FrameCode-VibeWork synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
FrameCode-VibeWork (FCVW) organizes AI-assisted project changes around minimum context, a plan, scoped execution, evidence and a version record. It is a Markdown governance and memory system, not an enclosing agent runtime. This account covers the whole shipped FCVW system at Git commit 847cdcedc02d08531961c8367f4e2ddb2dc392f1, with applicability cutoff 2026-11-08. SRC-1 is the sole frozen source. Python supports implementation findings; policies and skills support doctrine/design findings. No host/model execution trace, consuming-project QA run or causal experiment was supplied, and the analysis did not execute the target. The external host, model, application and user approvals remain dependencies rather than FCVW implementations. See the boundary and evidence and Source register.
Entering a governed change
A user asks an external host agent or human to change a project. That host owns the next step: read the entry point, classify session and event impacts, load applicable policies, create or resume a plan, implement, validate and close out. FCVW supplies the work contract, not an inference loop or filesystem sandbox. Plans retain scope, acceptance, risk, regression checks and rollback. Read-only work and narrowly specified trivial prose changes have no-plan alternatives; the representative workflow here uses the standard plan path. RT-RTE-governed-change.
Optional Python routing unions supplied session requirements, explicit events, file-derived events and active-plan context_files. It returns mandatory paths, reasons and reading ranges; it does not deliver all mandatory bodies. Complementary lexical retrieval can rank copied Markdown chunks and expand one graph hop. A versioned-change invocation requires event and file declarations, but application semantic impacts remain caller-owned. Empty selectors and omitted semantic facts can under-route: valid syntax does not establish complete task context. RT-RTE-context-delivery.
Context control is selective loading, not management of the host's full window. Skills load on demand; section reads omit unrequested bodies; archive links allow later reload. With the optional context budget, selection receives complete chunks and skips oversized candidates. Without it, ordinary excerpts are sliced at 1,200 characters. The estimate covers the serialized optional-result array, not mandatory documents, actual tokenizer usage or existing host turns. FCVW does not evict those turns or allocate isolated worker contexts. RT-CLM-context-controls, RT-RTE-mcp-checks, MEM-RTE-session-handoff.
Executing, reviewing and delivering
The representative multi-step workflow connects one active plan to independent work packages and main-agent review. Delegation requires current authority, real host tools, non-overlapping ownership and a reason to expect time or review-quality benefit. Otherwise one executor or sequential packages suffice. Each handoff names exact files, required skill, preserved behavior, evidence and rollback. FCVW prescribes collision checks and soft locks but does not enforce worker filesystem ownership or context isolation. The main agent reviews every return; a worker's success assertion cannot close the plan. RT-RTE-work-packages, workflow package annotation.
Review tests preservation as well as the requested feature. Insufficient blast-radius evidence, pending regression results or undocumented failing checks block prescribed completion. A failing or unavailable check can be handled only with an explicit limitation, residual risk, owner and decision; bypass additionally requires named authority and bounded follow-up. Recovery belongs to the host: split, reduce, defer, obtain required approval or perform authorized rework. Confirmed regressions should lead to a strengthened replay or guardrail, not merely another note. Delivery means scoped changes, decided acceptance, reproducible evidence, final regression dispositions, a changelog/release record and consistent plan status/directory. It is not demonstrated deployment or a mandatory commit/push. Workflow account, workflow change annotation.
Optional hooks strengthen only selected host paths. When enabled by the automation contract and environment, pre-edit can deny recognized edit-tool calls lacking an active/closing plan. It does not match edits to plan scope; shell edits, unrecognized tools and exempt paths are outside that dispatch. Stop can request repairs, but recursion, failures and timeouts include fail-open branches. These are host-honored protocol controls, not universal edit or completion barriers. MCP separately confines direct resolved file arguments within its root and caps section content, but that guarantee does not cover arbitrary host tools or other FCVW interfaces. Route references can flag out-of-root paths without reading them. RT-RTE-hook-lifecycle, RT-RTE-mcp-checks.
