesting, no symlinks. Upgrades flow via `git fetch upstream &&
git merge upstream/main` — the standard fork workflow.

Per the CEO's call, users can rename the fork to your-org/ops
or keep it as your-org/apexstack. GitHub handles the rename
cleanly. The install docs describe the rename as optional.

- Hero install block: "git clone ... .apexstack" →
  "gh repo fork me2resh/apexstack --clone"
- Hero CTA: "★ Star on GitHub" →
  "★ Star · ⑂ Fork on GitHub" (single button, per Q3.c)
- Hero install caveat: rewrote to explain "the fork IS your ops
  repo, no nested installs, no symlinks"
- SECTION 05 / INSTALL: full rewrite. 6 steps became 5. The 5
  steps are: star+fork, add upstream remote, configure
  onboarding.yaml, create the registry, start working. The old
  "symlink runnable layer" step is gone (no symlink needed). The
  old "wire CLAUDE.md with @-import" step is gone (CLAUDE.md is
  already at the fork root).
- README.md Quick Start: same 5-step rewrite, matched to the
  site flow.
- docs/multi-project.md: complete rewrite as a setup guide
  covering the fork flow, directory layout, daily workflow,
  upgrade path (git pull upstream), trade-offs, and FAQ. ~314
  lines, honest about the pros and cons.

Side effects (positive)
-----------------------
- Layers section UX bug from the earlier screenshot is gone.
  Layer 05 no longer has "content only fills left 1/3" — added a
  new .layer--wide CSS variant with a 2-column internal grid
  (description on the left, registry file list on the right).
  Fills the full width naturally.
- Hero metric #4 changed from "2 Operating modes" to "1 Fork ·
  your ops repo in one command" — celebrates the simplification
  instead of the dropped complexity.

Not in scope
------------
- The lifecycle demo script in the hero is untouched
  (it doesn't reference install or mode)
- Role files in roles/ are untouched (no mode refs)
- Workflow files (sdlc, code-review, deployment) are untouched
  (no mode refs)
- .claude/agents/ are untouched (no mode refs)
- Hooks in .claude/hooks/ are untouched (no mode refs)
- The apexstack.mode field in the existing onboarding.yaml.example
  — this is the site's own onboarding.yaml which gets removed
  with the mode config

Verification
------------
  grep -rn "single-project" --include="*.md" --include="*.yaml" \
    --include="*.html" --include="*.sh"
  → 2 hits, both intentional negation disclaimers

  grep -rn "mode:" --include="*.md" --include="*.yaml" \
    --include="*.html"
  → 0 hits

  grep -rn "\.apexstack/" --include="*.md" --include="*.yaml" \
    --include="*.html"
  → only in docs/multi-project.md explaining why the old pattern
    was dropped (historical context paragraph)

Multi-project neutrality
------------------------
  grep -iE "flat-?mate|curios|sharp ?pick|movetwo|yumyum" \
    on all 16 changed files
  → 0 hits

Refs me2resh/apexscript-org#100

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* refactor(#100): address Rex review — blockers + CTA contradiction + doc fork

Rex's delta review on 160148f found 4 blockers + 1 site CTA
contradiction + a few vestigial mentions. Plus the deeper grep
I ran caught two more: a stale .apexstack/ example in
golden-paths/pipelines/README.md, and docs/getting-started.md
still documenting the old .apexstack/ + symlink flow.

Blockers fixed
--------------
- B1: idea/SKILL.md rule 5 "Mode-aware — detect apexstack.mode"
      → "Single backlog — every idea goes into
         projects/ideas-backlog.md"
- B2: roadmap/SKILL.md rule 6 "Mode-aware — ROADMAP.md or
      projects/<name>/roadmap.md"
      → "One roadmap per project — always write to
         projects/<name>/roadmap.md"
- B3: status/SKILL.md had a "## Multi-project mode" section and
      a "Single-project mode default:" output format header.
      Renamed to "## Portfolio output" and "## Output format
      (one-project view with --project <name>)". No behaviour
      change — just language that doesn't frame /status as
      mode-switching.

