Qiju synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
From a session to a retained account
Qiju is a local session-record system for coding agents, not an agent orchestrator. It supplies a Python CLI and portable log, search and review skills for external hosts such as Claude Code, Codex, Kiro and Cursor. This analysis covers the whole shipped record loop, including those instructions, at SRC-1 commit a5ea9db29262ce7ec514cdc1dc5c5d017a6a3f89, with cutoff 2026-11-08. Its evidence is inspected implementation and doctrine/design. No actual store, host trace or causal experiment was supplied. Host orchestration, model internals, permissions and skill discovery remain external; source wiring is not observed handoff success (boundary, Source register).
The representative workflow is a requested checkpoint followed by a later history question. It is supported by connected capture and retrieval paths, not by joining unrelated capabilities or by an execution trace. The user supplies the checkpoint scope. The foreground host composes JSON through a temporary file and invokes log; Qiju then owns deterministic admission and persistence. The installed instructions ask for completed work, agreed next steps, rejected and undecided alternatives, artifact anchors and reasons not recoverable from files. This can preserve process knowledge that a code snapshot alone would omit, but it preserves a selected account rather than automatically acquiring the original conversation (workflow account, RT-RTE-record-capture, MEM-OBJ-body-md).
Capture normalizes metadata, repairs malformed or missing timestamps, redacts six named fields, checks schema and appends absent short/long copies under a shared home lock. A conflicting explicit project is rejected when a valid authoritative marker exists; markerless explicit projects and environment-root selection remain supported alternatives. Project selection is routing, not authenticated access control. The host/caller owns correction and retry after rejection. Staging cleanup follows successful capture and can refuse deletion with a warning without retracting the returned record ID (RT-RTE-record-capture).
These checks admit structurally valid history, not verified historical truth. Schema validation consumes types and metadata, not evidence that an action or agreement occurred. A user asking for a checkpoint does not approve each sentence, and the source label manual does not establish human authorship. Installed guidance nevertheless asks the host to privilege records as verified history. Artifact anchors and explicit decision statuses make checking possible and constrain requested composition, but no inspected factual-acceptance step closes that gap (EPI-RTE-shape-check, EPI-RTE-shape-admission, EPI-CLM-verified-history, RT-BAP-host-guidance).
Persistence and later selection
Record bodies, plans, descriptors and metadata share local JSONL and monthly Parquet files. Their effects differ: bodies supply interpreted history; next steps supply potential directions; timestamps, project names and IDs drive fixed filtering, ordering and copy-admission rules. Optional caller fields can survive capture without their meaning or provenance becoming known (MEM-OBJ-body-md, MEM-OBJ-next-steps, MEM-OBJ-search-descriptors, MEM-OBJ-address-metadata).
Retention is not semantic distillation. Explicit maintenance removes old short copies, sweeps stale staging and moves eligible long records into archives; archived history remains a normal retrieval input. Row-count and ID-set checks guard archive replacement, not every field's semantic equivalence. Sequential short/long appends are not one transaction. Retrying the same explicit ID can fill a missing tier, but existing copies are skipped without payload comparison, and an archived ID returns immediately. Short rotation does not check that a surviving copy exists after an earlier partial append. Thus the source supports preservation mechanisms, not universal losslessness or tested recovery (RT-RTE-store-maintenance, RT-RTE-record-capture, memory maintenance annotation, RT-CLM-lossless-records).
On a later question, the host requests search with scope, dates, metadata, tags or literal/regex terms. Keyword matching is deterministic OR matching, not semantic relevance ranking. The skill recommends a small candidate manifest followed by exact-ID hydration. This can keep bodies outside the model window until selected, while preserving an explicit reload address. It is not enforced: default search returns full records and show can be invoked directly. Limits and projections shape output, not a measured token budget or an indexed-performance guarantee. Qiju neither evicts accumulated host turns nor creates an isolated worker context (RT-RTE-record-retrieval).
