ByteRover CLI synthesis
Type: agentic-system-analyses/types/agentic-system-synthesis.md
Bounded synthesis
Request, authoring and delivery
ByteRover CLI supplies persistent project memory to coding agents. This account covers the repository at commit 1052ac1a5dd0fde4da8693d4712064f7876c269c, with an analysis cutoff of 2026-10-15. Its evidence is inspected implementation, shipped design text and attributed evaluation reporting from SRC-1, not executed workflows or causal experiments. The boundary is complete artifact, partial loop: first-party CLI, daemon, prompts and client adapters are included, but external coding agents, model providers, Cloud internals, transport dependencies and the shared UI implementation are not. Consequently, a wired target-side handoff is not evidence of successful host use. See the boundary and Source register.
The representative workflow begins with brv curate and a knowledge intent. ByteRover stores a UUID-correlated session and returns an authoring prompt; the calling agent generates HTML. On continuation, the CLI sends that HTML to the daemon. The writer checks topic structure, path and overwrite policy, then writes the topic. Writer rejection prompts caller-owned correction; four unsuccessful writer responses exhaust the CLI session, while malformed envelopes do not consume that budget. Success clears the session and returns the stored file path. ByteRover's own LLM does not generate content on this public route. These are implemented protocol steps, not a supplied successful transcript: RT-RTE-curate-session, RT-RTE-topic-admission and the workflow account.
The deliverable is a local context-tree topic, not an accepted factual conclusion. Typed fact validation checks attributes and markup, not the assertion against project evidence. No per-content human approval is required by this path; a generic caller identity or overwrite flag does not establish human semantic control. Topic success also does not guarantee atomic completion of logging, backups, signals and deferred indexing. A thrown dispatch preserves a response file, but does not prove that the daemon made no write; optional cleanup can follow a structured rejection, and cleanup errors can accompany successful delivery. See EPI-RTE-fact-validation, RT-RTE-topic-admission and workflow recovery.
Later selection and authority
A later brv query runs deterministic retrieval and rendering; the external caller owns synthesis and action. Search ranking, local-origin filtering, readable-file filtering and a document-count limit determine what arrives. Search counts therefore need not equal delivered documents. Topic assertions and reasons supply context; retained rules and decisions can supply directions, but rendered severity labels do not enforce compliance. Rationale can survive through bv-reason, without proving that every rule has a sound reason or that a reader uses it. See RT-RTE-query-retrieval, RT-BAP-topic-guidance and MEM-OBJ-bv-reason.
Delivery can weaken lineage without changing an assertion's words: the inspected renderer preserves link text but can discard its source URL. Focused brv read --raw is an uncached, explicit-path alternative that preserves available source bytes, not source truth. Query caches instead can reuse earlier rendered content. Their fingerprint scans markdown, so HTML-only changes need not invalidate it; the default 60-second cache TTL bounds reuse, not immediate withdrawal. These differences matter when checking a source or rejecting a topic: EPI-RTE-render-content, MEM-RTE-topic-read and MEM-OBJ-query-cache.
The public path is not the whole runtime. Internal curation runs a configured model in a task-scoped session, exposes selected tools and writes final HTML with overwrite enabled. SDK query-mode mutation guards and optional tool scheduling are scoped controls, not demonstrated OS isolation. Internal query can automatically supply broad manifest-selected context within a requested execution; its resolver ignores query relevance, and the inspected constructor does not inject search's signal store. Markdown-oriented summaries and abstract siblings cannot be assumed to cover all HTML topics. See RT-RTE-agent-curation, RT-CMP-content-generator, MEM-RTE-manifest-injection and MEM-OBJ-summary-index.