Non-blocking suggestions fixed
------------------------------
- S1 (site CTA contradiction): index.html:1732 final CTA was
      "one git clone, one symlink, and one config file away"
      → "one fork, one clone, and one config file away". The
      old line directly contradicted the "no symlinks" story.
- S3 (CLAUDE.md QUICK REFERENCE):
  - "Project registry (multi-project)" → "Portfolio registry"
  - "Per-project docs (multi-project)" → "Per-project docs"
  - "Live working copies (multi-project)" → "Live working
    copies (gitignored)"
  - "Multi-project guide" → "Full setup guide"

Also cleaned up while I was in there
------------------------------------
- inbox/SKILL.md rule 3: "Multi-project mode iterates the
  registry" → "Registry-scoped — only projects listed in
  apexstack.projects.yaml count"
- projects/SKILL.md error row: "Multi-project mode but no
  apexstack.projects.yaml" → "No apexstack.projects.yaml at
  the ops-repo root"
- projects/README.md directory layout: "ideas-backlog.md
  (multi-project mode)" → just "shared ideas backlog"

Out-of-scope but too tangled to leave
-------------------------------------
- golden-paths/pipelines/README.md: `cp` examples pointed at
  `.apexstack/golden-paths/...` as if apexstack were still a
  nested clone. Rewrote to "copy from ~/apexstack (your fork)
  to your managed project's .github/workflows/". Not strictly
  within the mode-removal scope but the README is now
  self-consistent with the fork-first install.
- docs/getting-started.md: rewrote Step 1 and Step 3 to match
  the new fork flow. Was still documenting the old .apexstack/
  + submodule + @.apexstack/CLAUDE.md @-import pattern —
  guaranteed to confuse a new adopter. Keeps parity with
  docs/multi-project.md and the site install section.

Verification
------------
  grep -rin "single-project" --include="*.md" --include="*.yaml" \
    --include="*.html"
  → 2 hits, both intentional negation disclaimers in
    docs/multi-project.md:5 and site/index.html:1367

  grep -rn "apexstack.mode\|mode: multi-project\|mode: single-project\|Mode-aware" \
    --include="*.md" --include="*.yaml" --include="*.html"
  → 0 hits

  grep -rn "\.apexstack/" --include="*.md" --include="*.html"
  → only in docs/multi-project.md historical context paragraph
    (explaining why the old pattern was dropped) and README.md
    negation disclaimer ("no .apexstack/ symlinks")

Refs me2resh/apexscript-org#100

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat(#101): /handover auto-append + /idea polish (audit-flagged gaps) (#9)

Closes the last sub-task of the apexstack positioning epic (#100).
The Explore agent audit flagged specific gaps in both skills; these
were deferred through PRs #4-#8 because they're independent of the
positioning work. Now fixed.

/handover — 3 gaps closed
-------------------------

1. **Auto-append to the portfolio registry** (was the CRITICAL gap
   from the audit). Previously the skill printed a YAML snippet
   and told the user "add this to apexstack.projects.yaml". Users
   forgot, or miscopied, or broke the YAML indentation.

   Fix: new step 6 prompts the user `Ready to add {name} to
   apexstack.projects.yaml? (y/n)`. On y, the skill:
   - Locates the registry at the ops-repo root (creates from
     .example if missing, with a warning)
   - Appends the entry via `yq` if available, otherwise plain-text
     append with careful indentation
   - Validates the result by parsing the YAML (yq or python yaml)
   - Rolls back on failure (never leaves the registry broken)
   - Confirms the append to the user with the derived roles list

   On n, prints the snippet for manual copy-paste and continues
   without writing.

2. **Dynamic "Next Steps"** derived from the risks found. The
   old template had generic placeholders like `{first concrete
   action — usually "run /audit-deps and triage criticals"}`
   which was embarrassing to ship.

