Exocomp synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Request and delegation
Exocomp is an enclosing runtime for model-directed software work. It supplies a planner, role-specific contractor processes, project tools and operator interfaces. This account covers the shipped repository at commit b36f7ba1f3e2eee25bbcd2711db7d2d5d4824251, with cutoff 2026-10-14. SRC-1 supplies inspected implementation and declared design. No target executable, test, provider or service was run. Historical session artifacts were available but were not used to establish operation at this revision. The whole-system functional boundary includes shipped callers and retained-state consumers, not an exhaustive audit of every helper. External model/provider internals, host isolation and generated frontend distribution are not established by that boundary. See boundary and evidence and its Source register. This synthesis uses accepted records without additional source reads.
An ordinary terminal request reaches a planner Session. The Session sends accumulated messages and role-filtered tool schemas to an OpenAI-compatible provider. It takes the first response choice, runs parsed tool calls sequentially, appends results and requests another completion. A response without tool calls returns control to the client; it is not a project-completion certificate. Code owns progression, while the model proposes actions. Model defaults and provider aliases do not pin exact weights or immutable endpoints. See RT-RTE-planner-chat and RT-CMP-planner-model.
Hire launches a distinct agent-mode process with a role, task and directory arguments, not the planner's full conversation. The child uses its role-default model; the parent's bookkeeping model does not prove which model the child invokes. Await waits for a registered child's termination and returns an extracted Quit report. The parent retains child messages for inspection, but the planner does not automatically receive the whole transcript. Separate conversations therefore limit context transfer without isolating shared filesystem effects. See RT-RTE-contractor-work, RT-CMP-coder-model and MEM-OBJ-agent-messages.
Specification, implementation and checking
The representative workflow is an operator's Go feature request: clarify and plan, hire architects, confirm the plan, hire coders, hire testers, delegate repairs and ask for refinement. These steps are connected by shipped instructions, not a coded phase machine or an observed successful lifecycle. One central artifact handoff is obstructed at this revision. Toolset converts DefineFunc to Definefunc, while Requirements uses exact membership and mixed-case definition dispatch. Neither spelling reaches the ordinary installed definition branch; the other Define* methods have the same problem. Direct public definition methods remain available, and existing specifications can be read, but successful Hire/Await cannot establish that architects produced the instructed new requirement artifacts. See WF-CLM-project-phases and MEM-RTE-requirement-update.
Coders can write project files and execute allowed programs. These effects occur before planner report acceptance. File writes check lexical paths, formatting and I/O, not behavioral conformance or passing tests. Program admission checks executable names and selected arguments; permitted Go code runs as an ordinary host process. Its internal file and network operations do not pass through the file-tool checks. Thus a planner without Files.Write can still possess code-execution capability, and a sandbox working directory is not an OS isolation guarantee. Lexical file checks also do not resolve symlinks. See RT-RTE-file-admission, RT-RTE-program-execution and RT-CLM-role-separation.
The installed tool map and each tool's HasMethod predicate supply real dispatch controls. They are not universally literal grant membership: an installed Humans tool admits Answer specially. Terminal commands and HTTP tool requests can invoke the same dispatcher without model selection or the chat busy-state gate. The inspected web handler adds no caller-authentication gate, while deployed network exposure remains unknown. See RT-OBJ-role-grants and RT-RTE-client-tools.
Human clarification has implemented advisory delivery. Ordinary qualified Answer calls store text, release a matching waiter and append generic tool messages; released Ask/Choose calls separately append their results and permit inference to continue. Direct Answer does not itself start inference or backup. Browser local validation does not prove asynchronous delivery, and its Waiting flag can refuse submission. Only a bare-name embedding call reaches the distinct out-of-bounds Session branch. Neither arrival nor the tool's human label authenticates the respondent or certifies a proposition, and requested plan confirmation is not an enforced write gate. See EPI-RTE-human-answers.