Retaining work and bringing it back
FCVW separates raw imports, historical evidence, reusable wiki understanding and current canonical project state. The curator asks the host to promote only reusable, sourced, nonduplicate knowledge, preferably by updating an existing page. Source review may confirm, update, supersede, invalidate or defer. These are content-based admission instructions, not evidence that a particular proposition was accepted. Retention and lifecycle linking remain distinct from acceptance. Without paired inputs, outputs and decisions, generated prose cannot be classified as an accepted new inference merely because it is called a synthesis. MEM-OBJ-wiki, EPI-RTE-wiki-review, EPI-RTE-wiki-integrate.
The continuity mechanism is a bounded host-authored session handoff from actual workspace, plan, decisions, validation and risks. A later host selects a relevant handoff by scope/date/unique ID, verifies it against disk and follows links. Rotation archives older sessions while preserving evidence; it does not implement transcript eviction or automatic deletion. The model-host branch can admit summaries without separate human approval per item. This supports a prescribed trace-fed durable-artifact route, not an observed automatic FCVW transcript service or improved capacity. MEM-OBJ-session, MEM-RTE-session-handoff, RT-RTE-asset-admission, profile verification.
Python builds disposable JSONL text indexes and typed graph/review state from files. Later requested retrieval and MCP section calls can return retained bodies; queue calls return plan identities, derived metadata and recommendations, while routes return references and digest calls return fingerprints. Non-writing tools can therefore consume memory without authoring it. The opt-in SessionStart hook instead pushes all sorted in-progress plan names and up to five pending names into additional context. This is coarse plan-identity supply, not targeted wiki selection or handoff-body loading. MEM-RTE-index-build, MEM-OBJ-knowledge-graph, RT-RTE-mcp-checks, RT-RTE-hook-lifecycle.
The trust policy is stronger than retrieval's own enforcement. Explicit authority and retrieval-scope metadata precede builder defaults. A separate validator can flag incompatible authority, but no mandatory validation-before-index chain is established. Retrieval does not rehash live sources; obsolete, superseded, contradictory and archived records are downweighted rather than universally withdrawn. Graph expansion may return only one target chunk, and ordinary slicing can omit conditions or rationale. These are reachable implementation countercases, not demonstrated malicious deployment. RT-CLM-context-controls, memory delivery annotation, EPI-RTE-evidence-check.
Checking evidence and changing future behavior
Deterministic checks establish narrower facts than semantic warrant. Source digests and direct declared dependencies identify review candidates, not changed truth or automatic revocation. Evidence metadata checks establish selected consistency/reference predicates; optional method guarding, URL non-fetching and local reference existence do not authenticate a claim or test execution. A review report has no automatic canonical-mutation authority. EPI-RTE-source-review, EPI-RTE-evidence-check, MEM-RTE-source-review.
Product QA adds a distinct stronger admission protocol. Approved expectations require approved requirements, criteria, decisions or owner confirmation; code/live behavior alone cannot approve intent. The host is instructed to observe selected application cases and ask the user about every expected/actual divergence before dependent action. The checker binds declared runs to selected parsed contract hashes and rejects incomplete or failed declared results. It cannot attest live execution, evidence truth, hidden coverage or consultation authenticity. A clear decision gate does not turn a failed test into a pass, and hashing does not cover unencoded behavior. MEM-OBJ-product-contract, MEM-OBJ-qa-run, EPI-RTE-live-qa, EPI-RTE-qa-check, EPI-RTE-divergence-decision.
For framework maintenance, the strongest supported contribution is an evidence-responsive method-revision procedure: diagnose failure/drift, patch the smallest behavior-changing section, replay the exposing task and retain updated guidance. Its theory-builder conditions remain separate: (1) localized revisable method content is afforded; (2) decisions depending on its content are claimed; (3) content-directed criticism and revision are requested, not shown performed with blame assigned to a particular theory; (4) retention for later rounds is afforded, not observed iteration. Criticism improving future capacity remains uninspected. A host-mediated reflective pathway is afforded because representations of operating failures can inform rule changes; exercised reflection is not established. Autonomous theory building remains uninspected, rather than disproved by the external user role. RT-RTE-asset-admission.
