@open-agent-toolkit/cli 0.2.73 → 0.2.74

package/assets/skills/oat-project-implement/SKILL.md

@@ -7,7 +7,7 @@ disable-model-invocation: true
 user-invocable: true
 allowed-tools: Read, Write, Bash(git:*), Bash(oat:*), Bash(oat project log:*), Glob, Grep, AskUserQuestion, Task
 metadata:
-  version: 2.3.8
+  version: 2.3.9
 ---
 
 # Implementation Phase

package/assets/skills/oat-project-implement/references/completion-and-closeout.md

@@ -890,64 +890,52 @@ closeout. The lite contract sets `PROJECT_RECAP_REACHABLE=false` and proceeds
 from the required reviews through its stored optional steps to `pr` and
 sequence completion.
 
-Before generating, inspect the active project's explainer runs. A fresh `project-recap` manifest for the current completed implementation deduplicates the lifecycle-tail run: reuse it and do not invoke the adapter again. Fresh means the manifest identifies recipe `project-recap`, belongs to this project, has a terminal outcome, and its recorded source hashes match the current approved implementation inputs. A merely present, incomplete, wrong-recipe, or stale manifest does not satisfy this check.
-
-Probe seam availability before resolving intent. Call
-`oat-explainer-kit/scripts/probe-recap-seams.mjs#probeRecapSeams` in
-`mode: unattended` with the exact seam inputs this tail would pass. The probe is
-pure and covers all five required seams — author, fact critic, browser session,
-visual critic, and set planner — so a host missing only the set planner is
-detected here instead of at the adapter's `E_SET_PLANNER_REQUIRED`. Pass the
-result to autonomous intent resolution as `seamProbe`. The resolver accepts a
-`seamProbe` only for autonomous `projectRecap`, so interactive resolution keeps
-its existing inputs and its recorded human decision.
-
-Resolve recap intent through `oat-explainer-kit`. When `OAT_AUTONOMOUS=1` and no fresh recap exists, run this recap gate exactly once; missing or stale persisted intent cannot suppress this autonomous gate. In autonomy, attempt the adapter run exactly once and only when the seam probe resolves every required seam and intent resolves to `generate`; an interactive `generate` still attempts the run regardless of the probe result, and a seam-less interactive attempt is still the `failed` outcome it is today. The autonomy gate and the interactive rule are two separate rules and are never read as one. Interactive mode honors the adapter's resolved persisted or workflow intent.
-
-In autonomy, a probe result of `seams-unavailable` means no provider is
-configured for a required seam. Autonomous resolution then returns a recordable
-`skip` with source `capability_probe`: record it with the warning, do not invoke
-the adapter, and continue closeout. That probe-driven skip record supersedes the
-intent resolved and persisted earlier in this run for the remainder of the run:
-persist it through the same `oat-explainer-kit` intent-persistence helper with a
-freshly captured state hash, and pass the skip — not the earlier `generate` — to
-the terminal-outcome guard as `--intent skip --skip-reason capability_probe`.
-Passing the superseded `generate` with no manifest raises `E_RECAP_OUTCOME` and
-blocks approval, which is the exact failure this gate exists to prevent. Interactive closeout is unchanged: a
-recorded interactive `generate` still attempts the recap and a run that fails
-for a missing seam is still the `failed` outcome it is today. An unattended completion on a host with no explainer seams
-configured never blocks final HiLL approval on a missing recap, and this skip is
-resolved before any run, never from a failed one. A probe result of
-`seams-invalid` means a seam is supplied but violates a resolution rule: report
-the configuration error and fail closed exactly as an invalid seam does today.
-Never convert a configured-but-invalid seam, or a run that failed after a
-passing probe, into a skip; that run stays `failed`.
-
-Invoke the `oat-explainer-kit` adapter first, then run its shared tracked-run finalizer in `dedicated` mode for a successful build. Use the adapter result and finalizer result as returned; do not improvise commits, durability evidence, or reruns. Outcomes `failed` and `built-not-durable` are recorded warnings, never blockers for final HiLL approval, completion reporting, or later PR steps.
-For an adapter invocation, construct exactly one brief-aware,
-provider-neutral author seam using
-`oat-explainer-kit/references/author-callback.md`. In-process callers pass
-`author`; JSON/CLI callers pass a validated `authorModulePath`. Supply it
-alongside the existing `critic` callback (or validated
-`criticModulePath`), and always invoke this implementation-tail recap with
-`mode: unattended`.
+Re-read the persisted `oat_project_recap` decision from `state.md` at this
+consumer boundary. A persisted `skip`, including `skip/failed_attempt`,
+suppresses manifest discovery and generation: do not bundle or author, do not
+invoke the adapter, and leave the selected run empty. Pass the persisted source
+to the terminal-outcome guard as `--skip-reason`; a failed-attempt skip also
+passes only the deployed consumer's containment-checked
+`failedAttemptEvidence`, sourced from the persisted project-relative
+`failed_attempt_evidence` locator.
+
+For a persisted `generate`, inspect the active project's explainer runs. Reuse
+a fresh satisfied `project-recap` package and do not invoke the adapter again.
+Fresh means the package identifies this project and recipe, has outcome `built`
+or `built-needs-review`, passes complete package validation, and its source
+hashes match the current approved implementation inputs. A merely present,
+`failed`, `incomplete`, wrong-recipe, or stale package does not satisfy this
+check.
+
+When no fresh package exists, invoke the `oat-explainer-kit` adapter's
+§ Generate with recipe `project-recap` and `mode: unattended`. Run
+`bundle` → host-agent authoring → `verify` → `record`; lifecycle generation
+never prompts. A missing browser records `built-needs-review` and does not
+block final HiLL approval.
+
+`built` and `built-needs-review` satisfy generation. `failed` and `incomplete`
+do not. In autonomy, retry once after either unsatisfied outcome. If the retry
+also fails, persist `skip/failed_attempt` with `failed_attempt_evidence` set to
+the attempted run's project-relative failed/incomplete `manifest.json` or
+`failure.json`. Re-read that persisted skip through the deployed consumer and
+continue without another bundle or authoring pass. Never turn the first failed
+attempt into a silent skip.
 
