AREX-Skill synthesis

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

Bounded synthesis

AREX-Skill combines a library of operating skills with DisCo, the enclosing application that constructs model sessions, supplies guidance and executes tools. This account concerns SRC-1, https://github.com/VectorSpaceLab/AREX-Skill, at commit ac3fe1afa80fb9a09775ecfb2b6cc3ba850a2db6, with an analysis cutoff of 2026-10-19. Its boundary is complete artifact, partial loop: the Git artifact is frozen, but not every shipped producer, evaluator, maintenance route and later consumer has been traced. The evidence is inspected implementation, instructions and attributed product reports. No target execution, example trace or causal experiment was inspected. External inference providers, dependency internals and separately installed consuming hosts are outside the inspected implementation boundary. See the boundary and Source register.

From request to operating context

A DisCo CLI request constructs services and a session, restores selected history, resolves a provider/model and enters interactive, print or RPC consumption. AgentSession assembles context and handles interception; external agent-core owns the inner model/tool schedule. The model proposes actions, registry dispatch selects executors, and local shell or custom operations realize effects. Print mode returns final text or JSON events, not a correctness certificate. Provider/model identifiers do not pin exact weights; provider-side parameter changes remain uninspected. These distinctions are supported by RT-RTE-session-prompt and RT-CMP-model-runtime.

Installed skills enter this context in stages. Role filtering precedes name-collision resolution; visible names, descriptions and locations enter a catalog. A body reaches the model through explicit invocation or a requested file read. Router guidance asks the model to compare likely capability areas and load a small useful set, rather than supplying an inspected embedding-retrieval mechanism. Hiding a description does not prohibit explicit invocation of a registered skill. Role and visibility filters do not confine filesystem access. Installed prose has instruction force through interpretation, not enforcement of every procedure: RT-RTE-skill-delivery, RT-BAP-skill-guidance and its memory annotation.

The same distinction bounds execution controls. Name-based allow/exclude filtering constrains the constructed model-tool registry, conditional on agent-core dispatching through it. It is not an OS sandbox or universal execution prohibition. The shipped RPC bash route separately reaches shell execution or extension-supplied operations without consulting that allowlist. Extension-side and provider-native operations are not exhaustively traced. See RT-RTE-tool-grants and RT-RTE-user-bash.

From draft to later Researcher

The representative workflow begins with an integrated repository-skill draft, not an invented reconstruction of its creation. The caller supplies source evidence, coverage and backend plans, prepared-environment information and import policy. Verification instructions require assertion-backed usability checks, content review, safe native cases, repair and focused rechecking. A fresh-agent test without research notes is prescribed when available; it is not evidence that a separate worker actually ran. Missing required accelerator evidence blocks the prior auto-import authorization: synthetic checks cannot substitute for it. The agent must obtain a new informed decision, verify the backend or narrow scope. These are model-interpreted policies, with local assertion/native-case references rather than a universal answer oracle. See WF-RTE-repo-verification.

Executable admission is narrower. The importer validates portable files, metadata, links, license-helper results and routing-handoff consistency, including digest and evidence-range checks. It does not judge whether cited passages support the operating claims or replay native verification. A manual-unrouted alternative permits unclassified admission with required runtime metadata. Overwrite flags do not authenticate the separate human approval requested by the workflow. Caught mutation failures trigger restoration attempts; failed restoration preserves recovery artifacts. This is best-effort multi-path recovery, not demonstrated crash atomicity: EPI-RTE-import-check and RT-RTE-repo-import.

Official-library installation/update is another route, with conflict/force, digest, staged-swap and no-op branches. Neither specialized route governs arbitrary direct file or shell writes. The retained graph separates interpreted instructions from executable scripts and routing metadata. Rich classification rationale and verification reports are not identical to the minimal runtime projection; later consumers do not automatically receive all admission reasons. Successful import makes guidance available to a new Researcher session. It does not show that session occurred, used the guidance or succeeded. See RT-RTE-library-update, MEM-OBJ-repo-skills and MEM-OBJ-routing-metadata.

