CQ synthesis

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

Bounded synthesis

From task experience to shared advice

CQ is a knowledge-sharing system for agents, not their enclosing runtime. This analysis covers its shipped prompts, integrations, clients, stores, remote admission and review, maintenance and source-level evaluators at commit 5765494a3840e5f5a7223c42101f41e85cb9d898, with an applicability cutoff of 2026-10-14. SRC-1 is the frozen repository. Evidence is inspected implementation and doctrine/design; no target execution, observed session, test result or causal experiment was supplied. External hosts, model services and the live deployment are outside that evidence boundary. See boundary and evidence and Source register.

The operational division is clear: the host agent chooses what to ask, reasons about candidates and decides how to act; CQ selects, transports and persists advice. Its five MCP operations are query, propose, confirm, flag and status. CLI and SDK routes expose related operations; the shipped Pi adapter maps use to shell commands. SQLite is the default local store, with alternative or injected stores and an optional remote service. CQ returns knowledge, not a completed application task, and does not mediate the host's other tools or model calls. RT-RTE-knowledge-query, RT-RTE-host-installation, RT-CMP-host-model.

A retained lesson consists of natural-language summary, detail and action, accompanied by symbolic tags, identifiers and evidence metadata. Detail can retain an explanation, but source references, raw task traces and checking rationale are not required. Caller content enters with unknown source warrant. Serialization, structural validation and later retrieval do not independently validate its claims. Arbitrary extension JSON has no established uniform meaning or application consumer. MEM-OBJ-insight, MEM-OBJ-selection-metadata, MEM-OBJ-extensions.

The skill requests querying before unfamiliar work and proposing stabilized discoveries during work. Retrospective reflection is a backstop: the foreground agent is asked to summarize selected session experience and draft shareable candidates. A conflicting skill subsection describes server extraction, but the dedicated command, MCP registration and Python stub support agent-led extraction, not that server endpoint. Candidate production and its semantic relation to the session remain unobserved. RT-RTE-session-reflection, EPI-CLM-reflection-server.

Permission, retention and remote eligibility

The representative workflow is interactive reflection with a configured remote store. Its requested progression is summary → candidate selection → safety screening → user approve/edit/skip → query for close matches → proposal or feedback → inline results and final counts/IDs. There is no delegated sub-agent: a later reviewer is another admission principal. The near-match query is limited retrieval, not a novelty certificate. Individual proposal errors are to be reported while the batch continues, not rolled back transactionally. These handoffs are supported by instructions and implementation, not an executed workflow. See workflow account and RT-RTE-session-reflection.

Three decisions must remain separate. Interactive user approval selects content under the instructed procedure. Software validates fields and credentials before persistence. Remote review subsequently controls eligibility for ordinary remote retrieval. The requested vulnerabilities, impact, bias and edge-case screen can prompt criticism and sanitization, but its result is not required by proposal handlers. Headless rules permit clean or appropriately caveated proposals without a user; direct APIs also do not attest approval. For presented hard findings, the interactive procedure defaults to sanitized content but lets the user explicitly choose the original. Consequently neither human approval nor removal of hazardous content is a universal software invariant. RT-CLM-proposal-approval, RT-RTE-knowledge-proposal, EPI-CLM-guardrails-layer, workflow account.

Successful remote proposals receive server identities and pending review state. A browser affords inspection and approve/reject/skip; it shows the retained lesson and selected metrics, not the originating conversation or safety-screen provenance. JWT authentication establishes a caller, not human presence or independent review; a shipped seed script approves through the same APIs. The pending check precedes an unconditional status write, so it does not establish atomic single-winner review. Ordinary server queries enforce approved status, including on the semantic selector. That gate does not establish factual acceptance or PRIVATE-tier confidentiality: public query filters status rather than ownership or tier. Rejection retains the payload for review readers while excluding ordinary retrieval; skip changes browser-local selection only. RT-RTE-remote-review, RT-CLM-approved-query, workflow account.

Local retention follows a different path. Go proposal failures involving transport, server errors or 401/403 can fall back locally, where content is immediately queryable without remote review. Python supports caught per-call failures after destination resolution, but exhausted discovery failures become an uncaught DiscoveryError before insertion. Thus outage retention is route-dependent. Later Go drain, explicitly invoked or started in the background at MCP startup, re-proposes LOCAL units and deletes successfully posted local copies. The remote successor has a new identity and awaits review. Transfer omits old evidence, flags and supersession links; Python also omits extensions. This is best-effort transfer, not faithful replication, exactly-once publication or uninterrupted availability. RT-RTE-knowledge-proposal, RT-RTE-local-drain, MEM-OBJ-flags, MEM-OBJ-extensions.

