AREX-Skill synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From research request to operating context
AREX-Skill combines a published operating-skill library with DisCo, its own coding-agent runtime and instruction-construction machinery. It is not merely a collection loaded by an excluded host. This account covers SRC-1, https://github.com/VectorSpaceLab/AREX-Skill, at commit ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6, with an analysis cutoff of 2026-10-15. The evidence is inspected implementation, design instructions and attributed public results. No target execution, example trace or causal experiment was inspected. The complete repository is frozen, but loop coverage is partial; external providers, destination agents and deployed state are not captured. See the boundary and Source register.
The ordinary traced request is disco -p "task" in default Researcher mode. The CLI constructs services and a session; the model chooses actions, and the host executes granted tools and returns their results. Default tools are read, bash, edit and write. Extensions can intercept input, alter context and change tool definitions. Provider inference is mediated by ModelRuntime, but exact weight bytes, provider-internal updates and deployed isolation remain uninspected. Print mode returns assistant text or events and disposes the runtime; saved session messages can support resumption. This is an implemented execution route, not evidence of a completed research task. See RT-RTE-foreground-task and RT-CMP-model-runtime.
Before the model reads a skill, the loader constructs an eligible registry. Ordinary discovery supplies names, descriptions and locations, not every body. The model is instructed to read matching guidance; the repository router progressively narrows area, family and repository, with a no-exact-match alternative. Explicit /skill:name invocation instead makes host code read and insert the frontmatter-stripped body. Role eligibility is executable registry filtering, while Creator/Researcher task separation is prompt policy. Neither prevents otherwise authorized direct file or shell access. Guidance delivery therefore establishes an available instruction channel, not accurate routing, obedience or confinement. See RT-RTE-skill-delivery, RT-CLM-mode-contract and RT-BAP-operating-guidance.
Constructing and admitting reusable guidance
The representative workflow is source-anchored direct construction: a Creator scopes capabilities and grounds them in evidence, requests approval of an exact specification, constructs a graph, checks it and repairs affected parts, then seeks separate destination/import and overwrite approvals. Direct construction applies when no existing workflow or bounded composition is adequate and the current run can produce and verify the graph. It is not the universal construction default. A source-only request does not acquire an invented downstream task merely to obtain an outcome trial. Strict verification blockers leave an unverified candidate. Child delegation is not mandatory. See the workflow account and RT-RTE-graph-admission.
The executable generic importer supplies a narrower guarantee than that procedure. It validates roles, names, links, namespace conflicts and staged/live graph structure, rejects existing targets without overwrite, and attempts restoration after caught precommit failures. It does not authenticate a human approver or establish that semantic checks ran. Sequential renames and fallible rollback are not all-reader atomicity or power-loss recovery. Direct writes and specialized importers do not inherit these checks. Managed published-library updates have a distinct staging, digest and conflict protocol; their source commit and digest state support drift detection rather than scientific endorsement. See RT-RTE-graph-admission, RT-RTE-library-update and MEM-OBJ-library-state.
Creator delivery is an installed graph plus requested construction evidence and an external Researcher handoff, not a solved downstream task. The importer does not enforce creation of the evidence record or handoff. A later Researcher can read selected guidance without automatically receiving the Creator conversation or all supporting reasons. Factual assertions and imperative procedures coexist in graph files but have different epistemic roles. Their packaging supplies neither common warrant nor proof of later reliance. See workflow successor annotations, EPI-OBJ-skill-assertions and EPI-OBJ-skill-policy.
What verification can decide
Paper recovery provides the strongest inspected acceptance mechanism: an implemented conjunction over source-boundary, validation, target-consistency, mechanism and metric checks produces an accept/refine decision for declared recovery usability. This is evidence-consuming acceptance architecture, not merely storage. Its license remains conditional on truthful, correctly scoped evidence and compatible metric units. The comparator can substitute the first available metric, accepts positive recovery when no paper value is supplied, and has separate proxy thresholds. A distinct hard-mode artifact validator rejects declared proxy/fallback success; that predicate must not be attributed to comparator acceptance alone. Experiment-gate internals remain uninspected. See EPI-RTE-paper-checks and EPI-RTE-paper-decision.