   New behaviour: a mapping table in the skill spec defines the
   derivation. Examples:
   - CVEs detected → `/audit-deps {name} — triage the {severity}
     {package} CVE before any new feature work`
   - Failing tests → `Fix the {N} failing tests in {module}
     before merging new PRs`
   - No observability → `/decide on observability ({top 2 options
     for this stack})`
   - Stale CI → `Re-enable CI — copy in golden-paths/pipelines/ci.yml`
   - Coverage unknown → `Set up test coverage reporting`
   - Backlog > 10 → `Triage the issue backlog with previous owner`
   - Missing README → `Write a minimum-viable README`

3. **Post-handover checklist** added to the assessment template,
   also dynamically tailored to the risks found. Items like "close
   {top risk} before the first feature PR", "add {name} to weekly
   /stakeholder-update rollup", "set up a coverage baseline".

Bonus gap closed: **derive the role list dynamically** from the
tech stack + CI config + security surface, instead of hard-coding
`[tech-lead, backend-engineer]`. Mapping table in the spec:
- backend deps → backend-engineer
- UI code → frontend-engineer
- CI config → platform-engineer
- deployment evidence → sre
- auth/crypto/secrets → security-auditor
- always → tech-lead

A typical handover ends up with 3-5 roles in the registry entry.

/idea — 3 gaps closed + 1 numbering bug fixed
----------------------------------------------

1. **Input validation for category**. The old step 2 asked for
   category but silently accepted anything. A user who typed
   `foo` got `Category: foo` in the backlog.

   Fix: the prompt now explicitly lists 1/2/3/4, accepts either
   the number or the word (case-insensitive), and **re-prompts
   on invalid input** with `Please pick 1-4 or type the category
   name.` Loops until valid.

   Same treatment for the description field — re-prompt if empty.

2. **Dedup check before appending** (new step 3). Prevents
   submitting the same idea twice.

   Implementation: a simple token-overlap heuristic. Normalise
   both titles (lowercase, strip punctuation), compare with a
   threshold (≥ 80% of the words in the shorter title appear in
   the longer one).

   If a match is found:

     ⚠ Similar idea already in the backlog:
       IDEA-025 — {existing title} (status: {status})

     Is this a duplicate? (y = skip, n = add anyway)

   On y: skip the append, suggest `/write-spec IDEA-025` if the
   user wants to work on the existing idea. On n: continue to
   step 4.

3. **Error handling for `gh issue create`** in step 6. Old
   behaviour: if the tracking issue creation failed, the idea
   was half-saved or the whole thing errored out. New behaviour:

     ⚠ Couldn't create the tracking issue: {reason}
       The idea is still saved in projects/ideas-backlog.md
       as IDEA-NNN.

     Try again? (y = retry, n = skip, gh = show the gh error)

   With a failure-modes table: missing auth scope, label not
   found, rate limit, network error, etc. — each has a specific
   recovery path.

   Guiding principle (now rule #8): **the backlog entry is the
   primary artefact; the tracking issue is a bonus**. Never lose
   the backlog entry because GitHub was flaky.

4. **Section numbering bug** fixed. The old file had steps 1, 2,
   4, 5, 6 — step 3 was missing because an earlier refactor
   deleted it without renumbering. With the new step 3 (dedup
   check) the sequence is now 1, 2, 3, 4, 5, 6 again.

Rules sections updated for both skills
--------------------------------------

- /handover: rule 4 is now "Auto-append to the registry (with
  confirmation)"; rule 5 is "Derive roles from the stack"; rule
  6 is "Derive next steps from the risks"; rule 10 is "Never
  break the registry".
- /idea: rule 6 is "Validate before accepting"; rule 7 is
  "Dedup before appending"; rule 8 is "The backlog is the
  primary artefact".

What's NOT in this commit
-------------------------

- The tabbed terminal animation (#102) — separate ticket,
  not part of the epic closure
- Any changes to the skill frontmatter
- Any changes to the `/handover` assessment output format
  beyond the Next Steps and Post-Handover Checklist sections
- Any changes to `/idea`'s backlog file format or ID scheme

This closes the epic #100. After this merges, apexstack is at
its v0.2 shape: positioning rewritten, lifecycle demo in the
hero, 19 roles wired in, multi-project-only + fork-first install,
and now the /handover and /idea skills that actually work
end-to-end.