Delivery, checking and future selection

Queries are principally requested memory reads. SQLite gathers domain-or-full-text candidates and ranks relevance times confidence; ordinary server queries rank approved domain matches. Go merges local first, deduplicates by ID with local winning, then truncates without a global rerank. A small limit can therefore exclude every remote candidate. Defaults and clamps bound item counts, not the host's token window. MCP query drops nonfatal warning/source metadata that SDK or CLI consumers can receive, weakening the host's ability to diagnose incomplete remote retrieval. Retrieved action text remains advice; CQ does not execute it. RT-RTE-knowledge-query, RT-BAP-knowledge-guidance.

Optional semantic retrieval uses retained numerical vectors, first limiting approved domain matches by distance, then ranking by (0.8 × relevance + 0.2 × (1 − normalized distance)) × confidence. Distance still affects final ordering. Provider identity, compatibility and quality are unresolved; similarity adds no truth warrant. Wrapped provider errors can bypass the intended SQL fallback, and embedding failure can block writes or feedback updates. MEM-OBJ-knowledge-vectors, RT-CMP-token-embedding, epistemic selection ledger.

The skill requests verification against current documentation, tool output or experience before reliance and feedback. The API consumes a caller's assertion rather than requiring that evidence. Confirmation raises confidence; supported flags lower it and retain an objection, without rewriting the insight or revoking review approval. Initial confirmation counts are admission defaults, not independent successful use. A sole matching flagged unit can remain deliverable even at zero confidence. Go-local duplicate flags can retain a reference; server and Python Client duplicate requests lack the required reference and fail before updating. Feedback therefore supplies wired future ranking influence, not calibrated correctness or guaranteed protection. RT-RTE-knowledge-feedback, MEM-OBJ-selection-metadata, MEM-OBJ-flags, EPI-OBJ-evidence, EPI-CLM-flag-protection.

For a task needing reusable starting-point advice, the inspected loop supports persistence, later selection and requested host checking. For a task needing accepted factual knowledge, reliable novelty detection, confidentiality or guaranteed withdrawal of wrong advice, those mechanisms alone are insufficient. Permission to submit, approved status and confidence do not establish evidence-consuming truth acceptance against a named criterion and intended scope. The inspected retention and indexing routes consequently do not establish lifecycle integration after such acceptance. The declared automatic PII, factual and staleness guardrail suite remains doctrine on the selected paths, not an implemented evaluator suite established by this analysis. RT-CLM-approved-query, EPI-CLM-guardrails-layer, epistemic bounded conclusion.

Experience-responsive capacity and its unresolved effects

The strongest supported experience-responsive contributions are persistent feedback that can change later advice ranking and the automatic Cursor failure-event → clipped retained record → later stop reminder chain. Under the memory comparison's trace-learning definition, the latter qualifies as wired trace-fed context production without needing a diagnosis or demonstrated improvement. It selects the latest conversation-keyed failure, bounds input/error text, emits it and clears it; session start sweeps old files. This is not automatic lesson admission. Discovery caches separately route HTTP calls using retained endpoint metadata. Neither these mechanisms nor reflection's requested summary implement host turn eviction, compaction checkpoints or isolated worker frames. RT-RTE-knowledge-feedback, RT-OBJ-failure-record, RT-RTE-failure-reminder, MEM-RTE-discovery-resolution, profile verification.

The independent property conclusions remain bounded:

  • Theory-builder conditions 1–4: localized candidate theory, content-dependent consumption, content-directed criticism and iteration are each uninspected for actual reflection reasoning. Requested extraction or stored advice does not establish any of them. RT-RTE-session-reflection.
  • Learning from criticism of consumed theory: future ranking capacity is wired, but no comparison establishes improved future action capacity, or attributes it to criticism of what a consumed theory says. Score arithmetic does not evaluate a prose objection's argument. EPI-OBJ-evidence, epistemic bounded conclusion.
  • Reflection: agent-led examination of session experience is afforded; exercised reflection remains uninspected. The source-native command name is not proof of the full reflective property. RT-RTE-session-reflection.
  • Autonomy: deterministic storage and score updates are wired; whole-loop autonomy is not established. External hosts own reasoning and action, and the selected interactive path requires per-content user decisions. RT-CMP-host-model, RT-RTE-session-reflection.
  • Self-improvement: over subsequent advice-consuming tasks at this boundary, no exercised improvement-directed self-change is established. Retained rank changes can reach future selection, but no trace shows subsequent behavior causally depending on them, and no independently assessed outcome establishes improvement. Static prompt installation is not experience-driven self-patching. RT-RTE-knowledge-feedback, RT-RTE-host-installation, epistemic bounded conclusion.

