Stigmergy synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Capture becomes evidence before it becomes knowledge
Stigmergy is a team-knowledge system that retains captures, compiles a linked wiki and supplies scoped material to later readers and models. This account covers its shipped acquisition, filing, maintenance, retrieval, answering and release paths at Git commit 0ca8bb1963ad9cf43eaf2c18329998023a9caa4f, with an analysis cutoff of 2026-11-08. It is code-grounded in SRC-1, not an observation of a deployed tenant. No target execution, authenticated production trace or causal experiment supports the account. Provider internals and external MCP hosts' own action selection and memory lie outside the boundary. See the boundary and evidence and Source register.
The operational division is between model proposals and software-controlled effects. External people or agents submit through MCP, a local bridge or Slack. Code binds identity and audience, retains original artifacts and queues work. A single writer processes tenant-bound operations; a model proposes wiki mutations rather than directly exercising Git credentials or shell access. Later questions invoke a separate read-oriented model loop. This is not an enclosing general-purpose agent runtime (RT-RTE-capture-filing, RT-RTE-scoped-answer).
Audience selection has several doors. Explicit audiences undergo membership/existence checks; omitted audiences need a valid default and may invoke a classifier under additional conditions. The classifier excludes public from its choices, but a public default can still be selected outside that branch. Classification therefore does not establish a universal prohibition on automatic public filing. Reader scoping similarly depends on correct principal/tenant binding and database-role isolation; it is not proof that every process or administrative path lacks broader access (RT-RTE-layer-routing, runtime controls).
Original bytes and readable evidence are distinct retained parts. Extraction produces text and provenance for a source page in the tenant's Git repository. That text need not be a verbatim rendering of the original medium: OCR and format parsers intervene. Retaining hashes, originals and extraction metadata supports lineage, not source truth or transcription fidelity (MEM-OBJ-artifact, MEM-OBJ-source, EPI-RTE-evidence-extraction).
A bounded proposal passes route-specific gates
The representative workflow follows one capture from receipt through filing and publication. A receipt means queued work, not a completed wiki change. At processing time, code rechecks the author's current publishing authority, serializes the write and creates a temporary worktree and layer view. The planner receives the source, provenance, packaged librarian instruction and selected existing content; other restricted layers stay outside its view (workflow account, RT-RTE-capture-filing).
Existing context is selected by entity/project names and lexical overlap, with bounded expansion of cited source evidence. Selection supplies retained content automatically before the planner requests particular pages; explicit reader tools elsewhere provide requested retrieval. Filing context is trimmed by dropping whole objects, not by maintaining a durable conversation summary. Omitted material remains stored, but the filing model has no tool to reload it during that call. The nominal 768-KiB limit does not guarantee successful rendering: compact-size trimming is followed by a separate indented-size rejection check, and excess evidence may be removed while one candidate remains (memory annotation, report-verification limit).
The model returns a typed FilingPlan. Writer checks enforce contracts, ACL relations, source references and capture-update preservation of cited blocks and local attributions. A bounded correction can replace a rejected proposal. Persistently lossy updates can be omitted individually; an invalid whole plan can restore mutable pages and retain only its source, still subject to fallback gates. Thus even a landed capture need not contain accepted wiki synthesis (RT-OBJ-filing-plan, EPI-RTE-filing-check, EPI-RTE-filing-retention).
These checks admit a structured knowledge product, not the truth of every sentence. The instruction requests source-backed synthesis, but without concrete input/proposal pairs the semantic relation of actual generated claims remains unresolved. Publication alone establishes no produced, accepted ampliative claim. There are narrower content-specific acceptance routes: alias classification admits eligible names at probability 0.75 for later lookup, while explicit entity merge requires a source assertion matching restricted identity grammar before consolidating pages. Neither independently establishes real-world identity; the merge check depends on imported source truth (EPI-RTE-filing-transform, MEM-RTE-alias-recording, EPI-RTE-merge-check, EPI-RTE-entity-merge).
