Compound Engineering Plugin synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Load instructions, retain host responsibility
Compound Engineering Plugin structures coding-agent work through skills and supporting programs. It is an extension, not the enclosing agent runtime. This account covers the shipped extension at commit 50532b00de523980010d97fe950965b35382a3b7, with an applicability cutoff of 2026-10-13. Its sole source, SRC-1, supplies inspected implementation and doctrine/design. No target execution, real project/session capture or causal experiment supports this analysis. The whole-system functional boundary includes distribution, project-learning producers and consumers, and repository-local improvement tooling; it does not imply an exhaustive behavioral audit (boundary, Source register).
Pi resource discovery and OpenCode registration deliver skill paths or bodies to the host. The install/conversion CLI instead writes host resources. Once a skill is loaded, the foreground model interprets its instructions; the host owns model calls, tools, native-worker scheduling and permissions. Registration is executable plumbing, not enforcement of the requested workflow. Static instructions, retained project evidence and controller authorization state therefore have different consumers and force (RT-RTE-host-loading, RT-RTE-resource-install, RT-BAP-skill-instructions, MEM-BAP-project-evidence, RT-BAP-controller-state).
Turn a request into bounded, verified work
The representative workflow is an explicitly autonomous new code-change request through lfg. Its shipped procedure connects planning, implementation, review, conditional learning capture and shipping; this is a prescribed workflow, not an observed successful session. The foreground agent selects a work source and admits the plan by its actual implementation scope and verification content. An explicitly blocked planning return remains terminal even when its artifact is readable (workflow account, RT-RTE-lfg-shipping).
Implementation runs through ce-work mode:return-to-caller, leaving the delivery tail with LFG. Independent units can use fresh native workers; dependent or unsupported-worker paths run serially or inline. Packets supply bounded plan excerpts, settled decisions, scope restrictions and verification obligations. Only summaries, evidence and changes should return from a worker's isolated context. The foreground orchestrator inspects the actual tree, runs authoritative verification and makes canonical commits. Shared-workspace ownership and isolated-base checks are model-facing policies, not plugin-enforced filesystem confinement. Reported pre-edit failures must be preserved because the final diff cannot reconstruct their timing (RT-RTE-native-work, workflow account).
The external-worker alternative adds executable protocol checks. Controller authorization binds recipient, unit/attempt/job identity, argv, packet digests and workspace before dispatch. Integration separately checks scope through host judgment, applies changes, runs supplied verification and rejects a failing or tracked-state-mutating check before canonical commit. Recovery restores tracked state or preserves blocked recovery information. These are wired checks on the bundled route, not proof that the verification command tests the intended behavior. The cooperative adapter approves permission requests; worktree isolation is not an out-of-workspace security barrier (RT-RTE-external-unit, RT-OBJ-run-manifest).
Child completion and caller acceptance remain separate. LFG allows one evidence-reconciliation retry for an otherwise implemented result, preserving the source and safe run identity rather than commissioning implementation again. Although ce-work can fall back natively when a required external engine is unavailable, LFG withholds shipping when a completed require return reports the wrong route. A disclosed prefer fallback can proceed (workflow account).
Admit review findings, then deliver without overstating completion
Review treats reviewer outputs as claims. Its helper implements schema, fingerprint and confidence partitions; semantic source checks and disposition remain model instructions. A nonempty evidence string is not authenticated quotation, and numerical confidence is not calibrated truth probability. Confirming a finding grants no new apply authority (EPI-RTE-review-findings). LFG applies only still-supported mechanical fixes within its narrower caller permission, verifies and commits them, and retains other actionable findings for human decision. A contract, permission, security or product-shape change cannot become automatic merely through high confidence (WF-RTE-review-admission).
With a remote, delivery requires pushed work, an open PR and a completed CI/review-feedback result—not merely a started watcher or a momentary green poll. Merge remains separately authorized. Without a remote, delivery is local commits with push/PR/CI watching skipped. Needs-human residuals are a decision handoff, not successful completion even if DONE appears. Some review-stage changes may already have been pushed before a later stop; stopping is not a promise of rollback (workflow account, RT-RTE-lfg-shipping).
Preserve reasoning for later decisions, with conditional warrant
The strongest supported learning contribution is an explicit route from experience to reason-bearing artifacts and later consumers. Capture keeps reasoning whose loss could cause recurrence, material risk or substantial rediscovery, rather than recording completion alone. Project solutions retain explanations and failed approaches; the glossary supplies canonical terminology. Ordinary qualifying content and glossary edits are model-admitted, while selected discoverability edits and ambiguous maintenance decisions have human gates (MEM-OBJ-solutions, MEM-OBJ-concepts, RT-RTE-learning-admission).