Closes me2resh/apexscript-org#101
Closes me2resh/apexscript-org#100

Co-authored-by: me2resh <ahmed.abdelaliem@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat(#103): mechanical enforcement of ticket-first, auto-review, onboarding

Test run of ApexStack on a real project exposed three SDLC compliance gaps
that all traced back to the same root cause: workflow rules lived as prose
in CLAUDE.md, .claude/rules/, and workflows/ — advice the model drops under
pressure. This PR moves the four highest-leverage rules into .claude/hooks/
where the harness executes them mechanically.

## New hooks

- require-active-ticket.sh — PreToolUse on Edit/Write/MultiEdit; blocks
  code-path edits when the active-ticket marker is missing. Exempts
  .claude/, docs/, projects/*/docs/, and any *.md file.
- auto-code-review.sh — PostToolUse on Bash; fires after PR creation and
  nudges Claude to invoke the code-reviewer agent (Rex) immediately.
- block-unreviewed-merge.sh — PreToolUse on Bash; blocks PR merge unless
  a Rex approval file exists at .claude/session/reviews/<pr>-rex.approved
  AND its SHA matches HEAD.
- onboarding-check.sh — SessionStart; warns on every new session until
  /onboard writes the onboarded marker.

## New skills

- /start-ticket <issue> — verifies the issue via `gh issue view`, writes
  the active-ticket marker, suggests a branch name per git-conventions.md.
- /onboard — day-one discovery pass (project identity, tracker repo,
  required CI checks, reviewers, UI work, deploy targets, sensitive topics).
  Writes the onboarding marker and .claude/project-config.json. Idempotent.

## Infrastructure

- .claude/settings.json registers SessionStart, new PreToolUse matchers for
  Edit/Write/MultiEdit and pre-merge, and a PostToolUse Bash matcher for
  post-PR-creation. Existing six-hook Bash block untouched. $schema preserved.
- .gitignore (new file — apexstack had none) excludes .claude/session/,
  .claude/project-config.json, .claude/settings.local.json, and workspace/*/
  per the multi-project model in docs/multi-project.md. workspace/README.md
  stays visible via a negation rule.
- .claude/hooks/README.md documents the four new flows, the exit-code
  semantics (exit 2 blocks in PreToolUse, nudges in PostToolUse), the
  session-state directory layout, the merge-gate trust model, and how to
  add new hooks.

## Review history

Rex approved across two passes. Three nice-to-haves from the first pass
were applied in a follow-up commit:

- onboarding-check.sh switched to an interpolating heredoc so the "no
  onboarding marker at" message prints the real absolute path and a
  copy-pasteable workaround.
- auto-code-review.sh comment expanded to cite harness version drift
  (Claude Code 2.x+ vs 1.x vs earliest builds) and flag when the older
  fallback paths can be dropped.
- Hooks README documents the merge-gate's local trust model: the approval
  file is session state, not a remote trust boundary, so branch protection
  and CODEOWNERS remain the adversarial layer.

Refs me2resh/apexscript-org#103

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat(#11): require explicit per-merge CEO approval — never infer from "go"

Closes the failure mode where I merged apexstack PR #10 after a plan-level
"go" that was not a per-PR merge approval. The rule existed in prose but
was ambiguous, and there was no mechanical enforcement for the human side
of the 2-reviews gate.

## Prose tightening

- `.claude/rules/pr-workflow.md` "Before PR merge" section rewritten with
  a concrete wrong/right example showing that plan-level "go" does NOT
  carry through to the merge step. Lists other destructive actions
  (force pushes, branch deletes, closing issues with dependents, external
  posts) that need the same per-action explicit approval.
- `CLAUDE.md` Quality Rules adds an explicit bullet calling out per-PR
  CEO approval, linking to the full rule.