Checks with bounded acceptance

The paper-recovery route supplies a substantive positive beyond storage and packaging: implemented code consumes evidence predicates and records an accept/refine decision for generated-skill usability on a declared full or proxy recovery target. Checks concern source/target consistency, validation, mechanism flags, artifacts and metrics. Their applicability depends partly on supplied optional inputs. Default metric tolerance is 0.05; proxy minimum is 0.8. Paper values over 1 are divided by 100 without equivalent recovered-value normalization, and a missing requested metric can fall back to the first available one. Meaningful comparison therefore depends on target identity and encoding. Logs and success labels are not independent execution attestation. A separate final hard-mode gate rejects proxy success even when the comparison accepts it. See EPI-RTE-recovery-check and EPI-RTE-recovery-decision.

For QA, exact match and token F1 are conditional computations against supplied answers, not authenticated scientific truth. Punctuation removal can collapse distinct numerical answers such as 1.0 and 10. Neither a scalar match nor recovery acceptance warrants every skill proposition, its explanatory mechanism or transfer. Refinement instructions do ask agents to use failures to revise modules and guidance, then repeat validation; that revision loop remains doctrine/design rather than inspected operation. See EPI-RTE-qa-normalization, EPI-RTE-qa-scoring and EPI-RTE-refine-guidance.

Retention, continuation and replay

The strongest implemented trace-fed contribution is automatic continuation-summary production and later model delivery. Default compaction transforms conversation and tool history into retained prose; repeated compaction can update an earlier summary. A recent verbatim tail remains, and original entries remain in the session tree outside active context. Optional branch summarization can carry selected abandoned-path information into a destination branch. Extensions may cancel or substitute these summaries. Ordinary model conversion delivers summary text, not arbitrary details; requested decision rationale does not guarantee fidelity. This supports an implemented trace-learning write route in the memory-comparison sense, not demonstrated improvement in future capacity: EPI-RTE-summary-generation, RT-RTE-context-compaction, MEM-OBJ-compaction-summary and MEM-RTE-branch-navigation.

Session JSONL stores an append-only tree, but the selected branch and latest compaction determine later context. Memory-only sessions and deferred initial flushing limit disk-durability claims. Custom state, display visibility and model-facing messages are separate parts; a hidden display is not necessarily context exclusion. See MEM-RTE-session-retention and MEM-OBJ-custom-entry.

Dynamic workflows use independent in-memory worker sessions and return selected values. Background completion pushes a report projection into the foreground; resumed interpretation instead requests a matching-prefix journal value and can skip live worker execution. Hash/index matching does not certify environmental freshness, and a checkpoint can use a headless default rather than human approval. Diagnostic truncation has a display consumer, not an established worker-context restoration route. Independent conversations do not imply filesystem isolation. See MEM-RTE-workflow-context, MEM-RTE-workflow-resume and MEM-OBJ-agent-history.

What the progression establishes

For continuation under context pressure, source code establishes summary retention and later delivery; preservation and benefit remain unobserved. For installing research guidance, it establishes structural admission and a later instruction channel, with richer verification requested by policy. For paper recovery, it establishes fallible evidence-consuming disposition within declared artifact/metric conditions. None licenses collection-wide executable correctness. The advertised verified-skill collection and benchmark gains remain attributed claims; no inspected experiment establishes an aggregate causal effect here, still less a routing, refinement or individual-skill effect. See EPI-CLM-verified-skills, EPI-CLM-accepted-graph and EPI-CLM-benchmark-gains.

These positives leave independent properties unresolved. For exercised construction, theory-builder condition 1 (a localized whole theory unit), condition 2 (a content-dependent decision), condition 3 (content-directed criticism with blame and changed reliance/revision), and condition 4 (that result shaping another round) are each uninspected. A generated theory and its complete dependent criticism/use chain were not individuated. Improved future capacity from criticizing a consumed theory is likewise uninspected. Reflection's two causal directions and autonomous performance of all internal theory-building operations are independently unestablished, not inferred absent. The relevant limits are stated in EPI-RTE-refine-guidance and the epistemic conclusion.

