Flopwire synthesis

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

Bounded synthesis

Flopwire preserves agent conversation evidence and makes it available to later sessions through retrieval and messaging. This account covers the whole shipped system at commit 057a4b87603cc878e81d9eb49298ddc4eb9eeb23, with an analysis cutoff of 2026-10-20. Its evidence is inspected implementation and attributed design in SRC-1, the frozen https://github.com/flopwire/flopwire repository. It includes the device daemon, optional team server, CLI/MCP tools, harness adapters, instructions, maintenance and evaluation drivers. External harnesses own model scheduling, permissions and interpretation. No execution trace or causal experiment is registered, and no target was run during this analysis. These findings establish implementation capacity, not successful operation or benefit. See the boundary and evidence and Source register.

The operational progression starts with an external harness recording a conversation. Flopwire's device daemon discovers changed sources, applies collection policy, and resumes parsing or reparses rewritten content. Accepted messages and progress cursors enter a local SQLite index. Admission is automatic under parser and policy rules; a human-authored prompt is not evidence of human approval of its capture. Extracted, compressed message text can survive deletion of the native transcript. Symbolic source identities, generations, cursors and lexical indexes govern subsequent selection. Unknown policy pauses collection; deny rules can purge local rows, and a failed parser-version dry run preserves earlier rows. These are structural controls over retention, rather than checks of what the conversation asserts. RT-RTE-transcript-index, MEM-OBJ-message-text, MEM-OBJ-index-metadata.

An enrolled device can also upload redacted source bytes to a shared archive, where asynchronous parsing supplies shared retrieval. The server uses Postgres and object storage, distinct from the local extracted-text store. Upload acknowledgement does not establish completed parsing or query visibility. Native-agent parsed-byte bounds have conditions: exports and empty policy take a different branch, and explicit recovery imports have separate snapshot and receipt logic. The inspected routes therefore support neither a universal native-watermark bound nor a demonstrated recovery bypass. Recovery receipts can remain unqualified after a successful upload. RT-RTE-shared-upload, MEM-OBJ-archive, MEM-RTE-import-recovery.

Flopwire automatically derives durable navigation from retained traces. Digests select short first-prompt and last-answer snippets and bounded activity fields; later folds consume separate bookkeeping. Commit projections retain command/result references and observed hashes for change lookup. This is a supported contribution beyond raw logging: trace-fed writes produce artifacts that later tools and code consumers use. Retained receipt sightings and historical sharing provenance also affect later delivery and restriction decisions. These routes support wired automatic trace learning in the comparison's write-and-later-consumer sense. They do not establish learning as improved capacity for future action. Original host-checkpoint production and alternative evaluator-fed revision routes remain unresolved. MEM-RTE-derive-navigation, MEM-OBJ-delivery-metadata, MEM-BAP-history-restriction, profile verification.

A later agent or user requests search, grep, session discovery, change lookup or an addressed read. Fixed lexical selection and filters return bounded evidence, leaving historical interpretation to the caller. The inspected local BM25/FTS and server lexical paths require no embedding model. Backend selection refuses invalid enrollment configuration and does not silently switch to local history after a shared-backend error; explicit local selection remains available. Caller self-exclusion is a narrower discovery policy: named-session, include-self, undetected-caller, read/raw and change paths prevent interpreting it as access denial. RT-RTE-history-retrieval, runtime scope controls.

What reaches the model matters. Digests and outlines are navigation, not complete accounts; outlines omit intervening assistant answers and tool results. Display cleaning and clipping can remove decisive context. Addressed reads can recover retained passages, while raw local reads require verifiable native bytes and do not substitute the independent server archive. A typical read focuses on 4000 bytes, with up to 24000; ordinary rendered MCP output also has a 24000-byte budget. These limits govern delivery, not the host model's whole context window. Captured Codex compaction can add a readable summary and novel replacement messages to later retrieval; opaque encrypted alternatives and the host's original selection and reuse remain unresolved. Static instruction renewal after host context changes does not reconstruct discarded conversation. EPI-RTE-text-render, memory retrieval annotation, MEM-RTE-capture-compaction.

