EQUIPA synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From task request to candidate work
EQUIPA is an enclosing runtime for model-driven software work. It selects agents, owns task phases and budgets, admits product changes, and retains experience for later invocations. Claude Code supplies an inner execution engine; EQUIPA also ships an Ollama tool loop. This account covers the shipped system at Git revision fa8f41a6948584ee53e804edbb08b7449d05cd14, with an analysis cutoff of 2026-10-15. Its sole registered source, SRC-1, supplies implementation, design instructions and attributed operational claims—not an observed execution or causal experiment. Whole-system scope includes interactive skills and maintenance programs, but does not mean every included path received equal inspection. Live configuration, databases, project state and external engine internals were not supplied. See the boundary and Source register.
The representative workflow starts with python -m equipa --task <id> --dev-test --yes for a Git project. Launch confirmation authorizes starting work, not each subsequent edit. The orchestrator pins the default branch, creates a task branch/worktree, and delegates bounded implementation to a developer invocation. A separate tester invocation receives task context, product state and the current cycle's diff rather than the developer's entire conversation. Python code parses its verdict and chooses retry or termination. On failure, addressed messages and selected failure context can reach another developer invocation. Cost, progress, overload and cycle limits can stop work instead of repairing it. This connected workflow is implemented, not an observed successful task. See RT-RTE-dev-test and the workflow account.
The runtime map extends beyond that selected path. Goals add planner/evaluator rounds; MCP dispatch returns a process identifier before completion. Developer, tester and planner/evaluator models choose within separate contexts, while host code owns phase transitions. RLM review keeps repository material in a REPL and supplies selected subquery results, replacing successive outer-call context rather than sharing full attention over the repository. These are distinct decision and context boundaries, not interchangeable agent capabilities: RT-RTE-goal-rounds, RT-RTE-rlm-review.
From candidate acceptance to delivery
A passing tester verdict is not delivery. Count/skip checks can reject inconsistent pass reports, but no_tests is also merge eligible; audit and no-change shortcuts further qualify universal testing claims. For changed Git work, dispatch separately checks branch state, review provenance, actual diff, pinned commits and default-branch integrity before merge. A security reviewer proposes findings; executable policy can block admission. Security rejection leaves work for operator resolution rather than guaranteeing another automated repair cycle. Delivery on this path means landing on the default branch and recording a merge-reconciled outcome. Clean no-op termination is a different result. See RT-RTE-gated-merge, RT-BAP-merge-policy and the workflow annotations.
These guarantees are conditional protocols. Doc-only exemptions, operator review opt-outs and non-Git tasks prevent extending them to every task or earlier filesystem effect. A Git worktree is not an OS sandbox. The optional Claude isolation route refuses setup failure and checks exported work before import, but depends on host users, permissions, systemd/sudo, cgroups and actual network enforcement. Ollama executes handlers as the orchestrator user and is refused when isolation is enabled. When allowed, role-filtered tool advertisement does not constrain returned names at the full handler table. External lifecycle commands form another effect surface; a returned blocked mapping is not a veto unless the caller consumes it. Thus isolated launch, tool restrictions and merge admission have different owners and covered paths: RT-RTE-isolated-agent, RT-RTE-ollama-tools, RT-RTE-lifecycle-hooks.
From task experience to later context
EQUIPA wires several automatic trace-fed writes to later consumers. Task output becomes sanitized episode prose; repeated failures become template lessons, and contrasted episodes can feed model-proposed SIMBA rules. Later prompt construction automatically selects lessons and episodes using role/project eligibility, active state, Q-values, frequency/recency and lexical relevance, with optional vector/graph ranking. SQLite holds both text and access metadata; stored vectors and graph edges do not imply separate storage engines or trained model weights. The model receives shortened historical text, not vectors, full source traces or necessarily the rule's rationale. See MEM-RTE-episode-feedback, MEM-RTE-lesson-maintenance, MEM-RTE-experience-retrieval and MEM-OBJ-embedding.
This separates two consequences. Historical prose supplies advice at inference; access metadata has executable force over ranking and eligibility. Uniform task-outcome rewards change episode priorities without isolating each episode's contribution. SIMBA's before/after success-rate arithmetic can drive pruning, but uncontrolled task mix and concurrent changes prevent causal attribution. Structural and injection-pattern checks are admission controls, not factual acceptance. Requested rationale is omitted by the inspected SIMBA INSERT. The retained episode-to-reflection and episode-to-rule semantic relations remain indeterminate without actual content instances. No accepted ampliative claim or post-acceptance knowledge integration is established in those inspected routes. See MEM-BAP-experience-context, MEM-OBJ-q-value and EPI-RTE-simba-rules.