Full capture writes the candidate before grounding. Mechanical checks inspect citations and scaffold; a read-only model is asked to check behavior against local code, merge claims against remote status and counts against enumerated support. The orchestrator adjudicates the evidence, correcting contradictions and qualifying or dropping unverifiable behavior. Neither pass is an unconditional hard stop. Lightweight omits semantic validation. Thus the system specifies evidence-consuming acceptance criteria, but persistence is not an acceptance receipt and the source's trusted-knowledge claim exceeds any universal semantic guarantee (EPI-RTE-capture-grounding, EPI-CLM-trusted-knowledge).
Planning requests retained learning through an inline path or a researcher. Metadata, lexical filtering and model relevance judgments select full documents; the researcher distills up to five findings, and the planner must open the dossier when dispatched. The cap bounds finding count, not tokens. Current evidence must challenge conflicting historical claims. Declared packs supply imported rule evidence, not workflow authority; trusted URL/ref cache reuse establishes availability, not immutable provenance, freshness or source warrant. Review can also turn a quoted rule contradiction into a finding candidate, not an unconditional release veto (MEM-RTE-learning-retrieval, MEM-RTE-pack-resolution, MEM-RTE-review-retrieval, EPI-CLM-pack-enforcement).
Full capture can reduce archived sessions into lossy skeletons, synthesize selected findings and incorporate them into durable learning. Handoffs separately compress prior conversation/workspace experience into an immutable continuity snapshot; resumption provides orientation and waits for current-user confirmation. Optimization retains measured-result records and a model-written strategy digest, then requests the next hypothesis batch to read that digest and the last ten experiments. Managed handoffs may disappear from temporary storage; optimization state is gitignored and does not travel with the branch (MEM-RTE-session-capture, MEM-RTE-handoff-resume, MEM-RTE-experiment-digest).
These three prescribed automatic experience-to-artifact-to-later-consumer chains support trace learning in the write-route sense, not demonstrated improved capacity. Requested research, continuity and digest reads are pull operations; automatic selection within a requested return does not establish push. Opaque host and ancillary alternatives prevent a pull-only system conclusion (profile verification). Refresh can investigate, revise, consolidate, replace or withdraw guidance and repair links. Independently supported guidance survives conflicting implementation as a possible product regression; age and unverifiability alone do not justify deletion. Stale metadata is not shown to suppress planning delivery (EPI-RTE-refresh-curation).
Distinguish improvement capacity from exercised improvement
Repository-local authoring affords changes to the plugin's own instruction texts, with provenance-directed criticism, mechanical tests and fresh-context behavioral comparisons. Behavioral evaluations are best-effort rather than CI gates. Implemented grading produces evidence within supplied predicate/rubric domains; it does not automatically admit or deploy a skill revision (RT-RTE-skill-authoring, EPI-RTE-eval-grading).
At this authoring boundary, theory-builder condition 1—localized editable theory content—is afforded. Condition 2—decisions depending on that content—is claimed. Condition 3—content-directed criticism and resulting revision or changed reliance—is claimed. Condition 4—retained criticism shaping a next round—is claimed. Those statuses do not establish that all four operated. Reflection has an afforded own-method representation/editing path, but both causal directions remain unobserved; reflective theory building is claimed. Autonomous theory building is uninspected. LFG's hands-off grant is a different, instruction-supported operational mode, not evidence that computation performs every internal theory-building role (RT-RTE-skill-authoring, RT-RTE-lfg-shipping).
Whether criticism improved capacity for future action remains uninspected. Own-skill authoring supports a marked dispositional self-improvement pathway, aimed at closing demonstrated behavioral gaps in future published/installed sessions. No evidence-to-update-to-distribution-to-dependent-operation trace establishes exercised self-improvement, and no outcome comparison establishes successful improvement. Generic optimization is independently scoped: localized hypotheses/digests are afforded, consumption and iteration claimed, content-directed criticism uninspected, and reflection/autonomy uninspected; optimizing an external target is not automatically plugin self-improvement (RT-RTE-skill-authoring, MEM-RTE-experiment-digest).
For a plan-to-PR request, the source supports a carefully separated chain of authority, verification and residual handoffs, conditional on host/model compliance. For later avoidance of known mistakes, it supports reason retention and requested reuse, not measured avoidance. For method improvement, it supports addressable edits and comparisons, not an exercised causal loop. Linked accepted learning instances, later decision traces and comparable outcomes would alter the reuse assessment. A traced criticism-driven method revision, installed changed version and subsequent behavior depending on that revision would alter the self-improvement assessment; success would still require objective outcome evidence. Implemented semantic admission/quarantine, current pack-ref validation or enforced worker permissions would change the respective guarantee assessments rather than retroactively validate existing prose (EPI-CLM-compound-loop, EPI-RTE-capture-grounding, RT-RTE-external-unit, RT-RTE-skill-authoring).
