Compound Engineering Plugin report verification

Type: agentic-system-analyses/types/agentic-system-verification.md

Verification

The reports support publication as a bounded source account, with the inventory limits below. I read every supplied member and its governing contract. The correction-answer inputs are empty and the structural report-check says none; there is no structural failure to route. I independently checked consequential source behavior rather than accepting the analysts' validation statements. SRC-1's checkout HEAD matches the boundary revision and its working tree is clean. No target code, tests, examples, hosts or model services were executed.

Runtime, authority and delivery

RT-RTE-host-loading, RT-OBJ-skill-files and RT-BAP-skill-instructions correctly separate registration from model compliance. SRC-1, .opencode/plugins/compound-engineering.js, registers bodies and session callbacks on V2, and paths/templates with conflict preservation on V1; .pi/extensions/compound-engineering.ts returns resource paths. Installer source selection, target writes and cleanup in src/commands/install.ts support RT-RTE-resource-install, not semantic approval or transactional multi-target rollback. The two components, RT-CMP-host-model and RT-CMP-external-model, correctly leave provider mutability and actual model resolution uninspected. Research tier policy and fixed-recipient dispatch do not establish immutable weights.

For RT-RTE-external-unit, RT-OBJ-run-manifest and RT-BAP-controller-state, I traced SRC-1, skills/ce-work/scripts/unit_workspace_jobs.py's runner-contract and authorization functions, skills/ce-work/scripts/cross-model-work.sh's authorization handshake, and the complete integration transaction in skills/ce-work/scripts/unit_workspace_transaction.py. Identity, argv, authorization bytes, packet digests and result-directory checks precede binding; integration checks exit status and snapshot stability before committing. Recovery distinguishes precommit restoration from postcommit finalization. These support wired protocol checks, not product correctness, universal confinement or actual successful operation. The report preserves ignored-state and cooperative-permission exceptions.

RT-RTE-native-work and RT-CLM-verified-change remain instruction-supported, not observed verification. SRC-1, skills/ce-work/references/execution-strategy.md, skills/ce-work/references/implementation-loop.md and skills/ce-work/references/return-to-caller.md, support bounded fresh-worker packets, actual-tree inspection, foreground verification, failure-driven reconsideration and return-time plan reload. The workflow annotations correctly preserve evidence that cannot be reconstructed from a final diff.

RT-RTE-lfg-shipping, WF-RTE-review-admission and the workflow account are supported by SRC-1, skills/lfg/references/work-return.md, skills/lfg/references/review-followup.md and skills/lfg/references/shipping.md. I checked the one evidence-reconciliation retry, mechanical-fix criteria, residual delivery, pre-capture push possibility, no-remote terminal and completed babysit requirement. The child's native fallback under require and LFG's subsequent mismatch veto are different admission points, not an unresolved contradiction. Needs-human residuals are not successful completion merely because DONE is emitted. Detailed child implementation remains outside this selected caller-facing trace.

Retained material and epistemic functions

RT-RTE-learning-admission, MEM-OBJ-solutions, MEM-OBJ-pack-rules, MEM-OBJ-concepts and their epistemic counterparts are supported by SRC-1, skills/ce-compound/SKILL.md, skills/ce-compound/references/assembly.md, skills/ce-compound/references/lightweight.md and skills/ce-compound/assets/resolution-template.md. I checked eligibility, overlap, destination selection, glossary side effects and write/check ordering. The reports correctly distinguish per-content decisions from merely initiating capture. Their reasons-bearing templates do not imply complete reason preservation through later distillation.

EPI-RTE-capture-grounding preserves the important qualification in SRC-1, skills/ce-compound/references/grounding-validation.md: retention precedes grounding, mechanical flags require adjudication, and Lightweight omits semantic validation. The script's checks in skills/ce-compound/scripts/validate-doc-claims.py concern citations and scaffold, not arbitrary factual truth. The ledger's wired mechanical checks and doctrine-only semantic disposition are compatible with the broader runtime's claimed admission policy. EPI-CLM-trusted-knowledge is an attributed claim under comparison, not an analyst certification of every retained document.

