feat: knowledge intake and consolidation (v1.36.0)#145
Merged
Conversation
Phase-0 artifacts for the knowledge intake and consolidation wave: design doc (consultant Variant 2, two shared seams), implementation plan (8 tasks in two tracks), and the variant audit trail with the raw consultant output. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…-note contract (I1, t_f8f5ef6a) Carry seam 1 of the knowledge-intake-and-consolidation wave: the capture-note contract module that owns the inbound staging vocabulary, plus the long-poll Telegram capture bot that writes only through it. Seam 1 (src/core/brain/capture/capture-note.ts): frontmatter kind (brain-capture), provenance (source, sender, capture timestamp), staging and archive path helpers, and read/write/list/archive functions. The layout mirrors the inbox-versus-processed distinction in governance/forget-plan.ts but lives in its own Brain/captures subtree so the signal inbox and the dream pass stay byte-identical for vaults that never capture. Bot (src/core/brain/capture/telegram-capture.ts): an explicit CLI runner verb (o2b brain telegram-capture run|catchup) long-polling getUpdates via fetch with no new dependencies and no webhook framework. It reuses the MarkdownV2 escaping from src/core/discipline/telegram.ts and a new bot token config key. A chat-id allowlist from config gates every update; each accepted text becomes one capture note through the contract; every rejected or malformed update is one explicit logged decision in a JSONL ledger; /catchup replies with captures since the last acknowledged one; a missing token is a typed error at startup. Nothing runs from hooks. The update-handling core is unit-tested with an injected transport (no real network); the runner verb is tested for the missing-token error and the disk-only catchup path. Deviations from plan.md (recorded per the codebase-over-plan rule): - No pre-existing "bot token config" existed (discipline delivery is handled externally by Hermes cron), so seam 1 introduces resolveTelegramBotToken / resolveTelegramCaptureAllowlist in src/core/config.ts following the existing config-resolver conventions; the token key is redacted by redactMapping. - Captures live in Brain/captures (+ /processed) rather than sharing the signal inbox directory, to guarantee the byte-identical opt-out for the dream and signal machinery. - The verb lives under "o2b brain" (where the knowledge features live), matching the curated command-manifest subset that omits recent verbs. - Rejected/malformed decisions are logged to a dedicated Brain/log/capture-decisions.jsonl sidecar instead of the validated brain-log event-kind enum, keeping the opt-out byte-identical. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…s (I2, t_b0bba8cb)
Add the report-then-apply pass that drains staged captures written by the
I1 Telegram bot. It walks the staging area through the seam-1 capture-note
contract and classifies each capture from STRUCTURAL signals only, never a
natural-language word list:
- source-reference: a url-shaped body (a single http(s):// token) ->
routed through ingestSource (src/core/brain/ingest/ingest.ts).
- obligation: a leading @obligation marker, optionally carrying a cadence
from the obligation vocabulary -> routed through the existing obligation
machinery (addObligation), with a named-constant default cadence.
- idea: everything else is an atomic idea -> create-or-merge note under
captured/, merging into an existing same-slug note instead of forking.
Dry-run is the default and writes nothing (regression-tested); --apply
executes the route and archives each processed capture through the
contract, so a rerun finds no staged captures and is a no-op (the
processed marker is the idempotency key, regression-tested). Each item is
reported with its action and reason; unroutable items (an obligation
marker without a title, a same-title collision) are reported and left in
place. Adds obligationExists to obligations.ts for the structural
collision check.
Deviations from plan.md (recorded per the codebase-over-plan rule):
- "explicit frontmatter" source detection reduces to url-shape here because
every capture carries the same brain-capture kind; url-shape and the
obligation marker are the active structural signals.
- Idea notes land in a vault-relative captured/ directory via the
create-note write envelope, which refuses the Brain machinery root; the
merge path appends to the existing note.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…tity gate (S1, t_40fa4e8d)
Give every deep-synthesis finding an additive causal-context field and a
decomposed, deterministic confidence, and gate emission on a shared
evidence-identity type so identity-less findings are a visible loss
rather than a silent drop.
deep-synthesis.ts and its types:
- Export EvidenceIdentity (path + kind + content hash) and the
hasEvidenceIdentity predicate. Identity is tied to retrievable
content: the hash is a sha256 of the evidence bytes. S2 (diarization)
consumes this shared type for stated-vs-evidenced lines.
- Each matched, readable note becomes one SynthesisFinding carrying its
evidence identity, a SynthesisCausalContext (typed relations,
superseded_by pointer, dangling-citation count), and a
SynthesisConfidence decomposed into support (incident positive
relations), opposition (incident contradictions), freshness (recency
in [0,1] from note age against the stale-age horizon), and coverage
(resolvable citations). Every component is computed from data already
flowing through the pass; no model call, no language analysis.
- The read-failure path that used to `continue` silently now records an
excludedFindings entry with a structural reason; the report carries
excludedFindingCount so the loss is always visible.
Additive only: the prior report fields, the `checked` dimensions, and the
steelman seed selection (buildStrongestObjection) are untouched, so
existing consumers and the CLI/MCP surfaces stay byte-identical.
Regression-tested on the prior output shape and the objection ordering.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…, t_28ba3fc4)
Add subject diarization: given a registry entity, assemble its document
set and emit a profile skeleton plus one needs-llm-step envelope for the
prose, with a deterministic stated-vs-evidenced section. No model is ever
called inline.
Core (src/core/brain/diarization.ts):
- Resolves the entity through the registry; an unknown entity is a typed
DiarizationError.
- Document set = the entity registry file plus every ingested source
page the subject appears in (anchored through the existing atomic-fact
machinery, not a word list).
- STATED claims are atomic assertions decomposed from the entity's own
registered body that anchor the subject. EVIDENCED signals are the
frequency and recency with which the subject appears across ingested
sources - pure counts and timestamps, never language analysis of the
claim text. Each gap line is classified stated_corroborated,
stated_unevidenced, or evidenced_unstated.
- Every line carries the shared S1 evidence-identity type
(EvidenceIdentity from deep-synthesis) and is gated on it; an
identity-less line is reported in excludedLineCount, never dropped
silently.
