v0.12.0

Findings of a week-long manual run on a real project (Web Studio 0.11.1, ten sessions, 250 agent runs) and of an
external PHP project. Nothing new to learn: the same commands, the same agents — cheaper, and harder to bypass.

• One story per session (coordination-rules rule 13). 80 % of the week's spend was the parent's context, not the
  agents' work: sessions ran to 966 k tokens with two compactions in ten sessions, and 85 % of requests re-read more than
  200 k. /story-done now hands off to /clear before the next story (continuing in the same conversation is no longer
  an offered option), /dev-story refuses to start a story past 50 % context without offering /clear first, and
  /code-review sends a re-review after fixes to a fresh session past that mark. Playbook § 10.25 says the same in two languages.
• Commands carry their namespace. A bare /code-review in a hand-off ran Claude Code's built-in review skill of the
  same name — nine general-purpose agents, no routing table, no appsec — and its findings travelled under the studio's name;
  once a security-sensitive story was closed on its word. Every hand-off in dev-story, refactor, code-review,
  story-done and every line /help prints now names /web-studio:<command> in plugin mode; coordination-rules § Subagents
  says why, next to the same rule for subagent_type.
• /help no longer starts the next command. It runs on Haiku, and a skill started from its closing question inherited
  that model: a /code-review and a /story-done ran on Haiku and closed a story with invented numbers. After the answer
  /help prints Run: /web-studio:<command> and ends the turn.
• The studio's review has to be in the log. /story-done DoD: a reviewer's SubagentStop after the branch's first
  commit, appsec-engineer among them for a security-sensitive story — otherwise NOT DONE (no studio review).
• A verdict is the reviewer's. After fixes /code-review sends the fix diff to the reviewer that raised each finding
  and moves NEEDS CHANGES to APPROVED only on its answer — never on a green CI (five of five verdicts had been flipped by
  the parent). A BLOCKING is downgraded only by the reviewer that raised it or by technical-director through /impact.
  Reviewers get the diff as one file and open every verdict with Read: N/M files; N < M is PARTIAL and never
  counts towards APPROVED (35 % of lead reviews had been cut off mid-diff).
• Engineers write the code, in every mode. The rule lived in dev-story and refactor; on an external PHP project
  the session drifted into "faster to fix it myself" from the third review round. It is now in coordination-rules
  § Subagents, principle 8 of the CLAUDE.md template, code-review Phase 5 and hotfix — and it has an observer:
  hooks/parent-write.sh (PreToolUse Write|Edit, warn-only) names the engineer the roster has for a code file the
  session writes itself and leaves a ParentWrite line in agent-audit.log; agent-stats.sh and /sprint-status print the count.
• Turn budgets and briefs. maxTurns 30 → 60 for the implementing Tier-3 engineers, 25 → 40 for the leads and
  appsec-engineer (go-engineer hit its limit on 55 % of runs, test-engineer on 50 %, backend-lead as a reviewer on 35 %).
  The dev-story brief has six fixed lines (Story · Read with line ranges · Write · Check · Skip · Report) because
  half of an engineer's turns went into grep/sed navigation of files the parent had already read; every agent's
  protocol gains point 8 — Read/Grep on the brief's paths, shell search only when the path is unknown, and a
  Checkpoint: line kept in the result so a cut-off hands the caller the point to resume from. After a cut-off the parent
  sends one resume line and runs no checks of its own until the SubagentStop (50 of 64 cut-offs had been answered with
  the parent's own git status and builds).
• The roster says what the tiers do. Leads review, design and name the specialist; the session dispatches. The Task
  tool and "routes work to …" leave the seven leads — measured on real projects, no lead ever spawned an agent.
• Hooks, four false positives closed and one gap: validate-commit blocks provider-shaped credentials everywhere but only
  warns on password = "…" shapes in test code and agent memory (three test files had been rewritten to please it);
  validate-push reads the commit from a git commit … && git push call instead of an empty rev-list (docs commits
  on the default branch cried wolf three times a day); written-paths resolves p='…'; open(p,'w') and
  Path(p).write_text(…) (27 document writes in one session had passed every guard); consent-guard stays silent on a
  one-line Status:/Started:/Actual: edit, on a command that sets the marker before writing, and on
  docs/ops/measurements/ (rule 12 asks for those). hooks/session-state.sh is executable — skills call it as a
  command and one call failed with Permission denied; the structure linter checks the bit for every hook a skill calls.