context-collection

name	context-collection
description	Land raw project information — meeting notes, existing specs, domain research, external tool exports, interview answers — into docs/context/ with provenance, so later phases and stateless agents can use it. Use this at project kickoff, whenever new source material appears (a document, a decision made in chat, findings from web/MCP research), and before any distillation work that lacks raw inputs.

Context Collection

Collection has one job: get information out of volatile places (chats, heads, external tools) and into the repository, faithfully. Completeness beats tidiness here — the distillation phase (see context-distillation) will filter. What collection must never do is silently interpret: a paraphrase that changes meaning is worse than a messy verbatim note, because downstream agents will trust it.

Landing zone

• Everything lands under docs/context/<topic>/, one topic per directory (e.g., docs/context/device-pairing/, docs/context/line-integration/).
• Each topic directory keeps an INDEX.md: one line per file — what it is and why it matters. Agents read indexes first; keep them current.
• Large binaries and originals stay in their system of record; store a link and an extracted-text or summary file here instead.

Provenance header (mandatory)

Every collected file starts with:

---
source: <URL, meeting name, person, or tool>
retrieved: <YYYY-MM-DD>
method: <verbatim | export | interview | ai-summary | web-research>
collector: <human name or agent/session identifier>
sensitivity: <public | internal | confidential>
status: raw
---

The method field is what lets a future reader calibrate trust: verbatim/export can be quoted as fact; ai-summary and web-research must be re-verified before becoming an agreement.

Collection rules

1. Raw over summarized. Prefer the original wording; if you must condense, keep the original alongside or linked. Mark interpretation as such.
2. Chat is a source too. When a real decision or requirement surfaces in a conversation, capture it here (method: verbatim, source: the conversation) — otherwise it evaporates when the session ends.
3. External systems via MCP. Pull from external tools (docs, tickets, designs) through the MCP servers configured for this repo (.mcp.json for Claude Code sessions; repository settings for the cloud agent). Record which server/tool produced the data in source.
4. No secrets, ever. Redact tokens, credentials, and personal data before a file lands. If sensitivity is confidential, confirm with a human that the repository is an acceptable home before committing.
5. Contradictions are content. When two sources disagree, collect both and note the conflict in INDEX.md — resolving it is a distillation/agreement decision, not a collection edit.

Definition of done for a collection pass

• New material sits under the right topic with a complete provenance header.
• INDEX.md updated.
• Conflicts and open questions listed at the top of INDEX.md.
• Nothing in the pass exists only in a chat window.