- Emits the profile skeleton (frontmatter + sections + a prose marker)
and exactly one needs-llm-step envelope mirroring the write-session
grammar (status, step, prompt, schema hints, target path) as plain
read-only data.
Surfaces: new CLI verb `o2b brain diarize <entity>` and, per the
deep-synthesis precedent for a read analysis, an MCP twin brain_diarize
(full tier, read-only). Registry baselines updated (parity frozen list,
mcp.test.ts and removed-tools.ts tool counts 106 -> 107); tool
description within the 300-char cap.
Deviations from plan.md (recorded per the codebase-over-plan rule):
- Profile notes target Brain/profiles/<entity-id>.md; diarization emits
the skeleton and envelope only and never writes, so the write happens
through the caller's own submit path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…se (S3, t_c5263e27) Add a deterministic, count-only rollup ladder to the dream synthesize phase. It never reads a fact's content - only how many facts exist at each rung since that rung last rolled up. When the count of new facts at a rung reaches its threshold, the ladder emits one needs-llm-step rollup envelope (the LLM writes the summary note; core never does) and records the counter reset in the dream report. rollup-ladder.ts (new): pure planRollupLadder over a persisted per-tier ledger. The rungs compose - a fired fact rollup increments the rollup rung's source count, so enough fact rollups cascade into one identity rollup in the same pass. The top rung is `identity`, literally the highest existing frontmatter tier weight. Named-constant thresholds (DEFAULT_FACT_ROLLUP_THRESHOLD, DEFAULT_ROLLUP_IDENTITY_THRESHOLD), config-overridable through a new optional `rollup:` block in _brain.yaml (resolveRollupThresholds). Ledger read/write helpers plus the Brain/rollup-ladder.json path. dream.ts: the ladder runs before the `changed` gate over the current fact count (preference artifacts); a fired rung counts as a state change. The DreamRunSummary gains an additive `rollups` field, and the dream summary log event gains a `rollups` key - both only when a rung fires. The ledger is persisted only on a firing, non-dry-run pass, so a below-threshold run writes no ledger, adds no report or log entry, and stays byte-identical (regression-tested). Triggers are pure counters, never language analysis. Config plumbing: BrainRollupConfig type, the optional `rollup:` block in the loadBrainConfig parser (hard-error on a mistyped threshold), and the resolve helper. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…R1, t_1dcbf352) Carry seam 2 of the knowledge-intake-and-consolidation wave: the keyed external-fetch helper, the Brave and Tavily providers, a full-page extract step, and the provider-pool wiring in research.ts. Every test mocks at the fetch boundary; no test path reaches the network. Seam 2 (src/core/brain/research/external-fetch.ts): env-gated HTTP with an injected transport, Bearer auth by default (or a named header), typed ExternalFetchError kinds (disabled, network, auth, http, payload), and a shared response cache keyed by the normalized request. The cache key is built only from method, url, sorted query, and body, never from the key, so no credential can leak into a cache key. Composed error messages are run through the shared redactor with the key as a literal, so an upstream error string embedding the key is scrubbed before it surfaces. Providers (research/providers/brave.ts, tavily.ts): each maps its response verbatim into ProviderSearchResult, inventing nothing. A provider joins the pool only when its key env (BRAVE_API_KEY, TAVILY_API_KEY) is set; a keyless pool reports itself empty explicitly (isEmpty) and leaves writeResearchReport untouched, so a vault that never configures a key behaves byte-identically to before this feature (regression-tested). runResearchPool aggregates results and carries provider failures as typed errors, never as content. Extract step (research/extract.ts): fetches full page text through the helper and reduces it mechanically to bounded plain text (no natural-language word list). findingFromExtract hands a verbatim prefix of that text to the existing citation-constrained pipeline, citing the fetched URL, so the report never carries invented content and the citation contract still holds. Deviations from plan.md (recorded per the codebase-over-plan rule): - Seam 2 honors "Bearer auth" as the helper default; Brave really uses the X-Subscription-Token header, so the helper supports a named-header auth scheme alongside Bearer rather than forcing every provider onto Bearer. - The provider contract (ResearchProvider, ProviderSearchResult) lives in research/providers/provider.ts so brave/tavily and research.ts share it without an import cycle; research.ts re-exports the key envs and pool API. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…y holdouts (G1, t_6832aac6) Add the CLI-first memory-graph repair lane and its graph-efficacy holdout harness. Both are deterministic and read only link structure; there is no natural-language word list anywhere. Repair lane (src/core/brain/link-graph/repair-lane.ts): candidate edges are ordered by identity strength (explicit references, session continuity, same-topic evidence, then opt-in inferred), with confidence and endpoints breaking ties. A confidence threshold and a hard per-run write cap are named constants. Dry-run is the default and writes nothing; apply requires the exact confirmation phrase (RepairConfirmationError otherwise). The lane never creates a dangling edge - a candidate whose endpoint does not resolve to a durable note is skipped - so the holdout gate stays satisfiable. A forward scan seeded from each note's current wikilinks and grown as the run writes makes a rerun after apply converge to zero writes (idempotence regression-tested). collectRepairCandidates draws the default candidates from structural signals only: explicit textual references (a note names another note's title without linking it), session continuity (notes co-referenced in one continuity event), and same-topic evidence (co-occurrence over shared references). Inferred candidates are not collected here; they are opt-in at the lane. Holdout harness (src/core/brain/link-graph/graph-holdout.ts): graph-neighbor holdouts measure graph lift (targets reachable only through the graph) SEPARATELY from direct recall (targets that are already a direct neighbor). Every target must resolve to durable memory and hydrate into bounded evidence (capped by a named constant); a dangling edge fails the gate. Both the resolve-and-hydrate pass and the dangling-fails case are tested. CLI: o2b brain repair-lane (dry-run default; --apply --confirm "apply repair"; --include-inferred; --json), wired through the brain verb dispatch and help text. Deviations from plan.md (recorded per the codebase-over-plan rule): - The lane takes caller-supplied candidates and a separate collectRepairCandidates derives them from vault structure, so the ordering/gating/idempotence core is unit-testable in isolation from candidate generation. - The holdout harness builds a canonical-key adjacency from note wikilinks rather than the Store snapshot, so it runs without a built search index; graph lift is a 2-hop-not-1-hop reachability measure over that adjacency. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
… (K1, t_6fc8663c)
Add a deterministic pre-promotion verifier gate to skill-proposal learning,
plus version tracking and same-name support merging, without changing the
human accept/reject flow (regression-tested).
