Sentry Skills synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Sentry Skills is an instruction and helper package for engineering and writing assistance, not an enclosing agent runtime. This analysis covers the repository artifact at commit d18b7aa8ba878354e5c348310230e652f7690f9c, with an applicability cutoff of 2026-10-14. Its operational boundary is complete artifact, partial loop: external hosts select instructions, call models and execute tools. The evidence is inspected implementation and doctrine/design from SRC-1, not supplied execution traces or causal experiments. No target code, evaluator or service was executed. The boundary and Source register state the inclusions and external dependencies.
From discovery to action
Installation exposes canonical skills and subagent prompts through plugin/discovery surfaces. A host delivers Markdown instructions to a foreground model or subagent; frontmatter has a separate host-parsed, symbolic consumption path. The model interprets workflow rules and chooses actions. Python helpers perform fixed fetching, transformation, validation or polling. Tool permissions, actual relevance matching, model versions and isolation belong to external hosts. The subagent alias opus does not pin a model revision or establish whether parameters change. See RT-OBJ-skill-md, RT-BAP-skill-consumption and RT-CMP-host-model.
Context control follows the same division. Skill authoring prescribes loading only the references needed for the current step, rather than all maintenance material. This is foreground selection, not an implemented token allocator or eviction mechanism. Document reader testing specifies a fresh reader receiving only the draft and question, with a manual fresh-chat alternative. It affords withholding author context, but host frame destruction, compaction and actual isolation are unobserved. See RT-RTE-skill-revision, its memory annotation, and MEM-RTE-reader-testing.
A repair cycle and its actual stopping condition
The representative iterate-pr workflow connects request, criticism, repair and delivery. With authenticated gh, uv and target-repository context, the foreground agent fetches CI and review feedback, diagnoses causes, edits code, verifies locally, commits and pushes, then examines new results. Helpers receive PR identities and return selected observations; no model-worker delegation is required. The foreground retains repair and recovery ownership. High/medium feedback is actionable, low-priority feedback requires user choice, and repeated failure after two attempts, unclear feedback or infrastructure problems require consultation. Local failure is a prescribed bar to pushing, not a code-enforced interception. See RT-RTE-pr-iteration and the workflow account.
The observations are narrower than a complete review transcript. Feedback retrieval is bounded and projects inline threads to their first comment, omitting self-comments and very short bodies. CI failure snippets are bounded; workflow-run matching does not enforce current-head SHA equality. Some fetch failures become empty lists. Instructions therefore require full failed-log inspection and cause identification before editing. Direct gh fallback is allowed, so helper classifications are not mandatory control points. These limits constrain what a clean fetched result establishes, independently of whether the repair policy is sensible. See RT-RTE-pr-iteration.
The monitor polls deterministically; it does not repair code. Pending non-human checks delay termination, then failures and human gates determine markers. Terminal failures can exit zero. ALL_CHECKS_PASSED means no received pending/fail bucket under its classification, not that every check explicitly passed. Full workflow success additionally requires a feedback fetch ten seconds after actionable CI passes and handling the low-priority choice. New actionable feedback reopens repair. Delivery is the pushed repair plus a success, stop or ask-user report, not proof of mergeability. See RT-RTE-check-monitor and its workflow annotation. Optional replies publish caller-authored claims and classify API acknowledgements; they neither verify those claims nor establish later agent receipt of the replies. See MEM-RTE-thread-reply.
Checks license different conclusions
The package does not supply one universal evaluator. Its structural validator checks skill format and recognized local references; a size warning is advisory, and passing does not establish semantic quality or security. Security scanning produces mechanical pattern leads, while model-directed context review filters intent and false positives before an advisory installation recommendation. Warden configuration is not inspected enforcement. See RT-RTE-skill-revision, EPI-RTE-pattern-scan and EPI-RTE-security-disposition.
Likewise, triage approval authorizes only a displayed archive batch; it is a prompt policy, not an argument-level mutation wrapper or proof that classification is correct. Reader testing checks selected comprehension, not document truth. Endpoint documentation requests source/live-response comparison, whose warrant cannot exceed the tested domain. Code simplification requests behavior preservation on both skill and subagent paths without establishing semantic equivalence. These are distinct operational and epistemic scopes, supported by RT-RTE-issue-triage, MEM-RTE-reader-testing, EPI-RTE-endpoint-check and RT-CLM-preserve-functionality.
Evidence can revise the guidance itself
The strongest supported improvement contribution is a prescribed experience-to-guidance route. Skill-writer separates working/holdout examples, interpreted findings, promoted instructions, provenance decisions and a maintenance contract. Model authors can choose deltas without per-change human approval; moving holdouts into the working set requires an explicit user decision. Later maintenance can read retained reasons, while ordinary task consumers normally receive the resulting rules rather than every justification. See RT-RTE-skill-revision, MEM-OBJ-working-set, MEM-OBJ-findings-log, MEM-OBJ-sources, MEM-OBJ-spec and EPI-RTE-example-retention.