Other checks answer different questions. Native PASS means selected-command exit success, not independent backend attestation or correctness of every instruction. QA EM/F1 compares normalized predictions with supplied reference strings, whose provenance is not authenticated here. Final reporting groups existing evidence and can write reports on failure; its positive-status expression tolerates a missing analysis decision. No observed accepted instance, produced accepted ampliative claim or post-acceptance lifecycle integration is established. See EPI-RTE-native-checks, EPI-RTE-qa-metrics and EPI-RTE-final-report.
Continuing, delegating and reusing work
The strongest wired trace-fed contribution is session summarization. Compaction turns selected conversation and tool activity, including prior summaries, into a retained checkpoint delivered with a kept tail to later model invocations. Branch summarization instead transfers selected exploration context to a navigation destination. Both automatically admit model-produced summaries after completion/cancellation checks; neither tests semantic equivalence. Tool-result serialization caps text at 2,000 characters, and retention/output budgets further limit what survives. Reasons are requested, not guaranteed preserved. Raw history remains in the event tree rather than wholly in the current window. This qualifies as an automatic trace-to-durable-context route, independently of whether future capacity improves. See MEM-OBJ-compaction, MEM-RTE-branch-summary, MEM-BAP-summary-context and profile verification.
Dynamic workflows supply programmed orchestration around model calls. Children receive fresh in-memory sessions with selected inputs rather than the parent's accumulated transcript. Their returned value crosses back; filesystem writes, diagnostics and journals can escape separately, so context separation is not OS isolation. Structured tasks enforce schema shape, including a schema-valid prose-recovery alternative, not factual truth. Persisted journals can replace fresh calls with matching cached values on resume. Matching uses a finite call-index/hash/prefix policy, not all environmental dependencies: current filesystem bytes and checkpoint defaults are not comprehensive invalidation inputs. See RT-RTE-workflow-agent and MEM-RTE-journal-replay.
Saved orchestration is different again: a selected script is retained and later executed by command or named nesting, rather than returning an old result. Command registration captures script content and does not prove refresh after replacement. Diagnostic history is a truncated display product, not the replay payload or an established model-learning input. These distinct products support local file/in-process storage, prose/symbolic forms, pull requests and automatic push selection; they do not constitute a complete memory inventory. See MEM-RTE-saved-workflow, MEM-OBJ-agent-history and profile verification.
Assessment at the supported horizon
For the scenario of preparing reusable operating guidance for later research sessions, the inspected design supplies a concrete construction protocol, structural admission, selective delivery and bounded paper-recovery acceptance. For interrupted or long-running work it supplies summary continuation, journal reuse and executable workflow reuse. Successful construction, faithful summaries, safe replay and improved downstream outcomes remain unobserved. The public verified-library characterization lacks a route-linked corpus audit here; reported benchmark gains are attributed package comparisons, not independently varied skill-component effects. See EPI-CLM-verified-skills and EPI-CLM-benchmark-gains.
Learning in the sense of improved capacity for future action is not established by the wired summary route. Paper feedback is requested to revise durable guidance, but no exercised feedback-to-revision-to-later-use chain is supplied. Theory-builder condition 1 (a formulated theory), condition 2 (content-dependent consumption), condition 3 (content-directed criticism with blame and changed reliance), and condition 4 (criticism shaping another round) each remain uninspected for particular theories at this boundary; none follows from graph packaging or procedural iteration. Whether criticism of a consumed theory improved future capacity is separately unestablished. Reflection's two causal directions and autonomous performance of all internal theory-building operations are also unestablished, not proved absent. See RT-RTE-graph-admission, RT-RTE-workflow-agent and the epistemic bounded conclusion.