## Mechanical enforcement

- `.claude/hooks/block-unreviewed-merge.sh` now requires TWO approval
  markers, not one:
    - `.claude/session/reviews/<pr>-rex.approved` — from the code-reviewer
      agent (existing)
    - `.claude/session/reviews/<pr>-ceo.approved` — from the /approve-merge
      skill, on explicit per-PR user invocation (new)
  Both must contain the current HEAD SHA. Any mismatch blocks the merge.
- New /approve-merge <pr> skill at .claude/skills/approve-merge/SKILL.md.
  Its one job is to write the CEO marker on explicit per-PR user
  invocation. Definition includes valid/invalid invocation triggers and
  an anti-pattern section describing the exact failure mode the skill
  exists to prevent.
- Marker paths anchored at `git rev-parse --show-toplevel` so invocation
  from any subdirectory (e.g. workspace/<project>/) still writes to the
  path the hook reads from.
- .claude/hooks/README.md documents both markers, why there are two, and
  the trust model: local session state, not a remote trust boundary. The
  failure mode closed is "invisible inference" (Claude decides "go"
  meant "merge"); the hook converts it into "visible rule violation"
  (Claude touched a file it wasn't supposed to). Grep-able instead of
  invisible.

## Dogfooding

This PR is the first one merged under the new rule. Flow:

  1. PR created, auto-code-review hook fires (already in place from #10)
  2. Rex reviews — APPROVED, flags 4 nice-to-haves, 1 of which is a real
     correctness bug (cwd-relative path in skill step 5)
  3. Fix commit pushed addressing the cwd bug; other 3 nice-to-haves
     deferred to a follow-up ticket
  4. Rex re-reviews the new HEAD — APPROVED, no further issues
  5. CEO is asked per-PR, explicitly, for merge approval on PR #12
  6. CEO replies "12 approved" — explicit per-PR approval
  7. /approve-merge 12 invoked — writes CEO marker at repo root
  8. Confirmation returned to CEO, separate per-action nod requested
  9. CEO replies "ship it"
  10. This squash-merge runs — both markers present, both SHAs match HEAD,
      merge-gate hook allows through

Rex reviewed twice. CEO explicitly approved twice (once for the merge,
once for executing the merge) — each moment discrete, each named the PR.

Closes #11

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat(#14): ticket-vocabulary rule + verify-issue-exists backstops

Closes the vocabulary-collision failure mode where Claude's internal
plan decomposition wears tracker clothing and the user reasonably
reads it as tracker state. Root cause identified on 2026-04-11 after
a friend of the CEO hit it on a real project: the agent presented a
10-item "Ticket N ... blocked by #M" plan in chat, none of which
existed on GitHub.

## Root-cause rule (primary fix)

- New .claude/rules/ticket-vocabulary.md reserves `Ticket`, `#N`, and
  dependency notation (`blocked by #N`, `depends on #N`, etc.) for real
  GitHub issues only. Prescribes safe vocabulary for in-conversation
  planning: `Step N`, `Item A`, plain bullets, phases. Includes a
  boundary-crossing rule — if plan items need tracker properties, STOP
  and run `gh issue create` before presenting. Contains the 2026-04-11
  anti-pattern verbatim and a corrected version side-by-side.
- CLAUDE.md Quality Rules links the new rule alongside the per-PR
  merge approval bullet (both are "don't fake process state" rules).

## Mechanical backstops

- validate-pr-create.sh extended: extracts the issue number from the
  PR title and runs `gh issue view` against the resolved tracker repo.
  Blocks PR creation if the referenced issue doesn't exist. Tracker
  repo resolves from .claude/project-config.json first, falls back to
  origin remote.
- New verify-commit-refs.sh on PreToolUse / Bash git-commit: parses
  the commit message from -m or -F args, scans for Closes/Fixes/
  Resolves/Refs/Related-to #N patterns, verifies each via gh issue
  view. Blocks on any missing reference. Interactive commits (no -m /
  -F) are skipped.
- settings.json registers the new hook under the existing git-commit
  matcher.
- hooks/README.md adds a "Ticket-Vocabulary Backstops" section framing
  both hooks as backstops, not primary fixes.