Verifier (src/core/brain/skill-proposal-verifier.ts): a pure gate that runs
before a draft reaches pending, validating the candidate against its OWN
supporting records - a present pattern key, distinct evidence count, count of
records that structurally support the pattern, and a confidence floor (named
constants). Every check is a fixed structural predicate; there is no natural-
language word list. A rejection records a concise reason naming the failed
checks.
learnSkillProposals integration:
- Each candidate passes the verifier before any write; rejections are appended
to a JSONL ledger (Brain/skill-proposals/verifier-rejections.jsonl) and
reported in the new verifierRejected result field. Support is re-derived per
record (recordSupportsCandidate) reusing the existing action/shape/hour
derivation, so the gate validates real support, not just a count.
- Same-name identity is the (pattern_kind, key) name_key, independent of the
specific records. A candidate colliding with an ACCEPTED proposal evolves
that skill (increments its version and its procedure's version, merges
support) instead of forking; a collision with a PENDING draft accumulates
support into it; a collision with a human-REJECTED name stays suppressed.
- New drafts carry name_key and version 1; acceptSkillProposal stamps the
version onto the accepted proposal and the generated procedure. Reading a
legacy proposal without a version field defaults to version 1, so existing
proposals are untouched by the field (byte-identical opt-out).
Deviations from plan.md (recorded per the codebase-over-plan rule):
- The verifier takes a projected VerifierCandidate (evidence as {id,
supportsPattern}) rather than raw continuity records, so it stays a pure,
independently-tested predicate while learnSkillProposals supplies the
structural support derivation it already owns.
- The verifier's minimum evidence tracks the run's effective minSupport, so a
legitimate minSupport-2 run is not rejected by a fixed floor while the
default (minSupport 3) still gates thin candidates.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The graph-efficacy holdout gate passed whenever no edge was dangling, so a target that resolved to durable memory but hydrated into an empty body still passed. The repair-lane design requires every target to both resolve AND hydrate into bounded evidence, so a resolved-but-empty target must fail. Pass the gate only when danglingCount and unhydratedCount are both zero, and surface unhydratedCount on the gate result so a failed gate names its reason alongside danglingCount, keeping the per-resolution hydrated/evidenceChars fields intact. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
…nd exclusions on CLI and MCP surfaces The S1 finding pass (t_40fa4e8d) added per-finding causal context, decomposed confidence (support/opposition/freshness/coverage), and the excludedFindings/excludedFindingCount evidence-loss ledger to the core DeepSynthesisReport, but neither the `brain deep-synthesis --json` verb nor the brain_deep_synthesis MCP handler emitted them, so the operator saw the pre-S1 dossier shape and the evidence loss stayed invisible. Add a shared synthesisFindingsJson serializer in core so both surfaces emit byte-identical snake_case findings, excluded_findings, and excluded_finding_count. Prior fields are unchanged. Refresh the stale t_04e94382 docstring reference on the verb and note the new fields in the MCP tool description within the 300-char registry cap. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
….36.0 CHANGELOG entry and link reference for 1.36.0, README What-is-new rewrite with a one-line pointer to the previous release, cli-reference section for the capture runner, inbox drain, diarize, repair lane, and synthesis JSON additions, and mcp.md bullets for brain_diarize and the additive deep-synthesis fields. Version propagated with scripts/sync-version.ts. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
solaitken
enabled auto-merge (squash)
July 19, 2026 18:32
|
No actionable comments were generated in the recent review. 🎉 ℹ️ Recent review info⚙️ Run configurationConfiguration used: defaults Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (2)
📝 WalkthroughWalkthroughVersion 1.36.0 adds Telegram capture and inbox draining, keyed web research and extraction, deterministic rollups, diarization, enriched synthesis findings, graph repair, skill verification/evolution, and corresponding CLI, MCP, documentation, and test updates. ChangesKnowledge intake and consolidation
Estimated code review effort: 5 (Critical) | ~120 minutes Possibly related PRs
Suggested reviewers: Sequence Diagram(s)sequenceDiagram
participant Operator
participant BrainCLI
participant Telegram
participant CaptureStore
participant InboxDrain
participant BrainVault
Operator->>BrainCLI: telegram-capture run
BrainCLI->>Telegram: poll updates
Telegram-->>BrainCLI: text update
BrainCLI->>CaptureStore: write staged capture
Operator->>BrainCLI: inbox-drain
BrainCLI->>InboxDrain: classify staged captures
InboxDrain->>BrainVault: route and archive captures
BrainVault-->>BrainCLI: drain report
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
There was a problem hiding this comment.
Actionable comments posted: 12
🧹 Nitpick comments (5)
src/core/brain/skill-proposals.ts (1)
573-581: 🩺 Stability & Availability | 🔵 Trivial | 💤 Low valueVerifier-rejection ledger grows without bound and swallows no write errors.
appendVerifierRejectionappends one line per rejection on every run with no rotation/cap, and a failedappendFileSyncwould surface as an unhandled throw that aborts the wholelearnSkillProposalspass. Consider a size/age cap on the ledger and/or isolating the append so a ledger-write failure doesn't fail intake. Low priority given intake volume.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@src/core/brain/skill-proposals.ts` around lines 573 - 581, Update appendVerifierRejection to bound the verifier-rejection ledger by applying an appropriate size or age rotation/cap before appending, and isolate filesystem failures so append errors are handled without aborting the surrounding learnSkillProposals pass. Preserve the existing JSONL entry format and vault path safety.src/core/brain/link-graph/repair-lane.ts (1)
365-380: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick winExplicit-reference matching has no safeguard against generic/short titles at fixed high confidence.