Feedback, maintenance and recovery
The strongest supported learning-related contribution is implemented nonparametric feedback and continuation memory. Returned search hits feed retained access/importance/maturity signals that later influence ranking and prune candidates. Successful rewrites separately update write-history signals. These are activity-responsive selection changes, not outcome-tested semantic corrections; a tier named validated is not factual validation. Full compaction and rolling checkpoints also automatically derive retained continuation material from conversation or tool activity. The memory record supplies a summary-loading witness; the rolling checkpoint instead retains counts, tool names and short assistant snippets in a sandbox variable and asks the model to read it after history clear. This supports trace-fed write-and-later-consumer mechanisms without establishing improved capacity, actual reload or restart durability. See MEM-RTE-signal-feedback, the admission annotation, MEM-OBJ-compaction-boundary and MEM-OBJ-checkpoint-progress.
Dream maintenance computes structural candidates and delegates semantic decisions to the caller. Age and importance predicates establish conditions on stored fields, not obsolete knowledge. Survivor edits pass through separate admission; finalize moves caller-selected files outside the active tree. A different archive service stores a searchable cue and verbatim payload with explicit expansion, rather than the dream format. Neither a registered reload tool nor an archive writer API establishes a deployed automatic lifecycle. See RT-RTE-dream-maintenance, EPI-RTE-prune-candidates, MEM-OBJ-dream-archive and MEM-RTE-archive-expansion.
Review is conditional and post-write: pending topics remain locally retrievable. Rejection can restore a backup or delete a file; missing backups limit restoration. Review statuses veto selected legacy Cloud exports, not all reads or every push family. Dream undo likewise restores selected logged state, not factual warrant, and can retain errors while marking a log undone. Direct editing, hub bundles and Cloud pull have different admission paths and do not inherit the HTML writer's checks. Thus neither mandatory pre-use approval nor universal rollback follows: EPI-RTE-review-decision, EPI-RTE-dream-undo, RT-RTE-direct-edit, RT-RTE-hub-install and RT-RTE-cloud-sync.
Assessment at this boundary
For a caller seeking persistent context with explicit authoring and retrieval handoffs, the source establishes a concrete retain-and-return protocol. For a scenario requiring factually warranted memory, immediate retraction or guaranteed restoration, the inspected paths establish narrower controls, not those guarantees. Shipped query-first/curate-after instructions request workflow order; they do not establish binding host enforcement. Some public wording also assigns LLM retrieval or synthesized answers to ByteRover, whereas the current public query route leaves synthesis outside it. See RT-CLM-project-memory, EPI-RTE-hook-instructions, EPI-CLM-agent-native and EPI-CLM-query-answer.
Reported benchmark accuracy supports attributed answer-scoring outcomes, not independently observed benefit of this frozen route. The reported query-time contrast also removes a separate justifier, so attribution is no finer than that treatment bundle; the curated tree remains fixed and does not isolate write-path contributions. A matched experiment varying curation independently, with question-level outputs and retained-state lineage, would change that assessment. See EPI-CLM-benchmark-results.
The accepted records do not establish a formulated-theory episode, so theory-builder conditions remain separate unresolved findings: 1, no supplied whole theory instance is established; 2, delivery is established but decisions depending on that theory's content are not; 3, structural rejection and activity scores do not establish stated content-directed criticism with blame and revision; 4, retained feedback is wired but retention of such criticism into a next round is not established. No improvement of future-action capacity through criticism of a consumed theory is demonstrated. These are not system-wide absence findings: workers and external model processing remain partly uninspected. See MEM-OBJ-bv-rule and the epistemic bounded conclusion.
The sidecar provides a bounded two-way operational self-representation: retrieval activity changes retained signals, and those signals affect later selection. That does not establish reflective criticism of the system's method. Internal model/tool execution can proceed without per-content human approval, but role-complete autonomous theory building is not established, especially where public authoring and semantic maintenance depend on excluded callers. These properties do not follow from each other. Nor does the feedback loop establish exercised self-improvement: for an objective such as future task accuracy, hits are not outcome evidence, and no supplied subsequent operation causally demonstrates dependence on an improvement-directed update. The supported contribution remains activity-responsive selection and retained continuation context. Traces connecting criticism, update, later consumption and independently specified outcomes would support stronger learning or self-improvement findings; persistence alone would not. See MEM-RTE-signal-feedback, RT-RTE-agent-curation and RT-RTE-dream-maintenance.