Self-improvement is likewise a dispositional host-mediated pathway at the boundary of FCVW skill/profile maintenance across sessions, with the independent objective of reducing evidenced failure, drift or cost. Retained revisions can enter later instruction paths, but subsequent operation causally depending on them is unestablished. Thus neither exercised nor successful self-improvement is demonstrated. Release upgrade implements a separate machinery-replacement path with dry run, apply/conflict/prune controls and backups; it is not evidence-driven semantic improvement or atomic rollback. RT-RTE-asset-admission, RT-BAP-policy-consumption, RT-RTE-framework-upgrade.
For a project seeking auditable continuity, the supported contribution is the explicit chain from scoped work to evidence, handoff and later checking, with executable selection and consistency support. For a scenario requiring universal write mediation, authenticated QA or demonstrated learning benefit, the present records do not establish those properties. The retrieval benchmark compares optional selection with a baseline against authored/caller labels, without executing a model; the local driver consumes test/status/byte-snapshot evidence, not downstream task-quality measurements. Actual configured-host traces, authenticated case evidence, paired revision/replay records and subsequent revision-dependent operation would change these assessments. Mandatory validator admission before indexing, live-source freshness checks or broader host interception would change the corresponding implementation guarantees. EPI-RTE-retrieval-benchmark, RT-RTE-local-checks, report verification.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| No supplied host/model or consuming-project execution evidence | SRC-1; RT-CMP-host-model, RT-RTE-governed-change, RT-RTE-work-packages | Shipped instructions and optional local implementations; external executor excluded | Actual compliance, context isolation, delivery reliability or task benefit | Configured-host traces linking loaded inputs, delegated visibility, actions, checks and delivered outcomes |
| Initial and mixed provenance, semantic novelty and acceptance unresolved | SRC-1; MEM-OBJ-wiki, EPI-RTE-wiki-review, MEM-RTE-session-handoff | Inspected curation/handoff contracts, not authenticated project claim histories | Complete lineage/admission inventory, accepted ampliative claim instances or observed trace learning | Paired original inputs, generated artifacts, admission decisions and later consumer records |
| Retrieval copies and trust controls are conditional | SRC-1; MEM-OBJ-context-index, RT-CLM-context-controls, EPI-RTE-evidence-check | Builder, retrieval, optional selector and independent validator | Universal freshness, withdrawal, rationale delivery, authority non-elevation or host-window bounds | Enforced validation/freshness chain, consumer-visible payload checks and explicit host-window accounting |
| QA evidence and consultation remain declared, with unsupplied linked payloads | SRC-1; MEM-OBJ-qa-run, EPI-RTE-qa-check, EPI-RTE-divergence-decision | Selected contract/table checker and QA instructions | Authenticated execution/consent, complete application coverage or classification of linked payloads | Actual scoped application traces, approved-source provenance, user decision events and referenced evidence |
| Navigation/catalog, full dependency/validator internals and historical claim instances are not exhaustively assessed | SRC-1; MEM-OBJ-knowledge-graph, RT-RTE-mcp-checks, report verification, profile verification | Typed graph and selected consumer/check paths; named available alternatives remain unresolved | Complete mechanism/value inventories, exhaustive enforcement or system-wide negatives | Bounded inspection of the remaining material producers, consumers and claim histories |
| Revision persistence does not establish improvement or causal attribution | SRC-1; RT-RTE-asset-admission, EPI-RTE-retrieval-benchmark, RT-RTE-local-checks | Exposed method maintenance and labeled deterministic evaluation, not subsequent host operation | Performed theory criticism/iteration, improved future capacity, autonomy, exercised or successful self-improvement, component-level causal benefit | Content-directed diagnosis/revision records, retained-result consumption, subsequent causally dependent operation and appropriately controlled outcome comparisons |
Both independent verifications report no blockers or additional publication limits. No surviving substantive cross-report conflict is identified; the stale memory integration-issue wording does not override the corrected MCP/hook records. This synthesis required no additional source reads. See report verification and profile verification.