The strongest supported check on historical claims is narrower than general truth recovery. Commit extraction pairs captured calls and results, rejects selected unsafe patterns, and preserves printed commit observations separately from whole-call failure. It conditionally derives observations under its encoded transcript interpretation. It does not independently authenticate Git state. Change lookup returns evidence addresses and exact/prefix distinctions; a prefix match establishes neither unique identity nor sole authorship. Arbitrary imported conversation claims have unknown source warrant. Neither ranking, digest retention nor projection storage is evidence-consuming acceptance of an ampliative historical explanation. The shipped recovery skill asks the external reader to distinguish proposals from approvals, inspect contradictions and recover decisive clipped text; software does not enforce the interpretation. EPI-RTE-commit-check, EPI-RTE-commit-store, EPI-RTE-change-lookup, RT-BAP-recovery-guidance, epistemic conclusion.

Coordination adds automatic supply of retained messages. At eligible hooks, recipient identity, pending state, leases and budgets select envelopes for insertion into context. Hook output is limited to five messages and 9000 encoded bytes; full stored bodies can be reread through inbox while retained. Body markup is escaped, but permission and sender-handling rules remain instructions interpreted by the model. Repeated unconfirmed instruction attempts can be treated as delivered, so bookkeeping is not proof of receipt. Opencode and cloud adapters use different insertion and confirmation mechanisms; a cloud busy-state check before push is not an atomic no-wake guarantee. Flopwire controls its message crossing, while the external host owns subagent contexts and return values. RT-RTE-hook-delivery, RT-RTE-adapter-delivery, RT-BAP-message-guidance, memory hook annotation.

The representative workflow connects evidence recovery to a local own-user request and correlated reply. The initiating agent establishes the historical session, checks presence when contact is needed, and sends a bounded question with optional evidence references. A recipient receives it at an eligible event, performs work under its host's controls, and can send a reply linked to the request. The requester inspects the answer and decides whether it suffices. Software can reject invalid sends, including replies to a closed done parent; it does not verify substantive completion. Unconfirmed leases permit redelivery, with model guidance against repeating work, rather than exactly-once execution. This is an instructed and wired multi-call workflow, with no observed exchange. Reading retained server threads instead supplies historical evidence and explicitly grants no new authorization to act or reply. WF-RTE-request-reply, MEM-RTE-thread-history.

Withdrawal has several scopes. Content masking refreshes local text and derived views, with durable sidecars protecting later rewrites. Native files and already delivered contexts remain separate. Collection denial, reversible server hiding, permanent deletion tombstones and asynchronous object purge have different effects. Historical provenance can restrict sharing even after source rows disappear. Independent backups have their own retention; an explicit pending-redaction override can preserve sensitive archive bytes in a backup. None of these mechanisms alone guarantees global erasure. MEM-RTE-redact-content, MEM-RTE-withdraw-history, MEM-RTE-restore-state, MEM-BAP-history-restriction.

For recovery of prior decisions, this architecture supplies persistent source passages, navigation and provenance checks, while leaving claim acceptance to the reader. For coordination, it supplies addressed handoffs and reply correlation, while leaving scheduling and work acceptance to the agents and hosts. Qualification drivers provide synthetic expected-answer checks and parser-fixture comparisons; they were not executed here and do not license general task reliability. Explicit binary upgrades validate checksums and runnable versions, without an established automatic evaluation-to-production promotion route or evidence of improvement. RT-RTE-workflow-trial, EPI-RTE-parser-parity, RT-RTE-binary-upgrade.

The retained records do not establish a theory-builder assessment for each of conditions 1–4: localized method text alone does not establish a formulated theory; delivery alone does not establish content-dependent decisions; content-directed criticism and retained iteration are unestablished in the inspected routes. No criticism-to-improved-capacity finding is supported. Reflection's two causal directions and autonomous theory building are separately unestablished, rather than inferred absent from opaque model processing. Exercised self-improvement is also unestablished: the trial and upgrade routes do not close an evidence-responsive self-change chain through subsequent operation causally dependent on that change. These conclusions do not negate the supported trace-fed update capacity. Route-linked execution evidence, comparison of future capacity, and an operative revision path from criticism through retained change to later behavior would alter these assessments. Complete characterization of unassessed evaluators could also change their scope. epistemic conclusion, RT-RTE-workflow-trial, RT-RTE-binary-upgrade.