mentionsTermonly filters titles shorter than 2 characters (Line 291); any title of length ≥2 (e.g. a common word) is matched at word boundaries across every other note's body and, if matched, always getsEXPLICIT_REFERENCE_CONFIDENCE = 0.9— well above the0.5write threshold and the strongest tier inorderCandidates. A vault with a generically-titled note could generate a large number of high-confidence spurious edges that get auto-applied ahead of the write cap. Consider raising the minimum title length or scaling confidence by title specificity (e.g. length, word count).🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@src/core/brain/link-graph/repair-lane.ts` around lines 365 - 380, Harden the explicit-reference matching in the repair flow around `mentionsTerm` and the `IDENTITY_STRENGTH.explicitReference` candidate by rejecting or down-weighting generic, short titles before assigning `EXPLICIT_REFERENCE_CONFIDENCE`. Ensure common two-character or otherwise nonspecific titles cannot automatically produce high-confidence edges above the write threshold, while preserving strong confidence for sufficiently specific titles.src/core/brain/capture/telegram-capture.ts (1)
220-256: 🩺 Stability & Availability | 🔵 Trivial | ⚡ Quick winHarden the long-poll loop against network hiccups: no fetch timeouts, no retry/backoff.
createFetchTelegramTransport'sfetch()calls have no timeout, so a stalled connection can block the poll loop indefinitely. AndrunTelegramCapturehas no try/catch aroundgetUpdates/sendMessage; any thrown error (including an HTTP error from either call) propagates out and terminates the whole long-running bot process, with no retry/backoff for what's meant to be a resilient poller.♻️ Suggested direction
async getUpdates(offset: number): Promise<TelegramUpdate[]> { const url = `${base}/getUpdates?offset=${offset}&timeout=${LONG_POLL_TIMEOUT_SECONDS}`; - const res = await fetch(url); + const res = await fetch(url, { signal: AbortSignal.timeout((LONG_POLL_TIMEOUT_SECONDS + 10) * 1000) });And in the poll loop, wrap
getUpdates/sendMessageso a transient failure logs and continues instead of crashing the process:- const updates = await opts.transport.getUpdates(offset); + let updates: TelegramUpdate[]; + try { + updates = await opts.transport.getUpdates(offset); + } catch { + cycles += 1; + continue; // transient failure - retry next cycle + }Also applies to: 263-289
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@src/core/brain/capture/telegram-capture.ts` around lines 220 - 256, Harden createFetchTelegramTransport and runTelegramCapture against transient network failures: add a bounded timeout to each fetch call, and wrap getUpdates/sendMessage in the runTelegramCapture loop so failures are logged and the poller continues with retry/backoff rather than propagating and terminating the process. Preserve offset handling and decision processing for successful updates.src/core/brain/rollup-ladder.ts (1)
144-186: 🎯 Functional Correctness | 🔵 Trivial | ⚡ Quick win
sourcecomputation hardcodesROLLUP_TIER.factinstead of "the tier below this rung".Works correctly today because there are exactly two rungs, but the ternary at line 159 will silently break the composition contract described in the module docstring if a third rung is ever appended (its
sourcewould keep readingproduced["fact"]instead of the rung immediately below it). Tracking the previous rung's tier explicitly (e.g. index-based) instead of hardcodingROLLUP_TIER.factwould make the composition correct-by-construction.♻️ Sketch: derive source from the previous rung instead of a hardcoded tier
- const entries: RollupLadderEntry[] = []; - for (const rung of rungs) { - const source = rung.tier === ROLLUP_TIER.fact ? factCount : (produced[ROLLUP_TIER.fact] ?? 0); + const entries: RollupLadderEntry[] = []; + for (let i = 0; i < rungs.length; i++) { + const rung = rungs[i]!; + const source = i === 0 ? factCount : (produced[rungs[i - 1]!.tier] ?? 0);🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@src/core/brain/rollup-ladder.ts` around lines 144 - 186, Update source computation in planRollupLadder to derive each non-base rung’s source from the immediately preceding rung’s tier rather than hardcoding ROLLUP_TIER.fact. Track the prior rung by index or equivalent while preserving factCount for the base rung and the existing produced-update ordering.src/core/brain/deep-synthesis.ts (1)
532-574: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low value
hasEvidenceIdentitygate is currently unreachable at both call sites. In both files, theEvidenceIdentityfed intohasEvidenceIdentityis always built from a non-emptypath, a non-empty constantkind, and asha256Hex(...)digest — which is never an empty string. So the identity check can never returnfalsehere, and the corresponding exclusion path (no_evidence_identityin deep-synthesis, the identity-gated drop in diarization'spush) is dead code today, despite being documented as a live "visible loss" gate.
src/core/brain/deep-synthesis.ts#L532-L574: theno_evidence_identitybranch (lines 546-549) can't currently fire; either drop the redundant check for this construction path or add a code comment noting it's defensive for a future evidence-construction change.src/core/brain/diarization.ts#L186-L222: same — thehasEvidenceIdentitybranch insidepush(190-193) can't currently reject a line built fromentityRel/source-page path +sha256Hex; same call: drop or document as forward-defensive.🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@src/core/brain/deep-synthesis.ts` around lines 532 - 574, The hasEvidenceIdentity checks are unreachable for the currently constructed identities. In src/core/brain/deep-synthesis.ts lines 532-574, remove the redundant no_evidence_identity gate or document it as defensive for future evidence-construction changes; apply the same choice to the hasEvidenceIdentity branch in push in src/core/brain/diarization.ts lines 186-222, preserving the remaining finding and diarization behavior.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Inline comments:
In @.codex-plugin/plugin.json:
- Line 3: Revert the manual version change in the plugin manifest and restore
its prior version value. Treat package.json as the sole version source, and use
scripts/sync-version.ts to propagate any intended version update to mirrored
files.
In `@src/cli/brain/verbs/diarize.ts`:
- Around line 62-65: Update the catch block in src/cli/brain/verbs/diarize.ts
(lines 62-65) to check flags["json"] and emit the { ok: false, error: ... } JSON
envelope for both DiarizationError and other operational errors, while
preserving plain fail() behavior otherwise. Apply the same conditional handling
in the RepairConfirmationError catch block in src/cli/brain/verbs/repair-lane.ts
(lines 60-65), using the existing error message for the envelope.