What reaches the host matters. Summary supplies stored title and locators, not a generated narrative summary or retained reasons. Full show can return reasons; actions omits them and the exact ID. The actions renderer deduplicates strings and prints unchecked tasks without consulting later body text for completion. It therefore establishes that an item appeared in selected plans, not that it remains open. Time/project checks, widened fallback and recovery of incomplete IDs are requested host conduct, not final-answer gates. Qiju delivers CLI data; the host owns the history answer or resumed action, with no supplied acknowledgement of correct downstream use (MEM-BAP-record-history, EPI-RTE-action-rollup, workflow account).
From history to proposed method changes
The strongest supported experience-based contribution is the checkpoint route itself: the requested host computation selects session experience into a durable account, software admits it without per-content human approval, and later retrieval can supply that account as context. This supports an afforded trace-fed behavior-shaping route, including conversational session material. It does not show faithful extraction, exercised learning or improved future capacity. Generic caller and legacy inputs remain unresolved alternatives, and tool use alone does not establish tool-trace or trajectory inputs (RT-RTE-record-capture, MEM-OBJ-body-md, MEM-BAP-record-history, profile verification).
A separate retrospective procedure asks the external host to inspect recent summaries, hydrate relevant records, diagnose recurring or high-impact mistakes and propose targeted skill wording. It cautions against invented details and misplaced blame. It affords evidence-sensitive method criticism, not just a log display. Application requires an explicit user request under the instruction policy; Qiju does not enforce that permission through a CLI veto. Recommendations have no wired durable queue, actual edits and later loading are not supplied, and direct init/update replaces skills from bundled package text independently of retrospective findings. Backups do not prove that local lessons remain active (RT-RTE-skill-review, RT-RTE-skill-install, EPI-CLM-review-improvement).
The runtime separates the relevant properties rather than assigning a grade:
- Theory-builder condition 1, localized content: afforded at whole-guidance/proposed-wording granularity; actual reviewed theory content is uninspected.
- Condition 2, consumption: claimed; requested decisions depend on guidance's failings, but actual dependence is not observed.
- Condition 3, content-directed criticism: afforded by requests to identify unclear, missing or counterproductive instructions and distinguish alternative causes; actual claim-specific criticism and changed reliance are not supplied.
- Condition 4, iteration: uninspected; no retained criticism shown to shape a next round. The conditions therefore do not establish a functioning theory builder.
- Improved future capacity from criticism of consumed guidance: uninspected, not established by retained history, recommendations or edits alone.
- Reflection: afforded for the host-plus-guidance pathway, in which operation records can represent mistakes and guide changes affecting later work; actual bidirectional use is unobserved.
- Autonomous theory building: uninspected; full computational performance of the internal operations is not evidenced, while application expressly requires user authorization.
- Self-improvement: the host-plus-installed-guidance pathway affords improvement-directed proposals toward preventing repeated mistakes and unclear instructions across sessions. Evidence shaping an actual admitted update, that update entering live authority, and later operation causally depending on it remain unestablished. Neither exercised self-improvement nor successful improvement is demonstrated.
These are the separate findings of RT-RTE-skill-review; the afforded checkpoint trace-fed route does not upgrade any of them.
Scenario-relative assessment
For a deliberately recorded checkpoint and a later question with useful scope or search terms, Qiju implements addressable local history with optional delayed body loading. For recovering an unrecorded decision, establishing that a plan is still pending, or treating a retrieved narrative as factually accepted, those mechanisms do not supply the missing evidence. For preventing recurring mistakes, it supplies a structured review-and-proposal procedure, not demonstrated beneficial adaptation (RT-RTE-record-retrieval, EPI-RTE-action-rollup, EPI-CLM-verified-history, RT-RTE-skill-review).