## Multi-line -m fix (second commit)

Rex's first review labeled a multi-line gap as "nice-to-have, rarely
fires". Testing revealed it fires ALWAYS on Claude-generated commits,
because Claude uses HEREDOC substitution for nearly every commit
message, which produces literal newlines in the -m value. The hook's
`sed -nE` processes line-by-line and couldn't match across lines —
result: the hook was silently inert on its own introducing commit.
The "dogfood success" in the PR #16 body was a false positive.

Fix: `COMMAND_FLAT=$(echo "$COMMAND" | tr '\n' ' ')` before all sed
extraction. Multi-line -m values now parse as a single logical line.
Scoped only to parsing — extracted MSG preserves original content.
Single-line paths unregressed (zero newlines → tr is a no-op).

## Smoke tests (post both fixes)

  1. settings.json jq parse                                        ✓
  2. validate-pr-create allows real #11 / #14 in title             ✓
  3. validate-pr-create blocks fake #999999 in title (exit 2)      ✓
  4. verify-commit-refs allows single-line real ref                ✓
  5. verify-commit-refs blocks single-line fake ref                ✓
  6. verify-commit-refs skips interactive commits                  ✓
  7. verify-commit-refs allows no-ref messages silently            ✓
  8. verify-commit-refs blocks multi-line HEREDOC with fake ref    ✓
  9. verify-commit-refs allows multi-line HEREDOC with real ref    ✓

## Why the rule comes first, hooks are backstops

Hooks can only see tool calls, not prose output. A fabricated `#N` in
a chat message is invisible to any hook. The vocabulary rule is the
primary defense; the hooks catch the downstream symptom at the moment
fabrication becomes durable (PR title, commit message). Linting
Claude's prose output was considered and rejected — it depends on
Claude remembering to check itself, which is exactly the failure mode
this ticket exists to prevent.

## Review history

Rex reviewed twice. First pass: approved lean with 3 nice-to-haves,
one of which (multi-line gap) turned out to be must-fix on closer
inspection. Fix commit addressed it. Rex re-reviewed the fix commit
and approved cleanly.

CEO approved per-PR explicitly (naming PR #16) before this merge,
then gave a separate explicit merge instruction in the same message.

Closes #14

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* feat(#13): rule-audit hooks — AgDR-for-arch, design review, red-CI, commit-format

Closes four previously-unenforced MUST rules from the apexstack rule
files. Ships them as hooks with narrow default path lists and
project-config overrides. Design decisions captured in AgDR-0001, the
first AgDR in the apexstack repo.

This is the third rule-enforcement PR in a single session, after #12
(explicit merge approval) and #16 (ticket vocabulary + verify-issue-
exists). Together, #11 / #14 / #13 convert 10+ prose MUST rules into
mechanical hooks.

## New hooks

- require-agdr-for-arch-changes.sh — PreToolUse on git commit. When
  the staged diff touches architecture files (infrastructure/, *.tf,
  Dockerfile*, docker-compose*, .github/workflows/), requires an AgDR
  reference in the commit message or a new AgDR file staged alongside.
  Closes the prose-only claim in agdr-decisions.md.

- require-design-review-for-ui.sh — PreToolUse on the merge step. When
  the PR diff touches UI files (.tsx, .jsx, .vue, .svelte, .css, .scss,
  .sass, .less, design-tokens), requires a design approval marker at
  .claude/session/reviews/<pr>-design.approved. Regex is .tsx$ / .jsx$
  EXACTLY — NOT .tsx? — so backend .ts files don't false-positive.
  Closes the prose-only gate in pr-quality.md and code-review.md.

- block-merge-on-red-ci.sh — PreToolUse on the merge step. Runs the
  pr-checks lookup and blocks on any failing/cancelled/timed-out/
  pending/in-progress check. Allows the "no checks reported" case with
  a NOTE (legitimate state for repos without CI). Closes the prose-only
  "No Red CI Before Merge" rule in pr-quality.md.

