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
3.0 KiB
Error Recovery Protocol
Shared procedure for any skill that spawns agents via the Agent tool. Referenced from the
## Error Recovery Protocol section of each orchestration/review skill. Those
skills keep the two load-bearing imperatives (surface immediately, produce a
partial report) inline, and keep their own Common blockers list inline —
those are skill-specific. This file holds the full four-step procedure.
Trigger 0 — verify the artifact before reading the response
Before applying anything below, and before treating a phase as complete: if the return contract named a path, check that the path exists. A named artifact that is not on disk is a FAILED PHASE, no matter how the agent's response reads.
This trigger exists because the three below are not sufficient. They fire on
BLOCKED, on an error, and on "cannot complete" — and an agent can fail in a way
that is none of those. Observed live: a devops-engineer spawned
as half of a parallel phase ran 26 tool calls and ~31,600 tokens, then
returned a single fluent preamble — "I'll verify the build infrastructure…" —
with no path, no summary, no BLOCKED item, and no file written. Nothing in
steps 1–4 matches that shape. An orchestrator reading the transcript sees a
plausible response and advances, and half a parallel phase disappears with no
signal at all.
The work was not lost — resumed with an explicit "you did not honour the return contract" message, the same agent completed in one tool call, because its investigation was still in context. Only the write-and-return step had failed. So the recovery is cheap; the detection is the hard part, and the detection cannot come from reading the response.
A fluent response is not evidence that a phase ran. The artifact is — observe the artifact, not the transcript. The rule applies to sub-agents exactly as it does to skills.
When the path is missing:
- Do not advance the pipeline or summarise the phase as done.
- Resume the agent, naming the unmet contract explicitly and telling it that "I could not write" is a valid answer but silence is not. Prefer resuming over re-spawning — the context is usually still there, and a re-spawn pays for the whole investigation again.
- If the second attempt also produces nothing, treat it as BLOCKED and apply steps 1–4 below.
If any spawned agent returns BLOCKED, errors, or cannot complete:
- Surface immediately — report "[AgentName]: BLOCKED — [reason]" to the user before continuing to dependent phases.
- Assess dependencies — check whether the blocked agent's output is required by subsequent phases. If yes, do not proceed past that dependency point without user input.
- Offer options via
AskUserQuestion:- Skip this agent and note the gap in the final report
- Retry with narrower scope
- Stop here and resolve the blocker first
- Always produce a partial report — output whatever was completed. Never discard work because one agent blocked.