- Around line 10-21: Update the missing-query validation in cmdBrainDiarize to
call usageError instead of fail, preserving the existing usage message. Import
usageError from ../helpers.ts and ensure this usage-error path exits with the
helper’s code-2 plain-text stderr behavior.
In `@src/core/brain/capture/inbox-drain.ts`:
- Around line 237-265: Separate plan.execute() from archiveCapture() handling so
a successful route is never reported as unroutable when archiving fails.
Preserve the routed result and surface archive-after-success failure distinctly,
ensuring the item is not marked safe to retry or re-executed on the next --apply
run; update the item counters and reason fields around the existing routing loop
accordingly.
In `@src/core/brain/capture/telegram-capture.ts`:
- Around line 121-136: Refactor renderCatchup into a planCatchup flow that
builds the reply and exposes the newest capture watermark without advancing it.
Update handleCaptureUpdate and runTelegramCapture so the Telegram path calls
writeCatchupWatermark only after transport.sendMessage resolves successfully,
while preserving CLI behavior by committing the watermark when stdout delivery
completes.
In `@src/core/brain/dream.ts`:
- Around line 346-358: Rebuild the rollup ladder plan after the mutation path
finalizes runId via nextAvailableDreamRunId, so every fired entry’s
envelope.target_path uses the unique final identifier. Keep the initial plan for
pre-mutation threshold evaluation if needed, but ensure the plan consumed for
rollup execution, ledger updates, and returned artifacts is recomputed with the
corrected runId; add coverage for two non-dry-run dream() calls sharing now with
a firing rung.
In `@src/core/brain/link-graph/graph-holdout.ts`:
- Around line 103-118: Guard the ensureInsideVault call in hydrateEvidence so
targets that escape the vault return { resolved: false, text: "" } instead of
propagating an exception. Keep valid-target parsing and fallback reading
unchanged, ensuring evaluateGraphHoldouts can continue processing remaining
holdouts.
In `@src/core/brain/link-graph/repair-lane.ts`:
- Around line 226-230: Update runRepairLane’s existing-edge check and linked
tracking to handle candidates whose endpointKey returns null, ensuring a
successfully applied candidate is recorded and skipped on subsequent runs.
Preserve the documented convergence behavior for arbitrary RepairCandidate
inputs without changing normal non-null key handling.
- Around line 134-136: Update noteAbsPath and the candidate handling in
runRepairLane so ensureInsideVault failures are caught and converted into the
existing skip-* decision flow rather than escaping and aborting the lane. Apply
this before processing candidate.source or candidate.target, preserve normal
existence checks for valid paths, and ensure invalid endpoints cannot create
edges.
In `@src/core/brain/research/external-fetch.ts`:
- Around line 129-139: Update normalizeRequestKey to include the request’s
accept value in the deterministic cache key, using the same normalization used
by request handling and a stable representation when it is absent. Preserve
exclusion of credentials and headers unrelated to response shape.
In `@src/core/brain/skill-proposals.ts`:
- Around line 129-152: Update the verifier input in the candidate gate around
verifySkillProposalCandidate so it uses the full support count from the
candidate rather than the capped candidate.records sample. Preserve the existing
record-level evidence checks, and ensure supportingCount(candidate) reflects all
records so minSupport values greater than six are evaluated correctly.
In `@src/core/config.ts`:
- Around line 854-858: Update resolveTelegramBotToken and the related
config-value handling around the additionally affected lines to verify
telegram_bot_token is a string before calling trim. Ignore numeric, boolean, and
other non-string values so they resolve to the existing null fallback rather
than throwing.
---
Nitpick comments:
In `@src/core/brain/capture/telegram-capture.ts`:
- Around line 220-256: Harden createFetchTelegramTransport and
runTelegramCapture against transient network failures: add a bounded timeout to
each fetch call, and wrap getUpdates/sendMessage in the runTelegramCapture loop
so failures are logged and the poller continues with retry/backoff rather than
propagating and terminating the process. Preserve offset handling and decision
processing for successful updates.
In `@src/core/brain/deep-synthesis.ts`:
- Around line 532-574: The hasEvidenceIdentity checks are unreachable for the
currently constructed identities. In src/core/brain/deep-synthesis.ts lines
532-574, remove the redundant no_evidence_identity gate or document it as
defensive for future evidence-construction changes; apply the same choice to the
hasEvidenceIdentity branch in push in src/core/brain/diarization.ts lines
186-222, preserving the remaining finding and diarization behavior.
In `@src/core/brain/link-graph/repair-lane.ts`:
- Around line 365-380: Harden the explicit-reference matching in the repair flow
around `mentionsTerm` and the `IDENTITY_STRENGTH.explicitReference` candidate by
rejecting or down-weighting generic, short titles before assigning
`EXPLICIT_REFERENCE_CONFIDENCE`. Ensure common two-character or otherwise
nonspecific titles cannot automatically produce high-confidence edges above the
write threshold, while preserving strong confidence for sufficiently specific
titles.
In `@src/core/brain/rollup-ladder.ts`:
- Around line 144-186: Update source computation in planRollupLadder to derive
each non-base rung’s source from the immediately preceding rung’s tier rather
than hardcoding ROLLUP_TIER.fact. Track the prior rung by index or equivalent
while preserving factCount for the base rung and the existing produced-update
ordering.
In `@src/core/brain/skill-proposals.ts`:
- Around line 573-581: Update appendVerifierRejection to bound the
verifier-rejection ledger by applying an appropriate size or age rotation/cap
before appending, and isolate filesystem failures so append errors are handled
without aborting the surrounding learnSkillProposals pass. Preserve the existing
JSONL entry format and vault path safety.