Testing capacity is substantive: tester instructions require go test, and permitted program execution can obtain test-specific evidence. Its warrant would remain bounded by the tests and their adequacy. No supplied result or enforced test-success gate establishes correct software here. Signoff is narrower still: an AST symbol-presence check permits implemented-status assignment, not acceptance that the requested behavior works. Bug Fix accepts closure without a repair check and suppresses the report from tool lists; its description remains in JSON. Changelog Remove appends an event rather than verifying or deleting a product change. See EPI-CLM-correct-results, EPI-RTE-symbol-check, EPI-RTE-signoff-disposition, EPI-RTE-bug-closure and MEM-OBJ-changelog-entry.
Delivery, recovery and later context
Contractor delivery uses report presence rather than task verification. The child selects its exit status using a phrase in its Quit message, while the parent ignores the Wait result and marks it finished when a nonempty report exists. A status-1 child can therefore supply an accepted report. No-report termination rejects Await delivery; cancellation, timeout and rejection do not roll back files. The planner chooses reassignment or repair. Final operator-facing prose is a third delivery boundary, distinct from child emission and Await return, and supplies no automatic correctness certificate. See RT-RTE-contractor-work and the workflow account.
Recovery writes session and agent JSON. Terminal and web startup can restore accumulated planner context and recorded contractors, not running child processes. Later inference supplies planner messages wholesale before provider adaptation; requested Await, project List/Search and operator GET expose selected retained material. Ignored backup returns prevent a chat-success-to-persistence guarantee. Debug snapshots are diagnostics with no established automatic future-inference reader. See MEM-RTE-session-recovery, MEM-RTE-operator-display and MEM-OBJ-debug-snapshot.
Requirement recovery changes the meaning of the available record. The writer retains only unimplemented specifications. Subsequent reads reconstruct implemented declarations from source with empty Behavior and prefer them over saved same-key specifications. Original intended behavior can disappear even before explicit Signoff; the next implemented view is not an archived acceptance decision. Bug visibility and richer operator JSON follow different retirement rules. These mechanisms control later information supply without validating its truth. See EPI-RTE-requirement-refresh, MEM-OBJ-requirement-behavior and MEM-RTE-project-records.
Customization likewise has distinct stages. Project role YAML replaces a same-process template, but parent registration does not prove the effective template of a fresh child initialized before its tool discovery. Skills.List refreshes the catalog, not loaded objects or injected messages. Qualified Load/Unload changes loaded state and emits notices; it does not insert or withdraw the body. Separate Session helpers or bare-name API use afford system-message insertion/removal. Body-text matching can leave older differing bodies, and no withdrawal rolls back prior effects. Embedded role skill grants are absent within the recorded YAML search boundary; granting a method alone would not repair the qualified-name delivery mismatch. See MEM-RTE-role-discovery, RT-RTE-skill-loading and RT-ABS-skill-grants.
Learning contribution and scenario-relative assessment
The strongest supported trace-fed contribution is the afforded Inquire route: recorded child user/assistant prose feeds automatic condensation; a returned summary can become a retained tool message, survive backup and supply later caller context. This qualifies as an automatic trace-fed durable context artifact, not demonstrated improved capacity. The original transcript remains. Projection omits tool/system messages and tool-call structure, so the summarizer lacks complete execution evidence; its semantic preservation is unchecked. No embedded role grants Inquire, so this is not a demonstrated default planner route. See RT-RTE-agent-inquire, EPI-RTE-inquire-projection and the accepted profile verification.
The unresolved question is whether consuming such retained material improves future action. No before/after or causal comparison establishes that outcome. Similarly, the records do not establish a formulated theory route meeting condition 1 (localized theory content), condition 2 (decisions depending on that content), condition 3 (content-directed criticism and changed reliance), or condition 4 (retained criticism shaping iteration). These are separately unestablished, not inferred absences. Criticism of a consumed theory improving future capacity is unestablished independently of summary production. Exercised reflection, autonomous theory-building and evidence-responsive self-improvement are also unestablished; model-selected task execution is not evidence for those properties. Explicit customization and executable installation do not close an evidence-to-own-organization-to-later-operation causal chain. See RT-RTE-agent-inquire, RT-RTE-skill-loading, RT-RTE-build-install and the epistemic bounded conclusion. This does not assert that successful testing, acceptance or these properties are impossible in every project workflow.