Withdrawal is consumer-specific. Inactive lessons leave generic automatic injection but remain available through explicit MCP queries. Those queries afford request-driven retrieval; the reports disagree over whether the overlapping handler's implementation status should be called afforded or wired, while agreeing that downstream client incorporation is uninspected. Neither interpretation establishes observed reliance: MEM-RTE-memory-queries, EPI-RTE-mcp-retrieval.
Recovery uses another retained chain: output checkpoints, sampled soft state and reconstructed session rows become selected continuation text. This is not restoration of a private model conversation. Sessions and hard checkpoints can coexist in history; regular history is omitted without anti_compaction_state, and consolidation can retain the first 400 words rather than the latest failures. Prompt trimming is approximate and can remove material after exposure counters were updated. Project history and initiative plans provide additional automatically selected context, with unresolved initial provenance and governance. See MEM-RTE-context-recovery, MEM-CLM-context-budget, MEM-RTE-project-context and MEM-RTE-initiative-memory.
From feedback to changes in EQUIPA itself
The strongest supported improvement contribution is an implemented route from operating feedback to EQUIPA's own later instructions, settings and retrieval priorities—not merely changes to an external product. ForgeSmith can change configuration and propose append-only prompt patches, with prompt auto-application disabled by default. GEPA validates and stores versioned candidates for conditional A/B delivery and baseline fallback. A separate autoresearch script directly deploys rewrites before benchmark-based keep/revert; it does not inherit ForgeSmith's manual/protected-section gates. ForgeSmith can return on empty findings before later maintenance phases. These distinctions matter to which changes can actually enter later work: RT-RTE-forgesmith-changes, RT-RTE-gepa-prompts, RT-RTE-autoresearch-prompts.
The bounded property findings are independent:
- Trace-fed learning mechanisms: automatic retained experience, continuation artifacts and enabled prompt/configuration changes have wired later consumers. This establishes trace-fed artifact updates, not improved capacity. Whole-task rewards, role-level before/after rates and lexical/file-list rubric scores do not establish that an individual episode, rule or prompt caused a gain: MEM-RTE-context-recovery, EPI-RTE-episode-values, EPI-RTE-simba-rules, EPI-RTE-rubric-scores.
- Theory-builder conditions 1–4: on the inspected SIMBA branch, condition 1 (a localized formulated theory) is uninspected at content-instance level; condition 2 (decisions depending on its content) is uninspected; condition 3 (claim-targeted criticism with placement of blame and changed reliance) is uninspected; condition 4 (retained criticism shaping another round) is uninspected. Repeated scoring is wired but does not settle these content requirements. Complete system assessment remains not determinable because method-writing paths and actual rule/criticism content are incomplete: EPI-RTE-simba-rules.
- Improvement through criticism of a consumed theory: not established. Neither the required content-directed criticism chain nor resulting improved capacity is supplied. This does not negate implemented feedback and revision mechanisms: EPI-RTE-simba-rules.
- Reflection: trace-fed narratives afford self-representation and feedback, but representation-mediated subsequent behavior is unobserved. The source does not establish exercised reflection merely by naming a reflection field: EPI-RTE-episode-memory.
- Autonomy: computation owns many task transitions and automatic admissions, but autonomous theory building is not established role by role; security rejection can require operator resolution. This is distinct from reliability or reflection: EPI-RTE-simba-rules, RT-RTE-gated-merge.
- Self-improvement: at the shipped-system boundary and across later task invocations, performance objectives such as task success percentage are specified independently of a particular rewrite. Evidence-responsive update determination, own-organization changes and live-capable later instruction channels are wired. This supports dispositional, improvement-directed self-change. No supplied trace establishes the fourth causal link—subsequent operation depending on the change—so exercised self-improvement is unestablished. Separate outcome evidence is also missing for successful improvement: RT-RTE-autoresearch-prompts, RT-RTE-gepa-prompts.