Limitations
| Limitation | Affected SRC IDs and record citations | Inspected boundary | Conclusion prevented | Evidence that would resolve it |
|---|---|---|---|---|
| Partial external loop and no execution evidence | SRC-1; Source register, RT-BAP-topic-guidance, RT-CMP-content-generator | Frozen repository and first-party adapters; excluded hosts, providers, Cloud, shared UI and transport internals | Deployed compliance, exact model behavior, successful end-to-end use, isolation and benefit | Frozen dependency/host evidence plus inspectable execution and outcome records |
| Additional context frame omitted from the reports' inventory | SRC-1; report-verification limit, RT-RTE-agent-curation, RT-RTE-context-control | Verifier inspected SDK agentQuery, sandbox binding and parent-child session linkage: child query, optional selected variables, returned text and finally deletion |
Exhaustive delegation-frame and input/return comparisons; actual grant/use remains unobserved | Separate accepted frame inventory with grants, selection, return and invocation evidence |
| Write-history feedback omitted as a separate trace-learning/source unit | SRC-1; profile-verification limit, RT-RTE-topic-admission, MEM-OBJ-runtime-signals, MEM-RTE-signal-feedback | Successful rewrites update importance, recency, update count and maturity for later selectors | Exhaustive signal-learning route and original-input comparisons; existing positive witnesses remain valid | Separate classification of rewrite activity, durable updates and later consumers |
| Reconciliation's compaction reload wording conflict, addressed by report verification | SRC-1; RT-RTE-context-control, MEM-OBJ-compaction-boundary, verification disposition | Runtime quotes post-boundary wording; memory includes the summary boundary. Verifier inspected loader and conversion, dismissing summary exclusion | Treating runtime's wording as proof the summary is discarded, or presenting the unchanged reports as an unambiguous exact history inventory | Aligned declaring-report wording and an exact retained-message account; the verifier already supports boundary inclusion |
| Available alternative writers, consumers and compression branches remain incompletely inspected | SRC-1; memory coverage, RT-RTE-agent-curation, RT-RTE-cloud-sync, RT-RTE-context-control | Registered ingest/map/swarm, Git VC/worktrees, migration, connector loading, alternate compression and downstream log consumers | Complete memory inventory, universal admission/rollback and system-wide negatives about checking or learning | Accepted producer/admission/later-consumer records for those branches |
| Format-divergent derivation and freshness | SRC-1; MEM-OBJ-summary-index, MEM-OBJ-file-abstract, MEM-RTE-manifest-injection, MEM-OBJ-query-cache | Markdown-oriented summary/fingerprint paths versus HTML authoring and dream loading | Universal derived-context coverage and immediate withdrawal after HTML edit/reject/archive | Complete cross-format wiring and invalidation evidence; exercised freshness tests |
| Semantic warrant, recovery and archive lifecycle remain bounded | SRC-1; EPI-RTE-fact-validation, EPI-RTE-review-decision, EPI-RTE-dream-undo, MEM-RTE-archive-expansion | Structural writer, conditional post-write review, latest-log undo and exposed archive service | Evidence-consuming factual acceptance, transactional restoration or mandatory automatic archival/reload | Recorded factual criteria/evidence and acceptance consumers; complete backup/log and archive-cycle traces |
| Benchmark attribution and improvement remain unverified | SRC-1; EPI-CLM-benchmark-results, MEM-RTE-signal-feedback, MEM-OBJ-checkpoint-progress | Reported answer scoring and bundled query ablations; wired activity/continuation updates | Current-route accuracy, isolated curation benefit, successful learning or exercised self-improvement | Matched independent interventions, concrete retained updates, later dependent operation and outcome evidence |