This supports a claimed trace-fed durable behavioral write, not automatic trace ingestion or demonstrated learning. Actual example origins, populated promotion records and original trace classes are unresolved. Retained experience is an input to revision; revised instructions have a different consumer in later task execution. Configured AXIS generation and model scoring do not install changes. Maintainer adoption requires completed working-set runs, no critical holdout regression, target improvement and the stated validation conditions. Actual reports, decisions and deployed read-back are not supplied. See RT-RTE-maintainer-eval and the profile verification.
Other retained outputs serve different purposes. Generated project guidance is intended to direct later agents. A SRED Work Summary is explicitly requested by a later organizer as evidence; a saved draft is requested for reader testing. Neither document handoff by itself establishes package self-revision. Host guidance loading affords or declares push, while requested maintenance/document reads are pull; the actual automatic host selector remains unknown. See MEM-RTE-project-guidance, MEM-RTE-sred-handoff and MEM-RTE-reader-testing.
For skill-writer, theory-builder condition 1, localized method content, is afforded; condition 2, decisions depending on it, is claimed; condition 3, content-directed criticism with blame and changed reliance, is uninspected; and condition 4, retained criticism shaping another round, is claimed, not observed. Learning as improved capacity for future action remains uninspected. Reflection is afforded through revision of its own method texts, while exercised reflective theory building and autonomous theory building remain uninspected. These findings are independent. The package affords dispositional self-improvement toward high-value input coverage with reduced wasted runtime tokens across later installed invocations. Evidence shaping an actual update, its entry into live host authority and subsequent operation causally depending on it are not established; successful improvement is separately unestablished. See RT-RTE-skill-revision.
For the selected repair scenario, the supported mechanism is a model-owned cycle with explicit verification and escalation policies plus bounded observation helpers. For later guidance revision, it is an evidence-retention and adoption procedure, not an exercised improvement result. A trace linking criticism, revision, admission, installed loading and later changed behavior would strengthen the latter assessment. Controlled comparisons would be needed to attribute benefit to particular rules or helpers. Host enforcement evidence and current-head-complete service observations would change the strength of the former assessment. Additional shipped revision mechanisms already prevent treating the selected skill-writer account as exhaustive, as the report verification explains.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Retained-revision and evaluator inventories are incomplete; report verification identifies prompt-optimizer evidence packs, concrete criticisms, candidate comparisons, hypothesis/edit/result logs and holdouts, plus a pr-writer adoption gate, without separate analyst records. | SRC-1; RT-RTE-skill-revision, RT-RTE-maintainer-eval, EPI-RTE-example-retention; verification limit | Selected analyst routes plus verifier inspection of prompt-optimizer and pr-writer evaluation material at the same pin | Exhaustive comparisons of retained criticism, revision routes or acceptance criteria; system-wide negatives about those mechanisms | Separate supported records covering these mechanisms and remaining consequential alternatives |
| Host execution, permissions, model resolution and context lifecycle are external and unobserved. | SRC-1; RT-BAP-skill-consumption, RT-CMP-host-model, MEM-RTE-reader-testing | Package prompts, manifests, configuration and context instructions | Actual activation, compliance, deployed enforcement, isolation or parameter mutability | Frozen host implementations/configuration and inspectable invocation traces |
| PR observations are selected and can be incomplete or mistaken for success. | SRC-1; RT-RTE-pr-iteration, RT-RTE-check-monitor, MEM-RTE-thread-reply | Inspected fetch, poll and reply helpers and PR instructions | Complete current-head feedback coverage, merge readiness, truthful reply content or later receipt of replies | Head-linked service traces, complete retrieval evidence, checked repairs and actual consumer inputs |
| No exercised revision-to-later-behavior or causal improvement chain is supplied. | SRC-1; RT-RTE-skill-revision, RT-RTE-maintainer-eval, MEM-OBJ-working-set, MEM-OBJ-findings-log | Prescribed evidence/replay/adoption paths; opaque AXIS results and baseline | Exercised content-directed criticism, improved future capacity, autonomous theory building or successful self-improvement | Provenanced examples, claim/blame/revision traces, adoption and live read-back records, later behavior and appropriately controlled outcome comparisons |
| Retained payloads, origins and alternative consumers remain partly unresolved. | SRC-1; RT-RTE-maintainer-eval, MEM-OBJ-document-draft, MEM-RTE-thread-reply, RT-RTE-skill-revision; profile verification | File witnesses alongside host-artifact, AXIS, reply and unread domain-leaf alternatives | Complete memory substrate/form/lineage/authority inventories, controlled push-selector or original trace-source classification | Recorded operative payloads, admission provenance, selector inputs/rules and later consumer delivery for each alternative |
| Check outcomes have local domains; operation and effectiveness were not observed. | SRC-1; EPI-RTE-pattern-scan, EPI-RTE-security-disposition, RT-RTE-code-refinement, MEM-RTE-reader-testing, EPI-RTE-endpoint-check, MEM-RTE-sred-handoff | Selected scanner, refinement, reader, endpoint and SRED instructions/helpers; SRED classification/template leaves not assessed | Detector completeness, guaranteed preservation, factual certification, exhaustive service retrieval or adequate coverage of every endpoint/SRED case | Scoped execution results with independent references, relevant leaf coverage and comparisons matching the claimed generalization |