Publication checks the lease, commits with an operation marker, advances the branch and records change evidence before queue finalization. Git and database completion are not atomic. On interruption, a later attempt can find the operation commit and reconstruct bookkeeping without generating another plan; this is forward reconciliation, not automatic reversal of a bad edit. Indexed availability follows asynchronously through webhook application or rebuild. Landed, searchable and answered are different milestones (workflow recovery, MEM-OBJ-knowledge-changes, RT-RTE-index-convergence).
Later consumers receive selected evidence, not an endorsement
The retained wiki combines interpreted prose with code-consumed controls. Page bodies and visible contradiction explanations supply evidence; metadata supplies identity, visibility and matching inputs. Fragment text and dense vectors support retrieval, not model-weight training. Automatic initial answer search and filing context selection coexist with explicit search/page/fragment requests. Page heads are excerpted, with fragment reload available to readers; a warning retained in the full page can be absent from its initial excerpt (MEM-OBJ-knowledge-body, MEM-OBJ-page-metadata, MEM-OBJ-contradiction-display, MEM-RTE-scoped-reading, profile verification).
For each question, AnswerService creates fresh state, searches in code, then lets the model use scoped read tools. Figures are checked against recorded tool evidence. Citation checks instead require a surfaced path and quote occurrence in a newly fetched scoped head/title plus fragments selected by recorded window numbers. Passing does not prove that every quotation previously appeared in delivered tool text or in the identical earlier snapshot. The report verifier confirmed this narrower domain; the runtime's surfaced-content wording must not be expanded into a prior-delivery guarantee (RT-RTE-scoped-answer, EPI-RTE-answer-check, report verification).
Occurrence also does not establish semantic support. A nonnumeric answer can contradict an occurring citation without violating these checks. Any untraced figure or failed verdict suppresses the draft, but exactly one citation-only defect can ship as partial. Budget recovery attempts an evidence-only completion under a stricter verified-only condition; otherwise code shapes a refusal. Raw search/read/list/describe products bypass this answer verifier entirely (EPI-RTE-answer-check, EPI-RTE-answer-disposition).
Maintenance changes future inputs; improvement remains a separate question
The strongest supported adaptive contribution is durable corpus and lookup change: filing, alias updates and accepted repairs can alter material supplied to future invocations. Scheduled garden repairs defined violations; explicit recompile regenerates derived pages source by source. Capture-update conflict preservation and named master-authorized resolution do not establish an identical invariant for garden's separate repair-body replacement. Ordinary source protection is likewise not global immutability: authorized deletion and history erasure deliberately remove material. Withdrawal sweeps dependent units and links; it does not decide that the withdrawn proposition was false (RT-RTE-maintenance-admission, EPI-RTE-contradiction-disposition, EPI-RTE-evidence-withdrawal, epistemic conclusion).
This architecture can carry evidence and reasons forward, but their exposure differs: contradiction explanations can return to filing, while generic mutation reasons retained in manifests have no established automatic model reader. Mixed capture origins also leave qualifying trace-fed learning unresolved; raw queue/audit/evaluation retention is not itself such a learning route. No inspected run establishes improved future capacity, or that criticism of a consumed theory caused such improvement (memory write side, RT-BAP-corpus-data, profile verification).
These findings do not establish or refute theory-builder conditions 1–4 as a complete system classification: the accepted runtime supplies no separate condition-by-condition assessment. Retained prose, delivery and correction feedback alone cannot substitute for evidence of theory-dependent decisions, content-directed criticism and iteration. Reflection and autonomous theory building remain separately unestablished; automatic scheduling and admission do not settle them. For self-improvement, corpus outputs can affect later organization and cannot be excluded merely as products, but exercised evidence-responsive self-change is not demonstrated here. The separate release pathway can replace code and packaged skill; its evidence-to-update determination and subsequent causal dependence are also unestablished (epistemic conclusion, RT-RTE-product-admission).
Release evaluation is more than ordinary filing lint. Golden QA reports finite-case outcomes without automatically revising production; filing parity compares configurations and external blind editorial judgments before a guarded staging deployment. An explicit operator bypass exists. No inspected terminal release packet establishes that this revision passed, was installed through that gate or improved later operation; configuration comparisons do not isolate component effects (EPI-RTE-qa-evaluation, EPI-RTE-release-evaluation, EPI-RTE-release-admission).