For the selected Go-feature scenario, implemented delegation, immediate editing and resumable context coexist with an obstructed new-specification handoff and weak completion criteria. Evidence of a revision-linked end-to-end run would establish which routes actually operated; explicit behavioral criteria, adequate test results and their consequential acceptance consumer would support stronger task-completion claims. Repairing Define* dispatch would change the specification-capacity assessment. Repairing qualified skill body delivery, with explicit replacement/withdrawal semantics, would change instruction-adoption guarantees. Demonstrating retained evidence causally selecting a change to Exocomp's own organization and governing subsequent operation would change the self-improvement assessment; successful external product work alone would not. The report verification accepts these bounded implementation findings, not efficacy or universal reliability.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Operation and benefit are unobserved at the pin; historical artifacts lack established revision provenance. | SRC-1; RT-RTE-planner-chat, RT-RTE-contractor-work, WF-CLM-project-phases | Shipped source and prompts, without target execution; historical artifacts not used as current-run witnesses | Successful lifecycle execution, reliability rates, compliance and causal benefit | Revision-linked traces with task criteria, artifacts, test results and appropriately controlled outcome comparisons |
| External model state, provider retention and arbitrary program-created memory remain opaque. | SRC-1; RT-CMP-planner-model, RT-CMP-coder-model, RT-CMP-summarizer-model, RT-RTE-program-execution | Exocomp request/execution integration, not dependency internals | Exact model identity/immutability and complete retained-memory or learning enumeration | Immutable model identities and scoped provider/program retention, update and consumer evidence |
| Host isolation and distribution success are not established. | SRC-1; RT-RTE-program-execution, RT-RTE-file-admission, RT-RTE-client-tools, MEM-RTE-operator-display; boundary | Lexical helpers, host-process launch, authored web routes; generated assets and GUI dependency not frozen | Universal confinement, authenticated deployed access and runnable-browser distribution claims | Deployment controls, process/isolation evidence and reproducible frontend/build artifacts at the reviewed boundary |
| Initial project-content provenance and per-content human control are unresolved. | SRC-1; MEM-RTE-project-records, MEM-RTE-role-discovery, RT-RTE-skill-loading, EPI-RTE-human-answers | Software import/admission paths and generic callers | Authenticated source warrant or manual approval classification from startup, caller identity or labels | Content-linked authorship, respondent authentication and explicit supply/edit/approval decisions |
| Transcript restoration does not establish loaded-skill or pending-wait restoration, nor guaranteed backup integrity. | SRC-1; MEM-RTE-session-recovery, RT-RTE-skill-loading, EPI-RTE-human-answers | JSON backup/restore and newly constructed runtime tools | Complete cross-restart extension/coordination recovery and guaranteed persistence or answer delivery | State-specific reload wiring and revision-linked backup/restart and concurrent-delivery evidence |
| Summary and import fidelity, detailed display behavior and acceptance coverage remain bounded. | SRC-1; EPI-RTE-inquire-projection, EPI-RTE-website-fetch, MEM-RTE-operator-display, EPI-RTE-symbol-check | Selected projection/conversion/display and AST paths; not all parser, rendering, test or project acceptance branches | Complete extraction/summary/declaration fidelity or a system-complete negative about epistemic acceptance | Source-to-output comparisons, remaining material helper inspection and actual evidence-consuming acceptance instances with named scope |
| The partial curation inventory omits separately classified role-template replacement and refreshed same-name loaded-skill replacement. | SRC-1; MEM-RTE-role-discovery, RT-RTE-skill-loading; profile limit | Recorded template-map overwrite and loaded-object adoption, independently of body insertion | Complete comparisons of extension/template revision mechanisms; supported existence findings remain valid | Separate classifications of both replacement mechanisms against their consumers; catalog refresh alone is insufficient |
| Some profile form and authority units group distinct parts or consumers. | SRC-1; RT-RTE-agent-inquire, EPI-RTE-inquire-projection, MEM-OBJ-changelog-entry, MEM-RTE-project-records, MEM-OBJ-debug-snapshot; profile limit | Inquire input/output, changelog duplicate/date consumers and diagnostic envelope/prose units | Part-by-part or consumer-by-consumer counts, or attributing every grouped force/form to every part | Split classification units for input learning versus output context, duplicate admission versus ordering, and envelope versus payload; individual positive findings remain supported |