- validate-commit-format.sh — PreToolUse on git commit. Validates the
  subject against ^(feat|fix|refactor|test|docs|chore|style|perf|build
  |ci|revert)(\(scope\))?: text. Matches the PR-title type list in
  git-conventions.md. Multi-line -m HEREDOC safe.

## AgDR-0001 (first AgDR in apexstack)

docs/agdr/AgDR-0001-rule-mechanization-hooks.md records:
- Options considered (broad defaults rejected, narrow + config chosen,
  one-PR-per-hook rejected)
- Rules that explicitly stay advisory (>80% coverage, no bare any,
  one-ticket-at-a-time, role trigger patterns, /decide triggers)
- Threshold decisions for each of the four hooks
- Consequences and follow-up tickets

Sets the pattern for all future apexstack AgDRs.

## Smoke tests (all passing; two bugs found mid-test and fixed)

  1. settings.json jq parse ................................ valid
  2. validate-commit-format:
     - feat: ok / feat(scope): ok / multi-line HEREDOC ok .. exit=0
     - "added thing" (no type) ............................. exit=2
  3. require-agdr-for-arch-changes:
     - no staged / non-arch file ........................... exit=0
     - Dockerfile + no AgDR ................................ exit=2
     - Dockerfile + "per AgDR-0042" in message ............. exit=0
  4. block-merge-on-red-ci:
     - "no checks reported" case ........................... exit=0 (NOTE)
     - all green CI ........................................ exit=0
  5. require-design-review-for-ui:
     - regex .tsx matches / .ts does NOT match ............. verified
     - PR with no UI changes ............................... exit=0

Two bugs were caught mid-smoke-test and fixed pre-commit:

- block-merge-on-red-ci.sh assumed "no checks configured" returns
  exit 8 but the pr-checks tool returns non-zero with "no checks
  reported" text. Fixed to pattern-match the text.
- require-design-review-for-ui.sh regex was .tsx? / .jsx? which
  matched plain .ts / .js (backend code). Fixed to .tsx / .jsx
  exactly.

## Review history

Rex approved cleanly on the first pass (commit d7c9641). Unlike the
previous two reviews where nice-to-haves turned out to be must-fix on
closer inspection, this time Rex tested the key regexes live against
filename fixtures and verified the severity calibration honestly.

Four nice-to-haves from the review, deferred to a follow-up ticket:

- ^Dockerfile misses monorepo paths (backend/Dockerfile). Fix:
  (^|/)Dockerfile. Same for ^docker-compose. Material for monorepo
  forks but mitigated by .architecture_paths project-config override.
- infrastructure/ unanchored — mild false-positive on paths like
  docs/infrastructure/. Fix: (^|/)infrastructure/.
- Conventional Commits breaking marker (feat!:) is rejected by both
  this hook and the existing PR-title regex. Consistent, not a bug.
- Dead code (FAILED_COUNT, PENDING_COUNT) in block-merge-on-red-ci.sh.

## Deferred to follow-up tickets

- docs/rule-audit.md — full audit table of every MUST with mechanization
  status. Pure docs, small scope.
- validate-branch-name.sh / validate-pr-create.sh warning→blocker upgrade.
  Breaking change, deserves its own ticket.
- commit_types project-config override.
- /approve-design skill (analogous to /approve-merge) for writing the
  design-review marker.
- The four nice-to-haves from this review (bundled as one small ticket).

## Dogfooding

All four new pre-commit hooks fired on this PR's own commit and allowed
it cleanly. The commit-format and AgDR-arch hooks became self-aware for
the first time — and correctly.

Closes #13

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* fix(#18): regex anchors for monorepo Dockerfiles + dead code cleanup

Addresses three of the four nice-to-haves Rex flagged on PR #17. The
fourth (feat!: Conventional Commits breaking marker) is tracked
separately as #23 because it needs a design decision.

## Changes

- require-agdr-for-arch-changes.sh: monorepo-safe anchoring. The
  original ^Dockerfile / ^docker-compose.* only matched root-level
  files and silently skipped backend/Dockerfile, web/docker-compose.yml,
  services/api/Dockerfile.prod — the common real-world layout.
  Updated to (^|/) prefix anchors.