For a capture that must remain inspectable despite failed synthesis, source retention, bounded rejection and operation-marker recovery are material capacities. For a question requiring proposition-level factual warrant, the occurrence gate is insufficient. These are scenario-relative conclusions, not a product ranking. Concrete source-to-plan-to-commit traces would resolve semantic transformation and delivery questions; authenticated release packets and later-use traces would establish exercise; objective-linked comparisons with independent component variation would be needed for stronger benefit attribution. Semantic claim-support checks would change the answer gate's license only within their demonstrated domain, not retroactively validate the source corpus.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Context-budget disagreement retained by reconciliation and report verification | SRC-1; RT-RTE-capture-filing, memory annotation, report-verification limit | Filing context selection/rendering in src/stigmergy/knowledge/context.py |
Guaranteed rendering within the nominal cap, or all-candidates-before-any-evidence eviction | Aligned size accounting and verified eviction policy; until then the documented rejection and final-candidate exception remain |
| Quote-domain disagreement prevents a stronger joint guarantee; verifier confirms refetch rather than delivered-text interpretation | SRC-1; RT-RTE-scoped-answer, EPI-RTE-answer-check, verification disposition | Surfaced-path membership, scoped head refetch and recorded fragment-number checks | Prior delivery or immutable snapshot identity for every passing quote; general semantic support | Verification against recorded delivered text/snapshots would establish the former; a separately supported semantic evaluator would be needed for the latter |
| Wiring is not observed activation or benefit | SRC-1; RT-RTE-capture-filing, RT-RTE-scoped-answer, EPI-RTE-release-admission | Frozen implementation and doctrine; no authenticated target run or causal experiment | Production reliability, exercised self-improvement, improved future capacity or component-level benefit | Provenance-linked execution and later-consumer traces, objective-linked outcome evidence and appropriately controlled comparisons |
| Original-source fidelity, authorship and trace origin remain unresolved | SRC-1; MEM-OBJ-artifact, EPI-RTE-evidence-extraction, MEM-RTE-alias-recording | Extraction dispatch and selected OCR/text handling; not all media parser internals or actual input/output pairs | Universal transcription fidelity, complete lineage or a positive/negative trace-learning classification | Format-specific evidence comparisons and original-input provenance linked through automatic writes to later consumers |
| Included operational and lifecycle details remain incompletely inspected | SRC-1; memory coverage, memory integration issues, MEM-RTE-workspace-replay | Selected entity/tombstone, listing/UI, reset, audit/report and roster/session surfaces | Complete memory inventory, exhaustive read/write comparisons or complete reset/erasure safety | Operative-part and consumer tracing of the named remaining branches, with corresponding records |
| Preservation and withdrawal guarantees are route-specific | SRC-1; EPI-RTE-contradiction-disposition, RT-RTE-maintenance-admission, EPI-RTE-evidence-withdrawal | Capture preservation, separate garden replacement and managed deletion/purge | Universal maintenance preservation or deletion from external copies, backups and prior contexts; no actual garden loss is shown | Cross-route preservation checks and traces; an independently bounded inventory and verification of external retention/deletion |
| Scope and model guarantees depend on external contracts | SRC-1; boundary, RT-RTE-layer-routing, RT-CMP-librarian-model, RT-CMP-embedding-model | Shipped principal binding, scoped services and remote named model clients | Deployment-wide confidentiality, immutable provider weights or stable endpoint behavior | Audited deployed role/identity configuration and provider version/privacy evidence within a separately declared boundary |
| Evaluation and release evidence is fixture/configuration bounded and unobserved in operation | SRC-1; EPI-RTE-qa-evaluation, EPI-RTE-release-evaluation, EPI-RTE-release-admission | Selected scorer/parity/deploy implementation; no terminal real packet or full helper audit | Universal semantic quality, reviewer authentication, proof this release passed or isolated component effect | Authenticated complete review/replay artifacts, helper inspection, installed-use evidence and contrasts matching the attribution sought |