For MEM-RTE-pack-resolution and EPI-RTE-pack-resolution, I independently read SRC-1, skills/ce-compound/scripts/packs-resolve.py's Git acquisition function. URL/ref hashing, immediate trusted-cache return and clone/fetch fallback support reusable access, not immutable provenance or current mutable-ref freshness. MEM-RTE-learning-retrieval, EPI-RTE-learning-delivery, EPI-OBJ-learnings-dossier and MEM-BAP-project-evidence match SRC-1, skills/ce-plan/references/agents/learnings-researcher.md and skills/ce-plan/references/research.md: conditional dispatch, different pack/local-corpus filtering, full relevant-document reads, up to five findings, mandatory dossier opening when dispatched and current-evidence conflict reporting. This establishes prescribed delivery, not semantic preservation or useful activation.

MEM-OBJ-session-skeleton and MEM-RTE-session-capture correctly retain the lossy lineage. SRC-1, skills/ce-compound/references/session-history.md, prescribes bounded archive filtering and historian synthesis; skills/ce-compound/scripts/session-history/extract-skeleton.py selects the active parent path and latest resolvable compaction boundary and truncates displayed summaries. It does not retain the entire original archive in its derivative. MEM-OBJ-handoff and MEM-RTE-handoff-resume match SRC-1, skills/ce-handoff/references/create.md and skills/ce-handoff/references/resume.md: temporary storage limits, metadata-only discovery, selected-source orientation and current-user confirmation, not restored action authority.

EPI-RTE-refresh-curation and the memory annotation on learning admission preserve the distinct conditions in SRC-1, skills/ce-compound-refresh/references/classify.md, skills/ce-compound-refresh/references/modes.md and skills/ce-compound-refresh/references/per-action-flows.md. Independent guidance can survive implementation conflict; unverifiability alone is not falsehood; consolidation, successor replacement and deletion differ. Stale metadata is not demonstrated exclusion at the planning reader. No blanket human-approval or automatic rollback claim follows.

MEM-RTE-review-retrieval, EPI-OBJ-review-findings and EPI-RTE-review-findings correctly connect distinct selection, checking and disposition functions. SRC-1, skills/ce-code-review/references/select-and-route.md, skills/ce-code-review/references/dispatch-reviewers.md, skills/ce-code-review/references/finish-review.md and skills/ce-code-review/scripts/findings-mechanics.py, support conditional historical-evidence selection, candidate partitions and separate semantic validation/apply authority. I checked counterexamples to a universal confidence threshold and to treating nonempty evidence strings as authenticated quotations. EPI-CLM-pack-enforcement consequently remains policy-mediated consideration, not an unavoidable release barrier. EPI-OBJ-debug-findings and EPI-RTE-debug-investigation faithfully scope SRC-1, skills/ce-debug/references/investigate.md's hypotheses, exact-symptom reproduction and discriminating predictions; they do not claim an executed causal diagnosis. EPI-OBJ-work-evidence correctly differs from a controller receipt.

Improvement, relations and omission disposition

MEM-OBJ-experiment-log, MEM-OBJ-strategy-digest and MEM-RTE-experiment-digest match SRC-1, skills/ce-optimize/references/persistence.md and skills/ce-optimize/references/loop.md: measurement markers, local gitignored persistence, integration-before-kept status and digest/last-ten read-back. The report withholds executable metric-validity and improvement claims. RT-RTE-skill-authoring matches SRC-1, .agents/skills/ce-skill-work/SKILL.md, .agents/skills/ce-skill-work/references/evaluate.md and .github/workflows/ci.yml: localized rule editing, provenance-directed criticism, fresh comparisons, explicit skip conditions and mechanical CI gates. Its theory-builder conditions, reflection and dispositional self-improvement are separately scoped; none is promoted to observed iteration or benefit.