Linked capture/retrieval traces would change the assessment of exercised handoff and rationale receipt. An evidence-consuming historical check would change the warrant assessment. A retained critique, authorized edit and later behavior trace would address iteration and live use; objective-based outcome comparisons would be needed separately for benefit, with attribution no finer than the intervention compared. Transactional persistence or verified surviving-copy controls would change the preservation assessment. None of those changes follows from source inspection alone.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No supplied operational or causal evidence; external host behavior remains uninspected | SRC-1; RT-CMP-host-model, RT-RTE-record-capture, RT-RTE-record-retrieval, MEM-BAP-record-history | CLI, shipped host instructions and documented external dependencies | Actual host compliance, cross-host continuity, recall quality, performance and causal benefit | Linked host invocations, retained records, returned fields and downstream actions; appropriately bounded comparisons |
| Structural admission does not authenticate history; claim inventory is not multilingual-complete | SRC-1; EPI-RTE-shape-check, EPI-RTE-shape-admission, EPI-CLM-verified-history, EPI-OBJ-record-metadata | Capture/schema and inspected English claims/skill bodies; other documentation not fully inventoried in the epistemic account | Fact-verified narratives, authenticated event times, accepted new historical knowledge or exhaustive public-claim enumeration | Record-to-artifact checks and recorded acceptance criteria/results; completion of the relevant documentation inventory |
| Output shaping can omit reasons or later status | SRC-1; RT-RTE-record-retrieval, MEM-OBJ-search-descriptors, EPI-RTE-action-rollup | Tier merge, filtering, renderers and installed search procedure | Complete recall, actual rationale receipt, verified pending tasks or host-context compaction | Hydration/consumer traces and explicit completion/retraction reconciliation; evidence of external context management if claimed |
| Durability and preservation are conditional | SRC-1; RT-RTE-record-capture, RT-RTE-store-maintenance, RT-CLM-lossless-records | Sequential append, first-ID precedence, rotation and row/ID archive checks | Cross-file atomicity, payload-equal retries, field-equivalent archival and universal no-loss recovery | Field roundtrip checks and fault/retry evidence under named filesystem and DuckDB contracts; surviving-copy guarantees |
| Privacy controls cover particular paths and fields, not all copies | SRC-1; RT-RTE-record-capture, RT-RTE-privacy-repair, MEM-OBJ-staging-record, EPI-RTE-store-withdrawal | Six-field capture redaction; literal scrub of current short and home long/archive; purge of home long/archive | Secret exclusion from optional metadata, IDs/extras, staging, imports, other workspaces, backups or external host copies; universal withdrawal | Broader admission/traversal controls and copy inventory with measured coverage; external-host evidence |
| Imported and opaque content retains unresolved provenance and meaning | SRC-1; MEM-OBJ-address-metadata, MEM-OBJ-redaction-metadata, RT-RTE-store-migration | Known import/rebrand route and retained extra fields; detailed normalization helpers not fully assessed | Complete form, lineage, authority, original admission or trace-source classification; semantic preservation of legacy history | Authenticated input instances and consumer paths, helper-level record support and field/meaning comparisons; rebranding can erase historical distinctions |
| Retrospective changes and their later consumption are not established | SRC-1; RT-RTE-skill-review, RT-OBJ-qiju-skills, RT-RTE-skill-install | Requested review, user-authorized application policy and independent bundled replacement | Completed criticism/iteration, retained rationale use, exercised reflection, autonomous theory building, improved capacity or exercised/successful self-improvement; complete revised-guidance read-back direction/signal | Actual critique and admitted edit, preserved reasons, later skill loading and dependent behavior, plus objective-based improvement comparisons |
| Optional scheduled maintenance is documented but not inventoried as an implemented installation | SRC-1; RT-RTE-store-maintenance, report verification limit | Explicit maintenance implementation; overview doctrine mentions optional macOS launchd job, but frozen tree lacks installer/job implementation | Scheduler wiring, unattended coverage or automatic-upkeep reliability; not evidence scheduling never existed | Frozen installer/job implementation and invocation evidence linking it to maintenance |