  Also dropped two ambiguous directory patterns:
  - (^|/)infrastructure/ — matched docs/infrastructure/notes.md and
    src/types/infrastructure/foo.ts as false positives
  - (^|/)terraform/ — same ambiguity, redundant with \.tf$
  Terraform files are caught unambiguously via \.tf$ / \.tfvars$ at
  any depth. CDK / Pulumi projects that use plain .ts/.py files in
  an infrastructure/ directory need to override via
  .architecture_paths in project-config.json.

- block-merge-on-red-ci.sh: deleted unused FAILED_COUNT and
  PENDING_COUNT variables.

- .claude/hooks/README.md: updated the architecture-paths section
  with the new regex list and an explicit CDK known-gap callout.

- docs/agdr/AgDR-0001-rule-mechanization-hooks.md: new "Post-ship
  amendments" section at the bottom with a dated changelog entry.
  The original decision block is untouched — sets a pattern for
  threshold refinement without rewriting historical records.

## Smoke tests

21 path fixtures tested against the new regex list, all passing.
13 should-match cases (including monorepo layouts at various
depths) and 8 should-NOT-match cases (including the previously
false-positive docs/infrastructure/ and docs/terraform-primer.md).

## Review

Rex tested the patterns against 26 fixtures (5 beyond my 21) and
verdict was approved, clean. Two surfaced nice-to-haves:

- README known-gap callout could name Helm/K8s/CFN/SAM, not just
  CDK. Deferred to the SAM-coverage follow-up ticket.
- Dockerfile permissiveness is intentional — catches .prod/.dev
  variants correctly.

Material finding (not blocking, follow-up): SAM template.yaml is
NOT caught by the defaults. FlatMate specifically is SAM-heavy
and would benefit from a project-config override OR a default
addition for (^|/)serverless\.ya?ml$ (the word "serverless" has
no library-code collision, unlike "infrastructure"). Tracked as
a new follow-up ticket to be opened post-merge.

Rex also offered an amendment rule-of-thumb worth codifying:
changelog append for regex/threshold refinement and scope
additions, new AgDR for decision reversals and deprecations.
To be folded into #19 (the full rule-audit doc) or a new tiny
docs/agdr/README.md ticket.

Closes #18

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* ci(#2): dogfood apexstack CI with pr-title, markdown, shell, link checks

ApexStack now dogfoods its own CI. Four GitHub Actions workflows added
so PRs to the framework repo get the same quality gates it ships to
adopters via golden-paths/pipelines/.

- pr-title-check.yml — copied verbatim from golden-paths. Enforces
  ticket ID in PR titles.
- markdown-lint.yml — markdownlint-cli2 on **/*.md with relaxed config
  (.markdownlint.json disables MD013/MD033/MD041 etc. because existing
  docs have bare URLs, long lines, inline HTML).
- shellcheck.yml — shellcheck on .claude/hooks/*.sh at warning severity.
- link-check.yml — lychee on site/index.html + **/*.md, with weekly
  cron for link-rot detection.
- .markdownlint.json — relaxed ruleset. Real structural problems still
  caught (duplicate headings, broken fences, malformed lists).

Optional html-validate.yml skipped — site/index.html is 2000-line
marketing HTML that would need non-trivial config.

Rex approved. No must-fix. Three nice-to-haves deferred:
- ShellCheck will flag pre-existing SC2059 in validate-pr-create.sh
- markdownlint-cli2-action could be bumped from v16 to v23
- pr-title-check validates ticket-ID presence only (stricter check
  lives in the local hook)

Refs #2

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

* docs(#19): full rule-audit table at docs/rule-audit.md

Full rule-audit table at docs/rule-audit.md. Every MUST / NEVER /
HARD-STOP rule across CLAUDE.md, 8 .claude/rules/*.md files, and 3
workflows/*.md files, each mapped to its enforcement mechanism (hook,
agent, or prose) and classified as mechanize…