Limitations

Limitation	Affected SRC IDs and record citations	Inspected boundary	Conclusion prevented	Evidence that would resolve it
Retrieval coverage diagnostics are not separately characterized in the inventories.	SRC-1; RT-RTE-history-retrieval; report verification limit	Verification traced coverage attachment and detail/summary/budget reduction in cmd/flopwire/retrieve.go, cmd/flopwire/coverage.go and internal/coverage/summary.go.	Complete comparison of retrieval metadata, diagnostic lineage and budget-dependent omission. Omitted issue counts cannot be read as zero; this mechanism qualifies reliance rather than accepting truth.	Accepted records tracing diagnostic inputs, projection, omitted/unknown fields and actual consumer output under each budget.
No registered execution or causal comparison; external host implementations are outside the frozen evidence.	SRC-1; boundary; RT-RTE-history-retrieval; WF-RTE-request-reply	Shipped Flopwire code and instructions, without executing them.	Observed delivery, accurate recovery, task completion, deployment reliability, benefit or generalization across host versions.	Route-linked traces and independently checked outcomes under declared host/configuration conditions; suitable comparisons for causal claims.
Extraction and command-grammar inspection is selective.	SRC-1; RT-RTE-transcript-index; MEM-RTE-derive-navigation; EPI-RTE-commit-check	JSONL extraction, selected parser and command/result branches, digest and projection consumers.	Complete transcript fidelity, semantic summary accuracy, exhaustive spoof resistance, authenticated Git history or sole authorship.	Accepted records covering remaining parsers/grammar and independently authenticated source/outcome evidence.
Original host compaction and opaque checkpoint alternatives remain unresolved.	SRC-1; MEM-OBJ-compacted; MEM-RTE-capture-compaction; profile verification	Readable extraction and supported replacement messages; opaque items are skipped.	Complete checkpoint form, lineage, admission, trace source, budget, selection or host reuse; validated summary content.	Frozen producer and host-consumer evidence covering the opaque alternative and its admission and transformation.
Adapter-retained storage and host consumption are unresolved.	SRC-1; RT-RTE-adapter-delivery; RT-RTE-hook-delivery	Flopwire hook/adapter insertion, leases, confirmation and busy checks.	Complete storage classification of host copies, proof of model receipt/compliance, exactly-once work or atomic no-wake behavior.	Host storage/insertion contracts and correlated traces through model consumption and task effects; atomic delivery controls where claimed.
Import, remote policy, family reconciliation and grant authenticity are incompletely traced.	SRC-1; MEM-RTE-import-recovery; RT-RTE-shared-upload; MEM-BAP-history-restriction	Native callback wiring, selected receipt/binding and restrictive provenance paths.	Universal upload bounds, recovered-content fidelity or complete remote sharing restriction.	Accepted producer-to-admission-to-consumer records for converters, verification, remote enforcement and provenance reconciliation.
Withdrawal and backup have distinct scopes and unaudited race/deployment behavior.	SRC-1; MEM-RTE-redact-content; MEM-RTE-withdraw-history; MEM-RTE-restore-state	Selected masking, hiding, tombstone, object-purge and restore implementations; native raw files remain separate.	Universal secret removal, global erasure, purge completion or successful deployed restore; already exposed context is not retracted.	Scoped end-to-end evidence including native sources, devices, archive repair, independent backups and completed deletion/restore jobs.
Explicit caller choices do not establish human content approval.	SRC-1; RT-RTE-policy-admission; MEM-RTE-redact-content; MEM-RTE-import-recovery	Software admission and authenticated requests, including password-gated sender acceptance.	Complete human-versus-software agency classification or proof of per-action human presence.	Evidence identifying who supplies or approves each retained-content decision, separate from software validation and I/O.
Evaluator and maintenance alternatives remain incompletely assessed.	SRC-1; RT-RTE-workflow-trial; RT-RTE-binary-upgrade; epistemic conclusion; profile verification	Selected synthetic qualification, parser oracle and upgrade paths; other evaluator/update consumers unresolved.	Complete memory-operation inventory, global absence of evaluation-driven revision, improved capacity, reflective/autonomous theory building or exercised self-improvement.	Accepted records for remaining update consumers, separate theory-condition assessments, and evidence connecting criticism, retained revision and later causal use to a declared objective.