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Retained-state and consumer inventory is partial; carries the report-verification inventory limit. | SRC-1; MEM-OBJ-solutions, MEM-RTE-learning-retrieval, MEM-RTE-experiment-digest, report limits | Principal learning/continuity paths; brainstorm delivery and dogfood, ideation and live-polish re-entry identified; pending-feedback, sweep and additional glossary/solution consumers lack complete leaf characterization. | Exhaustive memory, admission, expiry, selection and context-budget comparisons across shipped workflows. Supported principal routes remain valid. | Record-level inspection of each named alternative's operative parts, admission, persistence, selection, delivery and later consumer. |
| Own-method improvement coverage is selective; carries the report-verification improvement-plane limit. | SRC-1; RT-RTE-skill-authoring, EPI-RTE-eval-grading, memory write side, report limits | Authoring/evaluation pathway and retuning kernel/selected cut-pass findings; retune publication destination and complete later-consumer wiring unspecified in this account. | Complete improvement-plane inventory or comparison of all feedback routes; exercised self-improvement from retuning instructions. | Accepted records tracing retune archive, admission, retained measurement/reasoning artifacts and later consumption; execution/dependence evidence for exercised claims. |
| Some profile units group objects or consumption paths; carries the profile-verification limit. | SRC-1; MEM-OBJ-solutions, MEM-OBJ-pack-rules, MEM-OBJ-experiment-log, MEM-RTE-handoff-resume, MEM-RTE-experiment-digest, profile limits | Grouped storage, mixed experiment form and grouped requested-read units; partial axes retain local conditions and evidence bases. | Exact per-object counts, one-to-one unit comparisons, or inference that every grouped object shares every finding. | Finer units tied to separately recorded parts and consumers, preserving conditional Git retention and mechanism-specific evidence. |
| No observed compliance, accepted instance, useful activation or causal benefit. | SRC-1; EPI-RTE-capture-grounding, MEM-RTE-learning-retrieval, RT-RTE-skill-authoring, EPI-CLM-compound-loop | Frozen source implementation/doctrine; no real project/session or experimental capture. | Successful delivery/reuse, improved future capacity, exercised or successful self-improvement; semantic preservation, entailment or novel conjecture for particular generated prose. | Provenance-linked inputs, checks/dispositions, later decisions and outcomes; controlled comparisons at the claimed attribution level. |
| Host/provider behavior and deployed permissions remain external or opaque. | SRC-1; RT-CMP-host-model, RT-CMP-external-model, RT-RTE-host-loading, RT-RTE-external-unit | Inspected Pi/OpenCode loaders and cooperative external adapter; host scheduling, compaction, served models and provider state not frozen. | Uniform sandbox, exact served-model identity/fixity, complete model-window or host-memory classification. | Frozen host/provider contracts, model receipts and deployed permission/context traces for named routes. |
| Pack freshness, upstream warrant and session provenance are locally unresolved. | SRC-1; MEM-RTE-pack-resolution, MEM-OBJ-pack-rules, MEM-RTE-session-capture | Trusted URL/ref cache function and lossy Pi active-path/compacted-tail extraction; other copies/platform parsers not fully audited. | Universal current/immutable pack provenance, imported-claim certification, lossless original reasons or complete trace-source coverage. | Captured pack content/revisions and acceptance evidence; cache-refresh/copy audits; raw-to-derived mappings across supported parsers and host summaries. |
| Semantic warrant and withdrawal are policy- and mode-dependent. | SRC-1; EPI-RTE-capture-grounding, EPI-RTE-refresh-curation, MEM-RTE-learning-retrieval | Postwrite Full adjudication, Lightweight exception and refresh/current-evidence safeguard; no automatic stale exclusion demonstrated. | Persistence as acceptance, universal hard semantic gate, automatic stale suppression or rollback of every write. | Traced disposition/read-time behavior; executable gate, quarantine or stale-filter implementation covering the claimed modes. |
| Recovery, delivery and evaluation guarantees are bounded by selected paths and references. | SRC-1; RT-RTE-external-unit, WF-RTE-review-admission, MEM-RTE-experiment-digest, EPI-RTE-eval-grading, workflow account | Bundled integration transaction, LFG caller-facing native route and selected grading logic; full resume/fallback, child shipping, metric/decide scripts and rubric validity not completely audited. | Exhaustive recovery, universal shipping/test adequacy, authenticated peer independence, valid optimization objectives or component-level causal effects from grader output. | Remaining route/metric/transport inspection plus task-specific test/rubric validation, receipts and comparable execution evidence. |
No unresolved inter-report conflict was identified in the supplied reconciliation. No source reads were needed for this synthesis; it uses the supplied members only.