For self-improvement, the assessed pathway is Creator/verification feedback → revised retained guidance → later Researcher operation, with better task usability as the objective and construction-to-later-session use as the horizon. Policy requests evidence-responsive revision, and implementation supplies durable admission and live delivery. This supports a dormant, improvement-directed self-change pathway. It does not establish an exercised evidence-to-update link or subsequent behavior causally depending on the update; successful improvement needs separate outcome evidence. Trace-fed summaries do not close those attribution gaps merely by existing. A provenance-linked trace of criticism, exact edits, admission and a later dependent decision would change the exercised-path assessment; a suitably controlled outcome comparison would address benefit at its actual treatment granularity. See EPI-RTE-refine-guidance, RT-BAP-skill-guidance and EPI-CLM-benchmark-gains.

Limitations

limitation	affected SRC IDs and record citations	inspected boundary	conclusion prevented	evidence that would resolve it
Saved-workflow machinery admission remains partial; carried from report verification.	SRC-1; MEM-RTE-workflow-context, MEM-RTE-workflow-resume; verification limit	Save/registration, independent sessions and result replay are supported; full replacement, validation, initial provenance and nested-consumer admission are not audited.	Complete production-machinery admission comparison or delegation isolation across all workflow paths.	Records tracing saved-script authorship, replacement gates, validation and nested execution/resource inputs through actual consumers.
Specialized evaluators, refresh/extend and cross-host export/withdrawal remain partial; carried from report verification.	SRC-1; boundary, memory integration issues, epistemic coverage; verification limit	Selected repository verification, paper recovery and session/workflow memory routes, not every available shipped path.	Collection-wide correctness, complete evaluator coverage, universal withdrawal or system-wide absence of adaptation/knowledge production.	Source-grounded producer/check/admission/later-consumer records for the remaining paths, including exported-payload removal and task-specific outcomes.
Operation and improvement are not observed.	SRC-1; RT-RTE-session-prompt, WF-RTE-repo-verification, EPI-RTE-refine-guidance, EPI-CLM-benchmark-gains	Source implementation and policies; README results are reported operation only.	Reliable compliance, exercised theory-building/self-improvement, improved future capacity or causal skill/component benefit.	Provenance-linked run evidence through revision and later reliance, plus an outcome comparison with variation appropriate to the claimed effect.
Semantic support and input authenticity remain conditional.	SRC-1; EPI-RTE-import-check, EPI-RTE-recovery-check, EPI-RTE-qa-normalization	Structural import checks, declared artifacts/metrics and lexical QA comparisons.	Blanket scientific correctness, authenticated reproduction, mechanism fidelity or unrestricted semantic equivalence.	Individuated claim/source checks, authenticated execution provenance, verified reference answers and target/metric encoding checks over the intended domain.
Retained reasons, opaque alternatives and context fidelity are unresolved locally.	SRC-1; MEM-OBJ-routing-metadata, MEM-OBJ-custom-entry, MEM-OBJ-compaction-summary, MEM-OBJ-workflow-journal	Runtime projections, readable summary text and cache envelopes; not all index evidence, extension payloads or full worker-value consumers.	Complete memory-form, provenance, authority or trace-source inventories; guaranteed delivery of reasons or faithful summaries.	Records tracing each opaque branch and post-transformation consumer, index evidence projection and summary/source correspondence.
Execution and provider guarantees stop at local enforcement points.	SRC-1; RT-RTE-tool-grants, RT-RTE-user-bash, RT-CMP-model-runtime, RT-RTE-repo-import	Named registry filtering, supported direct shell path, request adapter and caught-error installation recovery.	Universal no-execution, deployed isolation, exact weight pinning, provider mutability facts or crash-atomic installation.	Deployment/executor contracts and evidence, complete alternate-path tracing, provider version/mutability evidence and crash/recovery observations.