For a Git task requiring bounded retries and policy-controlled landing, the inspected implementation supplies a connected progression with explicit rejection boundaries. For a task requiring independently validated correctness, complete context recovery, universally enforced containment or demonstrated improvement, the same source does not establish those outcomes. This is a scenario-relative evidence distinction, not a product ranking. A trace connecting task identity, actual tests, reviewed candidate, landed commit and reconciled telemetry would establish exercised workflow. Final-prompt exposure logs and content-linked update/consumer traces would resolve different memory and self-change questions. Controlled comparisons at the claimed component level would be needed to attribute gains; actual theories, challenged claims, blame assignments and retained criticism would change the theory-builder assessment. See RT-RTE-dev-test, MEM-RTE-experience-retrieval, EPI-RTE-simba-rules and EPI-CLM-self-improvement.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No observed operation or causal experiment | SRC-1; RT-RTE-dev-test, RT-RTE-autoresearch-prompts, EPI-CLM-self-improvement | Frozen implementation and attributed design/operation claims; no live state | Successful delivery, actual activation, realized cost/reliability, exercised or beneficial self-improvement | Inspectable executions linking updates and final consumers; outcome comparisons with stated controls |
| Host and model contracts remain external | SRC-1; RT-RTE-isolated-agent, RT-CMP-claude, RT-CMP-ollama, MEM-CMP-embedding-service | Launch/refusal/import logic and configurable service callers | Deployed containment, immutable model identity, provider parameter maintenance or weight-learning conclusions | Host configuration/enforcement evidence and versioned provider/model evidence |
| Testing and review have alternate paths | SRC-1; RT-RTE-dev-test, RT-RTE-gated-merge, EPI-RTE-hook-checks | Git workflow, no-tests/shortcut/review opt-outs, conditional hook results | Every task independently tested, per-change human approval, universal correctness or security | Route-specific execution and valid test/review evidence; stronger all-path enforcement for universal claims |
| Exposure and reward attribution are incomplete | SRC-1; MEM-OBJ-q-value, MEM-RTE-experience-retrieval, EPI-RTE-simba-rules | Task-keyed tracker, pre-trim counters, uniform rewards and role-level before/after rates | Authenticated full-text exposure, individual usefulness or causal rule effect | Final delivered-context records, durable invocation-level exposure lineage and independently varied comparisons |
| Recovery and budgets are selective targets | SRC-1; MEM-RTE-context-recovery, MEM-CLM-context-budget, MEM-OBJ-persisted-output | Checkpoint/session selection, history gate, truncation, optional offload helper | Full restoration, numeric latest-checkpoint invariant, hard storage/prompt cap or established offload reload | Consumer traces and enforced size/ordering policies; wired reload caller |
| Retrieval integrations remain locally unresolved | SRC-1; MEM-RTE-experience-retrieval, MEM-OBJ-lesson-graph-edges, MEM-RTE-episode-feedback, EPI-RTE-context-reorder | SIMBA source mismatch, mixed graph ID spaces, graph-error sort behavior, omitted embedding config and unused ordering weights | Coherent semantic graph, universal bonus/vector activation, or effective audit-weight propagation/attention benefit | Typed identity and caller/consumer traces; corrected integration paths and measured retrieval/attention comparisons |
| Actual claims and content-directed criticism are unavailable | SRC-1; EPI-RTE-simba-rules, EPI-RTE-episode-memory | Generation requests, structural/safety gates and later delivery; incomplete method-writing coverage | Accepted novel knowledge, complete theory-builder verdict, criticism-driven learning or exercised reflection | Content instances, evidence-consuming acceptance, explicit criticism/blame and retained next-round dependence |
| Initial provenance and governance are unresolved | SRC-1; MEM-RTE-project-context, MEM-RTE-initiative-memory, RT-RTE-role-extension | Known later selection/transformation, instruction-driven edits and trusted-source loaders | Universal human approval/authorship, complete decision withdrawal or cross-worktree plan consistency | Content-specific admission history, governance and all relevant consumers; resolved interactive tool integration |
| MCP implementation-status conflict, carried from both verifications | SRC-1; MEM-RTE-memory-queries, EPI-RTE-mcp-retrieval; report limit, profile limit | Shared registered lesson handler: memory says afforded, epistemic says wired; both leave client incorporation uninspected | One agreed cross-report implementation status or wired downstream model reliance; inactive-row visibility and afforded pull remain supported | Owner-level reconciliation of route boundaries/statuses; deployed client-consumption evidence for stronger reliance claims |
| GHOST finding-verification branch lacks its own route analysis | SRC-1; RT-RTE-forgesmith-changes, epistemic coverage, report limit | Verifier found model scout over resolved security findings, parsed verdict/evidence, timestamp or failed-resolution disposition and follow-up task in forgesmith.py |
Complete comparisons of security-finding rechecking, decision revision and follow-up generation; omission is not absence of those operations | Declared route analysis separating check, disposition, consumers and warrant; operation/soundness evidence remains separately necessary |
| Remaining included maintenance and consumer coverage is partial | SRC-1; RT-RTE-lifecycle-hooks, RT-RTE-forgesmith-changes, EPI-RTE-rubric-scores, EPI-OBJ-training-marker; report verification | Arbitrary plugins, complete optimizers/rollback, distinct rubric producers and consumers, sweeps, training, scheduling, housekeeping/setup/migration/sync and opaque host state are not exhaustively traced | Complete improvement/measurement/training/cleanup inventory or system-wide negative; imported training metadata does not prove training | Record-supported producer/admission/consumer coverage and evaluation semantics for each unresolved alternative |
No target-source reads were performed for this synthesis; its evidence remains the handed members.