A linked session/candidate/review/later-use trace would change the assessment of operation and consumption. Recorded challenged claims, blame placement, changed reliance or revision, and reuse in another round would bear separately on theory building. An independently specified objective and controlled later-outcome comparison would bear on improvement and attribution. Implemented approval attestations, tier-aware access predicates, atomic review transitions, evidence-requiring acceptance, withdrawal rules or identity/history-preserving transfer would change the relevant architectural assessment; they would still require their own execution evidence. These resolving conditions follow the limits of the cited proposal, review, query, drain and feedback records, not a product ranking or adoption recommendation.

Limitations

limitation affected SRC IDs and record citations inspected boundary conclusion prevented evidence that would resolve it
Undeclared independent prose-criticism part; carries the unresolved reconciliation conflict and both verification limits SRC-1; MEM-OBJ-flags, EPI-OBJ-evidence, report verification limits, profile verification limits Retained mixed feedback containers, Go-local serialized delivery and separate arithmetic/prose consumers Complete part-addressed criticism inventory, independent admission/consumption/lifecycle comparisons, or uniform form/authority for either container. This does not imply absent criticism or arithmetic evaluation of its argument. Separate analyst declaration of the prose part with its admission, consumer, authority and lifecycle evidence.
No observed operation or causal comparison SRC-1; RT-RTE-session-reflection, RT-RTE-remote-review, RT-RTE-knowledge-feedback, RT-BAP-knowledge-guidance Source and instructions only; external host execution and live service excluded Successful extraction, compliance, factual checking, later reliance, exercised theory building or self-improvement, and causal benefit Linked execution artifacts and scoped independent outcome comparisons, including the retained result and later consumer.
Requested checks and publication decisions lack supplied factual acceptance evidence SRC-1; RT-CLM-proposal-approval, RT-RTE-remote-review, EPI-CLM-guardrails-layer, EPI-CLM-reflection-server Interactive/headless prompts, direct admission, browser/API review and selected retrieval paths Universal human approval, effective automated content guardrails, accepted produced claims or lifecycle integration; host response to conflicting reflection instructions Recorded criterion, intended use, scope and consumed evidence for acceptance; corresponding enforced routes and reconciled shipped instructions where claimed.
Provider and initialized-state provenance unresolved SRC-1; RT-CMP-token-embedding, MEM-OBJ-knowledge-vectors, MEM-OBJ-node-info, MEM-OBJ-extensions Opaque provider, pre-existing vectors/stores/cache files and arbitrary extension payloads Complete lineage, provider compatibility/quality, authenticated initial state or extension meaning and authority Versioned provider and initialization evidence; concrete payloads and their actual consumers.
Partial alternative-path and operational-envelope coverage SRC-1; RT-RTE-knowledge-query, MEM-RTE-store-replacement, RT-RTE-host-installation, MEM-RTE-maintenance-migration Inspected built-in selections, exposed Store capability, selected adapters and maintenance scripts; custom internals and host processing uninspected Exhaustive cross-client/store/adapter equivalence, complete memory inventories, host compaction or isolation, migration reference/vector consistency, live-writer safety and deployed reliability Scoped alternative implementation coverage, host contracts and traces, and maintenance/deployment evidence.
Delivery and transfer do not preserve all context or uninterrupted availability SRC-1; RT-RTE-knowledge-query, RT-RTE-local-drain, MEM-OBJ-selection-metadata, MEM-OBJ-flags Local-first bounded merge, MCP warning omission and re-proposal/delete transfer Complete novelty search, faithful identity/history replication, exactly-once transfer or continuous access to the transferred lesson End-to-end selection/provenance evidence and an identity-preserving, idempotent transfer/availability contract with supporting implementation.

Source-read log: none. This synthesis uses the supplied members and their retained passages; no ambiguity required direct frozen-source inspection.