EPI-OBJ-eval-artifacts and EPI-RTE-eval-grading distinguish encoded deductions from model grading. I checked SRC-1, tests/skill-eval-cell/grade.ts and tests/skill-eval-cell/judge.ts: missing-trailer failure, separate workspace/shim predicates, shuffled anonymous grading and incomplete-grid failure. The supplied concrete deduction holds over its stated premises. EPI-RTE-pi-conversion's concrete text transformation also holds under SRC-1, src/converters/claude-to-pi.ts; it does not establish equivalent host behavior. EPI-RTE-legacy-cleanup is supported by the Pi branch in src/commands/cleanup.ts and guarded rename in src/targets/managed-artifacts.ts, not proven host-cache invalidation.

I checked every declaration's evidence layer and status, each Part of: relation and all five supersessions in reconciliation. Capture, session enrichment, maintenance and later planning can remain parts of the broad learning container; work evidence is an object within native implementation; evaluation is within authoring; caller review admission is within LFG. The five supersessions identify the same solutions, pack rules, glossary, pack-resolution route and planning-retrieval route without discarding lens-specific findings. The other records have distinct targets, producers or consumers. No evidenced-absence records are declared. The reconciliation's postwrite grounding, stale-marking, mutable-cache and fallback dispositions preserve the findings rather than selecting stronger statuses. No unresolved inter-report conflict needs correction.

Selected omission questions came from the frozen top-level tree and shipped producer/consumer dependencies: whether another consumer makes the learning loop planner-only; whether another retained-state path makes the memory inventory complete; and whether archive-driven retuning warrants a stronger exercised-improvement conclusion. SRC-1, skills/ce-brainstorm/references/dialogue.md, adds pack-to-dialogue/dossier delivery with separate context budgets. skills/ce-dogfood/references/phases.md, skills/ce-ideate/references/post-ideation-workflow.md and skills/ce-polish/references/live-loop.md expose additional checkpoint/re-entry paths. skills/ce-retune/references/cut-passes.md requests measurement artifacts and a reason-bearing write-up for later readers. These corroborate the disclosed partial inventory and improvement affordances, not a contradiction of the bounded conclusions or evidence of actual operation.

Then write the finding down where the next person will hit it: the mechanism, the before and after, the measured numbers, and the hypotheses that died. --- skills/ce-retune/references/cut-passes.md

Further omission expansion stopped because these alternatives have supported local dispositions, the emitted findings do not assert complete route enumeration or system-wide absence, and no identified material challenge remains undisposed. Full parser, alternate adapter, live endpoint and other workflow audits would expand inventories rather than presently overturn the supported claims. Missing real sessions, stores, provider evidence and causal comparisons remain beside the claims they already qualify; they do not establish absent learning, acceptance or adaptation. EPI-CLM-compound-loop remains an intended benefit compared with source-supported paths, not a measured future-capacity result.

Blockers

none

Limits

• The retained-state and consumer inventory is partial: MEM-OBJ-solutions, MEM-RTE-learning-retrieval, MEM-RTE-experiment-digest and the memory coverage account cover principal learning and continuity routes but not complete classifications of brainstorm, dogfood, ideation, live-polish, pending-feedback and sweep state. SRC-1's skills/ce-brainstorm/references/dialogue.md, skills/ce-dogfood/references/phases.md, skills/ce-ideate/references/post-ideation-workflow.md and skills/ce-polish/references/live-loop.md show additional delivered context and disk-backed re-entry. The supported project-learning, session-enrichment and experiment-digest findings hold; withhold exhaustive memory/admission/expiry and context-budget comparisons across every shipped workflow.
• Own-method improvement coverage remains selective: RT-RTE-skill-authoring, EPI-RTE-eval-grading and the memory write-side account do not fully characterize the separate retuning route. SRC-1, skills/ce-retune/SKILL.md and skills/ce-retune/references/cut-passes.md, prescribe archive-based baselines, independently contextualized audit/defense, measured passes and retained measurement/reasoning artifacts. The publication destination and complete later-consumer wiring remain unspecified in this inspected account. The authoring pathway and independently uncertain learning/reflection/autonomy findings hold; withhold a complete improvement-plane inventory or comparative ranking of its feedback routes. No exercised self-improvement follows from these instructions.