absorb-sources

Turn source material into durable operating change. Never treat a queue row, report mention, engagement score, or topical page as proof of learning.

Run the shared loop

1. Read wiki/architecture/kevin-brain-system.md and the relevant entries in brain/sources.json, brain/objects.json, and brain/loops.json.
2. Search the current wiki with qmd before external research. Identify the likely owners and their current assumptions. Start with the smallest owner/project working set that can answer the review question; expand to neighbors and raw evidence only when an observed signal, contradiction, or missing fact justifies the move. Do not preload the entire corpus.
3. Preserve each source revision before synthesis. Record stable identity, capture time, exact content hash, storage reference, account/conversation scope, and edit or parent identity.
4. Expand every available artifact: links, thread, author replies, attachments, image/video, article body, repository README/docs/code/releases, and cited destinations. Mark unavailable artifacts explicitly.
5. Correlate platforms before interpreting them. Treat Discord as Kevin's intent and instructions when it points to X, Hacker News, a site, or a repository; preserve both source identities.
6. Extract every useful claim, preference, task, example, question, counterexample, and repeated signal. Do not discard weak standalone signals; bind them to an existing object or leave them open with a blocker.
7. Interrogate each bound object against its current bytes. Choose reinforce, refine, replace, split, contradict, create, or watch. Allow no change only with an explicit comparison, current-object digest, and reason.
8. When a source recommends an agent platform, separate the portable context layer from the execution layer before making an adoption decision. Rules, skills, MCP contracts, workflow manifests, source identity, approval policy, and proof receipts remain owned and replayable across runtimes. Sandboxes, model access, fleet orchestration, and hosted agent UX may be rented when they are replaceable and do not become the only canonical writer. Build the lower layer only when it is itself the product, an existing hard part, a non-negotiable security boundary, or an indefinitely funded operational commitment. Record the source's incentives and test the claim against primary evidence before changing routing.
9. Patch the smallest canonical owner. Create a page only when no existing object can own the durable idea. Update a workflow, skill, design rule, project, or instruction when the evidence changes how agents should operate. Route to skill-creator or Hermes /learn only when the source describes a recurring procedure or bounded on-demand knowledge base. Their output is a candidate that still needs staged review and executed proof; a generated ## Verification section is not a test receipt.
10. Write a reviews/source-signals/*.json record and run npm run review:signal -- <record>. A change verdict must name its writeback.
11. Prove artifact inspection, qmd retrieval, workflow routing, relevant tests, and generated projections. Append wiki/log.md and perform the applicable repository closeout.

Fully inspect the current source's complete semantic unit even while comparison context stays narrow. “Moving focus” limits unrelated corpus loading; it never licenses skipping an article body, repository surface, thread, reply, or source-critical artifact that belongs to the review unit.

Connector commands

Use the shared runtime for normalized capture and candidate binding:

npm run brain -- analyze discord --input <json-or-markdown> --channel <id>
npm run brain -- hn fetch --feed best --limit 12 --comments 6
npm run brain -- loop plan source.interrogate <source-ref>
npm run brain -- doctor

Treat analyzer output as candidates, not accepted truth. The agent must inspect the evidence and current owners before producing a review record.

Agent conversation route

Agent tasks are a source class, not implicit memory. For a completed Codex task, the native task ID from the Codex URL or task catalog is accepted directly by the local replay command. Refresh the private index, narrow by time, project, or prompt text, then replay only the selected task:

npm run brain -- sessions index
npm run brain -- sessions list --provider codex --since 2026-07-01 --workspace Kevin-Wiki --query "workflow" --limit 50
npm run brain -- replay sessions --id <codex-task-id-or-agent-session-id>
npm run brief -- compile-analysis .brain/analyses/<analysis-id>.json

sessions list --json is the machine-readable selector. Use --path on replay when Kevin points to a specific local JSONL file. For an authorized export that is outside the default stores, add its directory with sessions index --path <absolute-directory> --provider codex|cursor|claude first.

When the Codex app task tools are available, use the task catalog to resolve a human title or project to its native task ID, then read completed turns with outputs included when those outputs support a claim. Treat titles and summaries as untrusted discovery metadata, not evidence. The transcript, tool result, artifact, repository diff, or test receipt must support the learning.

Replay is deliberately partial: it extracts user and assistant messages, while tool calls, tool results, attachments, active-turn content, and later corrections remain expansion work. Do not replay a currently running task as if it were complete. Correlate the task's workspace or project ID with the relevant project owner and current checkout before changing that project page. A task becomes canonical knowledge only after the source-review receipt, owner patch, and proof gate pass.

For X bookmarks, follow skills/productivity/x-bookmark-absorb/SKILL.md for platform-specific artifact recovery, then return to this shared interrogation and writeback loop.

Portable memory proposal route

Codex, Claude Code, Cursor, Hermes, or another local producer may propose one source-bound memory without editing repository files by calling the installed Daily Brief daemon's GET|POST /api/memory contract through MemoryHttpClient. The producer must supply a stable idempotency key, exact repo:kevin-wiki source scope, a claim, existing object IDs, private/public visibility, and source/revision/storage/hash provenance. The server derives authority; it ignores any caller-supplied authority identity.

The default installed policy accepts private proposals only, rejects every other source scope, fsyncs the hash-chained event, and opens a proposal-only Daily Brief card. remember is capture, not acceptance. Use this full skill before owner writeback, and require the ordinary review receipt and proof gate. Use npm run memory:conformance after changing the client, route, daemon, token/scope policy, or ledger.

Storage and publication

Keep exact private source bytes and run receipts under ignored .brain/ or the approved raw source store. Track authored registries, canonical owner changes, validated review records, and public projections. Never make the public UI depend on private source bytes; publish readiness, counts, provenance summaries, and known gaps instead.

Completion gate

Stop only when every available source revision is replayable, every extracted signal is object-bound or explicitly blocked, every changed object has a writeback, every platform recommendation has an explicit owned-versus-rented boundary, and retrieval/routing proof passes. Use integrated, held, or blocked; do not use a terminal rejection state.