For self-improvement, the assessed boundary is DisCo plus retained guidance used by later sessions, with the intended objective of more capable downstream task execution. Construction and installation can place instructions in its behavior-determining organization; paper feedback supplies a requested evidence-responsive revision path. But no inspected subsequent operation causally depends on such a revision, and no outcome evidence demonstrates improved capacity. Exercised self-improvement is therefore unestablished, separately from wired trace learning. A run linking evidence, a concrete update, admission, later consumption and changed decisions would alter that assessment; matched outcome comparisons would be needed for successful improvement. Input/output summary checks, environment-sensitive replay tests, authenticated recovery evidence and a corpus-linked verification audit would address the other specific uncertainties. No new source reads were needed for this synthesis.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Additional task-oriented feedback doctrine is omitted from the main mechanism inventory. | SRC-1; report-verification limit, RT-RTE-graph-admission, EPI-RTE-paper-decision. | PassNet feedback instructions describe graph analysis, pattern checks, GPU evaluation, parsed feedback and repair; individual script predicates, evaluator and retained-product lineage were not traced. | Complete evaluator, feedback-directed revision and admission-route comparisons across task-oriented collections; observed operation or learning from this additional doctrine. | Trace its scripts, evaluator contract, retained updates and later consumers at the frozen revision, with run evidence for operation. |
| The partial curation inventory omits explicit saved-workflow deletion and workflow-file/sidecar cleanup. | SRC-1; profile-verification limit, MEM-RTE-saved-workflow, MEM-RTE-journal-replay. | Accepted records establish those local deletion paths; the profile retains a valid managed-collection withdrawal witness but does not enumerate these operations. | Complete forgetting/withdrawal comparisons across saved orchestration and workflow state; not the supported local decay finding. | A curation inventory covering these recorded deletion mechanisms and their remaining copies/consumers. |
| Available shipped alternatives and opaque payloads remain incompletely assessed. | SRC-1; memory coverage inventory, MEM-OBJ-journal, MEM-OBJ-context-files, RT-RTE-graph-admission. | Specialized import/export recovery, individual graphs, task-oriented loops, extension replacements and use-changed instruction files are included but partial or uninspected. | Whole-system completeness or negatives about acceptance, trace learning, criticism and authority; full classification of opaque alternatives. | Producer/admission/consumer tracing and provenance for those particular alternatives. |
| No inspected execution closes delivery, activation or benefit chains. | SRC-1; Source register, RT-BAP-operating-guidance, MEM-BAP-summary-context, EPI-CLM-benchmark-gains. | Implementation and doctrine only for the local routes; benchmark statements are reported operation. | Observed reliance, learning as improved capacity, exercised self-improvement or component-level causal gain. | Provenance-linked executions and update/consumer traces; controlled comparisons matching the claimed attribution. |
| Check predicates depend on unauthenticated evidence and do not share one universal gate. | SRC-1; EPI-RTE-native-checks, EPI-RTE-qa-metrics, EPI-RTE-paper-checks, EPI-RTE-final-report. | Selected helpers inspected; experiment-validator internals and mandatory scorer-to-comparator linkage are not established. | Corpus-wide verified quality, independently measured recovery, mechanism proof, transfer, or equating report existence with acceptance. | Authenticated inputs, gate-internal inspection, explicit target/unit contracts and artifact-linked acceptance/validation traces. |
| Retention can omit reasons or reuse stale context. | SRC-1; MEM-OBJ-compaction, MEM-RTE-branch-summary, MEM-RTE-journal-replay, MEM-RTE-saved-workflow. | Summary serialization/budgets, finite replay hashes and registration-time script capture inspected. | Semantic preservation, complete environmental replay validity or automatic refresh of registered commands. | Paired original/derived context checks, explicit dependency coverage and version-aware replay/registration evidence. |
| External deployment and dependency behavior is not captured. | SRC-1; boundary, RT-CMP-model-runtime, RT-RTE-foreground-task. | Local DisCo drivers inspected; providers, destination hosts, live installed state and full dependency internals excluded or uninspected. | Exact loaded context, weight-update behavior, destination-agent compliance and deployment isolation guarantees. | Frozen deployment configuration, dependency/provider contracts and relevant execution traces. |