 Always include the selected or attempted recap's outcome and run path in the
-implementation completion report. Report a probe-driven skip as `skipped` with
-its reason and the unavailable seams instead of an outcome and run path. If
-`summary.md` exists, append or refresh its single concise `Explainer Outcome`
-section using the manifest and build record, or the recorded skip when no run
-exists; never append a second outcome section. If no recap was attempted,
-reused, or recorded as skipped, leave the summary unchanged.
-
-Before recording final approval, invoke the shared
-`oat-explainer-kit/scripts/check-terminal-outcome.mjs` guard with the resolved
-intent and, for `generate`, the selected or attempted manifest. The only
-terminal generated outcomes are `built-durable`, `built-not-durable`,
-`built-needs-review`, and `failed`. Missing records and `incomplete` block
-approval; do not substitute a warning or infer an outcome from filesystem
-presence. A `skip` intent requires no manifest; pass its recorded source as
-`--skip-reason` so the receipt states why no recap exists.
+implementation completion report. Report a persisted skip as `skipped` with
+its reason and no run path. If `summary.md` exists, refresh its one concise
+`Explainer Outcome` section from `manifest.json` plus `qa/result.json`, or from
+the persisted skip; never append a duplicate.
+
+Before recording final approval, invoke
+`oat-explainer-kit/scripts/check-terminal-outcome.mjs` with the persisted
+intent. For `generate`, pass the selected package's canonical `manifest.json`.
+For `skip/failed_attempt`, pass the validated canonical path and evidence kind
+returned by the deployed consumer together with the active project as
+`--project-root`; other skip sources carry no evidence.
+The outcome vocabulary is `built`, `built-needs-review`, `failed`, and
+`incomplete`; only the first two satisfy generation. Missing packages block
+approval.
 
 **Lite completion algorithm:** Dispatch incomplete `pre_approval` steps in
 stored order, retain `approval: not_required`, keep both approval reachability