🪄 Autofix (Beta)
Fix all unresolved CodeRabbit comments on this PR:
- Push a commit to this branch (recommended)
- Create a new PR with the fixes
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: 8c8c6ed3-398b-4653-ad3b-2b80c9fd8e96
📒 Files selected for processing (73)
.claude-plugin/plugin.json.codex-plugin/plugin.jsonCHANGELOG.mdREADME.mddocs/brainstorm/knowledge-intake-and-consolidation/cli-output/claude.mddocs/brainstorm/knowledge-intake-and-consolidation/cli-output/prompt.mddocs/brainstorm/knowledge-intake-and-consolidation/design.mddocs/brainstorm/knowledge-intake-and-consolidation/plan.mddocs/brainstorm/knowledge-intake-and-consolidation/variants.mddocs/cli-reference.mddocs/mcp.mdopenclaw.plugin.jsonpackage.jsonplugin.yamlplugins/codex/.codex-plugin/plugin.jsonplugins/hermes/plugin.yamlpyproject.tomlsrc/cli/brain.tssrc/cli/brain/help-text.tssrc/cli/brain/verbs/deep-synthesis.tssrc/cli/brain/verbs/diarize.tssrc/cli/brain/verbs/inbox-drain.tssrc/cli/brain/verbs/index.tssrc/cli/brain/verbs/repair-lane.tssrc/cli/brain/verbs/telegram-capture.tssrc/core/brain/capture/capture-note.tssrc/core/brain/capture/inbox-drain.tssrc/core/brain/capture/telegram-capture.tssrc/core/brain/deep-synthesis.tssrc/core/brain/diarization.tssrc/core/brain/dream.tssrc/core/brain/link-graph/graph-holdout.tssrc/core/brain/link-graph/repair-lane.tssrc/core/brain/obligations.tssrc/core/brain/paths.tssrc/core/brain/policy.tssrc/core/brain/research/external-fetch.tssrc/core/brain/research/extract.tssrc/core/brain/research/providers/brave.tssrc/core/brain/research/providers/provider.tssrc/core/brain/research/providers/tavily.tssrc/core/brain/research/research.tssrc/core/brain/rollup-ladder.tssrc/core/brain/skill-proposal-verifier.tssrc/core/brain/skill-proposals.tssrc/core/brain/types.tssrc/core/config.tssrc/mcp/brain/knowledge-tools.tstests/cli/brain-deep-synthesis.test.tstests/cli/brain-inbox-drain.test.tstests/cli/brain-repair-lane.test.tstests/cli/brain-telegram-capture.test.tstests/core/brain.dream.rollup.test.tstests/core/brain/capture/capture-note.test.tstests/core/brain/capture/inbox-drain.test.tstests/core/brain/capture/telegram-capture.test.tstests/core/brain/deep-synthesis.test.tstests/core/brain/diarization.test.tstests/core/brain/link-graph/graph-holdout.test.tstests/core/brain/link-graph/repair-collect.test.tstests/core/brain/link-graph/repair-lane.test.tstests/core/brain/research/external-fetch.test.tstests/core/brain/research/extract.test.tstests/core/brain/research/providers.test.tstests/core/brain/rollup-ladder.test.tstests/core/brain/skill-proposal-lifecycle.test.tstests/core/brain/skill-proposal-verifier.test.tstests/core/config-telegram-capture.test.tstests/mcp/brain-tools-parity.test.tstests/mcp/deep-synthesis-tool.test.tstests/mcp/diarize-tool.test.tstests/mcp/mcp.test.tstests/mcp/removed-tools.test.ts
Comment on lines
+62
to
+65
| } catch (err) { | ||
| if (err instanceof DiarizationError) return fail(err.message); | ||
| return fail((err as Error).message ?? String(err)); | ||
| } |
There was a problem hiding this comment.
🎯 Functional Correctness | 🟡 Minor | ⚡ Quick win
Emit JSON failure envelopes on operational errors when --json is requested.
As per learnings, when the user requests --json output, the CLI should emit a { ok: false, error: ... } JSON envelope for operational failures rather than printing plain text via fail().
src/cli/brain/verbs/diarize.ts#L62-L65: Update thecatchblock to checkflags["json"]and emit a JSON envelope if true.src/cli/brain/verbs/repair-lane.ts#L60-L65: Update theRepairConfirmationErrorcatch block to emit a JSON envelope whenflags["json"]is true.
📍 Affects 2 files
src/cli/brain/verbs/diarize.ts#L62-L65(this comment)src/cli/brain/verbs/repair-lane.ts#L60-L65
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@src/cli/brain/verbs/diarize.ts` around lines 62 - 65, Update the catch block
in src/cli/brain/verbs/diarize.ts (lines 62-65) to check flags["json"] and emit
the { ok: false, error: ... } JSON envelope for both DiarizationError and other
operational errors, while preserving plain fail() behavior otherwise. Apply the
same conditional handling in the RepairConfirmationError catch block in
src/cli/brain/verbs/repair-lane.ts (lines 60-65), using the existing error
message for the envelope.
Source: Learnings
Comment on lines
+854
to
+858
| export function resolveTelegramBotToken(configPath?: string): string | null { | ||
| const env = process.env["TELEGRAM_BOT_TOKEN"]?.trim(); | ||
| const raw = env || discoverConfig(configPath).data["telegram_bot_token"]?.trim(); | ||
| return raw !== undefined && raw.length > 0 ? raw : null; | ||
| } |
There was a problem hiding this comment.
🩺 Stability & Availability | 🟡 Minor | ⚡ Quick win
Guard against non-string config values before .trim().
?.trim() only guards null/undefined; a YAML value like telegram_bot_token: 123456 (parsed as a number) or a boolean would throw a TypeError here instead of degrading to the intended null/[] fallback.
