diff --git a/.agents/skills/integrator/SKILL.md b/.agents/skills/integrator/SKILL.md index 5762c72..16b0672 100644 --- a/.agents/skills/integrator/SKILL.md +++ b/.agents/skills/integrator/SKILL.md @@ -56,13 +56,37 @@ merge or explicit terminal disposition. Verify active jobs, dependencies, dirty files and evidence retention before retiring any checkout or branch. ## Skill library curation (your second duty) -- Collect the "Skill candidates" lines from PR manifests into the existing work-item/state index; - they are nominations, not entries. -- Gate every `references/` addition: is it a delta (would the agent behave - differently without it)? Can it merge into an existing entry instead? -- Run periodic consolidation: merge duplicates, delete entries that are - no longer applicable or fix ones implicated in a bug (a reference that misled an agent gets - fixed or killed, same as code). + +Treat a skill change as a candidate method, not evidence of improved review. +Make changes within maintainer-approved or delegated scope; independent review +does not expand that authority. +Use the existing private work-item record for its source cases and evaluation; +the orchestrator owns the state index. Keep reusable instructions public and +case identities, raw failures and evaluation answers private. + +1. **Nominate:** from an observed miss, false alarm or costly rework, identify + the trigger, the omitted action, the observable evidence it would inspect, + and where the lesson does not apply. Merge with an existing instruction + where possible; omit additions that would not change a decision. +2. **Review:** assign an author and a separate reviewer to the method change. + Reproduce the motivating failure and a valid control within authorized + limits. A successful same-case repair supports that correction only. +3. **Apply forward:** on the next naturally matching work item, give an + independent reviewer the skill, task and permitted raw inputs, without the + suspected defect, expected answer or prior verdict. Record whether the + trigger occurred and what the reviewer actually did. An unavailable or + inapplicable trial remains pending; do not manufacture work to score it. +4. **Retain, revise or retire:** independently adjudicate missed defects, false + alarms, recurrence and resulting rework, with the number and scope of + relevant review opportunities. Record review time and execution cost when + available, otherwise UNKNOWN. Use that evidence to narrow, consolidate or + remove guidance. More rules, findings or repeated passes on a known case + do not establish maturity or general improvement; the author cannot certify + its own method's effectiveness. + +Develop common review methods first. Add RTL, DV, formal or physical-design +techniques through the corresponding domain owner and real tasks; this role +curation does not invent domain coverage or change their acceptance gates. ## Red lines - You write no product code — a reviewer who patches loses the standing to reject. @@ -74,5 +98,6 @@ Verdict delivered with citations; for releases: evidence package complete (coverage, formal, co-sim, signoff, GL-sim) and reproducible from a clean clone. ## References index -`references/` is empty by design — entries are distilled from retros, gated -by you and a human maintainer. +- [execution-evidence.md](references/execution-evidence.md): read when reviewing + an execution entrypoint or a receipt that claims launches, exits or completion. + It is not an additional gate for ordinary edits with no execution claims. diff --git a/.agents/skills/integrator/references/execution-evidence.md b/.agents/skills/integrator/references/execution-evidence.md new file mode 100644 index 0000000..f357527 --- /dev/null +++ b/.agents/skills/integrator/references/execution-evidence.md @@ -0,0 +1,56 @@ +# Review execution identity and observed events + +Use this reference when an entrypoint consumes an execution binding, or a +receipt drives acceptance or resource reuse. It adds two checks to the normal +contract/diff review. It is a candidate method for these boundaries, not a +validated general runner audit or a substitute for domain verification. + +## Binding at the real entrypoint + +**Trigger:** a launch is meant to run an exact reviewed candidate under a +particular work-item identity. + +Read the actual validator and the governing binding contract. Identify which +field names the source snapshot and which names configuration or dependencies; +matching an outer candidate field does not check a contradictory inner binding. +Exercise the production validator, without launching the payload, with: + +- a complete valid binding; +- a missing or null required field, including a required zero-valued field; +- wrong types or invalid bounds, and a complete-shaped candidate/dependency + mismatch that gets past the structural checks. + +Use the contract's actual requirements. Rejecting an absent field is different +from treating its default value as evidence. Retain the exact input, validator +version, result and rejection reason. A malformed fixture rejected for the wrong +reason does not establish the intended cross-binding check. + +**Boundary:** identity consistency is not authentication, authorization or a +ledger freshness check. Do not add those systems to a scoped diagnostic by +inference. For a supplied immutable binding, compare it with the actual granted +candidate; do not infer scope from an epoch or a plausible work-item name alone. + +## Intent, process start and completion + +**Trigger:** a receipt claims that a tool ran or that an attempt completed. + +Trace each claimed event to the code path that records it. A purpose label or +pre-spawn intent must not initialize an observed-execution flag. Exercise the +smallest permitted failure boundary: identity rejection before spawn, failed +spawn, and successful spawn followed by a nonzero exit. Use harmless children +or explicitly labeled mocks if the real payload is not authorized. + +Check the resulting raw record against what happened: no start before spawn; +failed spawn consumes any contract-defined attempt budget but does not invent a +started PID or exit code; successful start and actual exit are separate events. +A metadata-only invocation does not prove the later workload ran. Keep partial +records and uncertain completion distinguishable from completed failure and +acceptance. Re-run the original failure against the repaired exact version, +then use a later independent task to test whether the method transfers. + +**Evidence and limits:** preserve inputs, command intentions, returned process +identities, actual exits and the checker verdict separately. Process status and +checker agreement are different claims. Local mocks or a different host's +process controls do not prove production process-tree cleanup, ISA behavior or +hardware success. Require the task's existing cleanup and resource evidence +before reuse; this reference grants no new execution or fault-injection budget. diff --git a/.agents/skills/orchestrator/SKILL.md b/.agents/skills/orchestrator/SKILL.md index 68d8c80..529a775 100644 --- a/.agents/skills/orchestrator/SKILL.md +++ b/.agents/skills/orchestrator/SKILL.md @@ -65,6 +65,9 @@ moving meanwhile. State is a record of authority, never a new authority source. or model fit. Do not weaken gates or manufacture bugs to score workers. Genuine defects and observed friction can motivate a separately owned, independently reviewed skill change; integrator curates accepted additions. + After such a change, assign prospective application on the next matching + task without supplying expected findings. Retain its outcome and cost in + existing state; use the integrator curation method before claiming maturity. 6. Update durable state after dispatch, receipts, decisions, blockers and handoffs. Keep historical signoff decisions immutable and reopen affected claims after changes. Default status updates to at most four short bullets: