DonchitosandClaude Opus 5
d05699707f
`project.yaml` at the repo root is now the single source of truth for a project's configuration, replacing `production/stage.txt`, `production/review-mode.txt` and `.claude/docs/technical-preferences.md` (all three still read as a fallback; nothing is force-deleted). **`modes.rigor` — one question instead of six.** `minimal`, `standard` or `full` fronts six knobs: `modes.workflow`, `docs.density`, `qa.level`, `modes.story_granularity`, `modes.review_mode` and `team.size`. It defaults to `minimal`, which is a change from 1.0's behaviour: the heavier tier was measured costing several times more to reach working code without producing a better result, and returning nothing when a fresh developer inherited the project. Pin `modes.rigor: standard` to keep 1.0's process level. Also in this release: - `/settings` to view or change any setting, and `project.local.yaml` for per-developer preferences that should not be committed. - `system_overrides` to hold one system to a higher standard than the rest of the project, and `testing.strict` to set per test type whether missing evidence blocks a story or warns. - A story that changes anything the player sees is not closed until the game has been launched, observed, and a screenshot retained in `production/qa/evidence/`. A parse check is not a run. This holds at every rigor level, including `minimal`, where automated tests are waived but the look is not. - The code root is engine-specific and enforced as such: Godot `src/`, Unity `Assets/`, Unreal `Source/<Module>/`. Two of the three will not build a project laid out any other way. - Migration tooling for existing projects: `.claude/scripts/migrate-v1-config.sh` converts a 1.0 layout, non-destructively by default, with a report to review before `--finalize` deletes anything. - Ask-before-write holds regardless of personal automation settings. Upgrading from 1.0 or earlier: see UPGRADING.md. Full list of changes: CHANGELOG.md. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_016xvzG3i7pWQg4LjyDwWMXd
2.5 KiB
paths
| paths | |
|---|---|
|
Agent Memory Rules
Agent memory is the one place an agent may write without asking — and the only
one. .claude/agent-memory/ is gitignored, per-agent, and never ships. That
exemption exists because the memory is the agent's own working notes about how to
do its job, not a project artifact.
It does not extend anywhere else. The Collaboration Protocol still governs
every file under src/, design/, docs/, production/ and assets/: ask
first, naming the path. Writing a memory is never a substitute for asking, and a
memory must never be used to record something the user declined to have written.
Never record a temporary absence as a durable fact
This is the failure this file exists to prevent, and this repo is unusually exposed to it.
CCGS ships as a template the user clones as their game. So on day one every
agent observes the same things — no src/, no GDDs, no engine configured, no
tests/performance/ — and every one of those observations is guaranteed to
stop being true. A memory reading "this repo is the framework, not a game;
perf requests have no target" is accurate when written and actively harmful two
weeks later, when it tells a future run to return BLOCKED on a project that now
has a build. The agent will not re-derive it; that is what memory is for.
So:
- Prefer recording a method over a state. "Confirm an engine project and a captured profile exist before profiling; return BLOCKED naming what is missing rather than estimating" is durable. "There is no game here" is a timestamp.
- If you must record a state, state what invalidates it, on the same line:
INVALIDATED WHEN: project.yaml declares an engine, or src/ contains game code.A reader with no other context must be able to tell whether the note still holds. - Never record absence of a game, of assets, of tests or of config as a settled property of the project. Those are the states CCGS exists to move a user out of.
Disclose the write
Say in your response that you recorded a memory and where. An unmentioned write
is indistinguishable from no write, and the user cannot correct a note they do
not know exists — the same reason a skipped check has to announce itself
(.claude/rules/skill-authoring.md, obligation 3).
Verify before you trust
A recalled memory reflects what was true when it was written. If it names a file, a path, a setting or a count, check that it still holds before acting on it. Memory is a starting point for investigation, never evidence.