🛡️ Proposed fix
export function resolveTelegramBotToken(configPath?: string): string | null {
const env = process.env["TELEGRAM_BOT_TOKEN"]?.trim();
- const raw = env || discoverConfig(configPath).data["telegram_bot_token"]?.trim();
+ const cfgValue = discoverConfig(configPath).data["telegram_bot_token"];
+ const raw = env || (typeof cfgValue === "string" ? cfgValue.trim() : undefined);
return raw !== undefined && raw.length > 0 ? raw : null;
}
export function resolveTelegramCaptureAllowlist(configPath?: string): string[] {
const env = process.env["TELEGRAM_CHAT_ALLOWLIST"]?.trim();
- const raw = env || discoverConfig(configPath).data["telegram_chat_allowlist"]?.trim();
+ const cfgValue = discoverConfig(configPath).data["telegram_chat_allowlist"];
+ const raw = env || (typeof cfgValue === "string" ? cfgValue.trim() : undefined);
if (!raw) return [];Also applies to: 867-877
🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
In `@src/core/config.ts` around lines 854 - 858, Update resolveTelegramBotToken
and the related config-value handling around the additionally affected lines to
verify telegram_bot_token is a string before calling trim. Ignore numeric,
boolean, and other non-string values so they resolve to the existing null
fallback rather than throwing.
Collaborator
Author
|
CodeRabbit triage: 14 findings are being fixed in a follow-up commit (usage-error exit code, JSON failure envelopes, distinct archive-after-route state in inbox-drain, catchup watermark advancing only after a delivered send, rollup envelope rebuilt with the final run id, guarded vault-escape handling in the holdout harness and repair lane, null endpoint-key idempotence, accept-aware cache key, full support count for the verifier gate, non-string config guards, long-poll timeout and retry, index-derived rollup rungs, minimum title length for explicit references, defensive-gate comments). Two deliberate skips: the mirrored-version finding is answered inline (bump was done via sync-version.ts, CI sync check green, same as the withdrawn finding on PR #144), and the verifier-rejection ledger growth note is declined because a ledger append failure aborting intake is this repo's fail-loud convention; rotation is deferred. |
- diarize: usageError (exit 2) for missing entity; JSON error envelope under --json - repair-lane CLI: JSON error envelope for confirmation failures under --json - inbox-drain: split route/archive handling into a distinct archive-failed state; idempotent idea-note merge so a re-run never duplicates - telegram-capture: advance the catchup watermark only after a successful send; long-poll fetch timeout and resilient poll loop that logs transport errors instead of crashing - dream: rebuild the rollup plan after a run-id collision so target_path matches the corrected run_id - graph-holdout: treat a vault-escaping target as dangling instead of throwing - repair-lane core: return null on vault escape (skip-unsafe-path); refuse null-endpoint-key candidates (skip-null-endpoint-key); raise the explicit-reference minimum title length to a named constant of 4 - external-fetch: include the accept type in the cache key - skill-proposals: carry the full supporting record set to the verifier gate - rollup-ladder: derive each rung's source from the previous rung by index - deep-synthesis / diarization: document the defensive evidence-identity gates Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Collaborator
Author
|
Review follow-up landed as 9ca0257: 14 findings fixed (usage-error exit code, JSON failure envelopes, distinct archive-failed state with idempotent idea merge, catchup watermark committed only after a delivered send, rollup plan rebuilt with the final run id, guarded vault-escape handling in both the holdout harness and the repair lane, explicit skip for null endpoint keys, accept-aware cache key, uncapped support count for the verifier gate, long-poll timeouts with a continue-on-transport-error loop, index-derived rollup rungs, a named minimum title length for explicit references, and defensive-gate comments). The non-string config guard turned out already safe: the YAML parser coerces scalars to strings, now locked by a regression test. Full suite 6624 pass. |
There was a problem hiding this comment.
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (1)
src/core/brain/capture/telegram-capture.ts (1)
293-355: 🩺 Stability & Availability | 🟠 Major | ⚡ Quick winNo backoff on repeated
getUpdatestransport errors.The deferred-commit fix here correctly resolves the prior critical finding (watermark now advances only after
sendMessagesucceeds). However, on agetUpdatesfailure the loop justcontinues straight into the next cycle with no delay. Since this runs unbounded in production (runTelegramCaptureis invoked with nomaxCycles, only ashouldStopsignal handler insrc/cli/brain/verbs/telegram-capture.ts), a persistent failure (bad token, DNS outage, Telegram API downtime) turns into a hot loop hammering the API as fast asfetchcan fail, risking additional rate-limiting/bans and pinning CPU.🛠️ Suggested direction: exponential backoff on consecutive transport errors
+ let consecutiveGetUpdatesFailures = 0; while (opts.maxCycles === undefined || cycles < opts.maxCycles) { if (opts.shouldStop?.() === true) break; cycles += 1; let updates: TelegramUpdate[]; try { updates = await opts.transport.getUpdates(offset); + consecutiveGetUpdatesFailures = 0; } catch (err) { decisions.push( logTransportError(vault, -1, null, `getUpdates failed: ${messageOf(err)}`, opts.now), ); + consecutiveGetUpdatesFailures += 1; + const backoffMs = Math.min(30_000, 1_000 * 2 ** (consecutiveGetUpdatesFailures - 1)); + await new Promise((resolve) => setTimeout(resolve, backoffMs)); continue; }🤖 Prompt for AI Agents
Verify each finding against current code. Fix only still-valid issues, skip the rest with a brief reason, keep changes minimal, and validate. In `@src/core/brain/capture/telegram-capture.ts` around lines 293 - 355, Single-line getUpdates failures currently retry without delay, causing a hot loop during persistent transport outages. Update runTelegramCapture to track consecutive getUpdates transport failures and await an exponential backoff before retrying, resetting the failure count after a successful getUpdates call; preserve decision logging, stop handling, and normal polling behavior.
🤖 Prompt for all review comments with AI agents
Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.
Outside diff comments:
In `@src/core/brain/capture/telegram-capture.ts`:
- Around line 293-355: Single-line getUpdates failures currently retry without
delay, causing a hot loop during persistent transport outages. Update
runTelegramCapture to track consecutive getUpdates transport failures and await
an exponential backoff before retrying, resetting the failure count after a
successful getUpdates call; preserve decision logging, stop handling, and normal
polling behavior.
ℹ️ Review info
⚙️ Run configuration
Configuration used: defaults
Review profile: CHILL
Plan: Pro
Run ID: 825d10fd-8854-4c8f-9f97-6fee79bacf4c
📒 Files selected for processing (25)
src/cli/brain/verbs/diarize.tssrc/cli/brain/verbs/inbox-drain.tssrc/cli/brain/verbs/repair-lane.tssrc/cli/brain/verbs/telegram-capture.tssrc/core/brain/capture/inbox-drain.tssrc/core/brain/capture/telegram-capture.tssrc/core/brain/deep-synthesis.tssrc/core/brain/diarization.tssrc/core/brain/dream.tssrc/core/brain/link-graph/graph-holdout.tssrc/core/brain/link-graph/repair-lane.tssrc/core/brain/research/external-fetch.tssrc/core/brain/rollup-ladder.tssrc/core/brain/skill-proposals.tstests/cli/brain-diarize.test.tstests/cli/brain-repair-lane.test.tstests/core/brain.dream.rollup.test.tstests/core/brain/capture/inbox-drain.test.tstests/core/brain/capture/telegram-capture.test.tstests/core/brain/link-graph/graph-holdout.test.tstests/core/brain/link-graph/repair-collect.test.tstests/core/brain/link-graph/repair-lane.test.tstests/core/brain/research/external-fetch.test.tstests/core/brain/skill-proposals.test.tstests/core/config-telegram-capture.test.ts
🚧 Files skipped from review as they are similar to previous changes (16)
- src/cli/brain/verbs/diarize.ts
- tests/core/brain/link-graph/repair-collect.test.ts
- tests/core/brain/capture/inbox-drain.test.ts
- tests/core/brain/research/external-fetch.test.ts
- src/cli/brain/verbs/telegram-capture.ts
- src/cli/brain/verbs/inbox-drain.ts
- src/core/brain/diarization.ts
- src/core/brain/research/external-fetch.ts
- src/core/brain/rollup-ladder.ts
- src/core/brain/link-graph/repair-lane.ts
- src/core/brain/capture/inbox-drain.ts
- src/cli/brain/verbs/repair-lane.ts
- src/core/brain/deep-synthesis.ts
- src/core/brain/dream.ts
- src/core/brain/link-graph/graph-holdout.ts
- src/core/brain/skill-proposals.ts
…lures A failed getUpdates retried immediately, so a persistent transport outage (bad runtime token, DNS failure, API downtime) turned the long-poll loop into a hot loop that hammered the API and pinned the CPU, since production runs with no maxCycles. Track consecutive getUpdates transport failures and await an exponential backoff (base 1000 ms doubling, capped at 30000 ms) before retrying, resetting the counter after any clean poll. The sleep is injectable so tests assert the growing waits without spending wall-clock time, and a stop requested during a backoff ends the loop promptly instead of issuing another poll. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Collaborator
Author
|
Incremental review finding addressed in f673a46: repeated getUpdates transport failures now back off exponentially (base 1000 ms doubling, capped at 30000 ms, named constants), the counter resets after a clean poll, and a stop request during backoff ends the loop promptly. Full suite 6627 pass. |
itechmeat
approved these changes
Jul 19, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
feat: knowledge intake and consolidation (v1.36.0)
Summary
The vault now captures from the operator's pocket and consolidates what accumulates: an inbound Telegram bot lands phone captures in a staged area, a report-then-apply drain pass classifies and routes each capture, keyed Brave and Tavily providers plus a full-page extract feed the citation-constrained research pipeline, and on the consolidation side facts roll up into higher tiers on pure counters, entities get diarized profiles with a stated-vs-evidenced gap, synthesis findings carry causal context and decomposed confidence behind an evidence-identity gate, the link graph gains a capped idempotent repair lane with efficacy holdouts, and skill proposals pass a deterministic verifier before ever reaching the pending queue.
Why this matters
flowchart LR subgraph Intake["Intake: content reaches the vault"] TG["Telegram bot<br/>(allowlist, /catchup)"] -->|"capture contract"| ST["Staged captures"] ST -->|"classify + route<br/>report then apply"| RT["source page / note / obligation"] WEB["Brave / Tavily / page extract<br/>(env-gated, keyed)"] --> RP["Research pipeline<br/>(no uncited claims)"] end subgraph Consolidation["Consolidation: accumulation becomes insight"] F["Facts"] -->|"count trigger"| RL["Rollup ladder<br/>facts to identity tier"] E["Entity documents"] --> DZ["Diarized profile<br/>stated vs evidenced"] SY["Synthesis findings"] -->|"evidence-identity gate"| OK["Emitted + excluded counted"] LG["Link graph"] -->|"dry-run, capped, idempotent"| RE["Repair lane + holdout gate"] SP["Skill proposals"] -->|"verifier before pending"| PQ["Versioned, merged, human-approved"] end Intake --> ConsolidationConcrete benefits
What ships
src/core/brain/capture/,o2b brain telegram-capturesrc/core/brain/capture/inbox-drain.ts,o2b brain inbox-drainsrc/core/brain/research/external-fetch.ts,providers/,extract.tssrc/core/brain/deep-synthesis.tssrc/core/brain/diarization.ts,o2b brain diarize, MCPbrain_diarizesrc/core/brain/rollup-ladder.ts, dream synthesize phasesrc/core/brain/link-graph/repair-lane.ts,graph-holdout.ts,o2b brain repair-lanesrc/core/brain/skill-proposal-verifier.ts,skill-proposals.tsMCP surface: 106 -> 107 tools (
brain_diarize). Every new surface is opt-in (env, config, or explicit verb) and byte-identical when unused, each with a regression test.Test plan
bun run fmtcleanbun run lintexactly 134 warnings / 0 errors (baseline)bun run typecheckcleanbun testfull foreground suite green (6605+ pass / 0 fail)bun run scripts/sync-version.ts --checkpassesSummary by CodeRabbit
o2b brain telegram-capturefor allowlisted Telegram intake with staged capture and/catchup.o2b brain inbox-drainfor dry-run-first classify-and-route with optional apply.o2b brain diarizewith stated-vs-evidenced output.o2b brain deep-synthesisJSON/MCP reports with causal context, decomposed confidence, and excluded-findings ledger.o2b brain repair-lane.brain_diarizeMCP tool and expanded MCP/JSON surfaces.