/mine-for-context

name: mine-for-context
description: "Deep context extraction from source material. Creates reference bundles, builds extraction plans, tracks what's been extracted, and discovers new targets — people, subjects, patterns, connections. The archaeologist that turns raw sources into structured knowledge. Can be invoked by alive:session-history for targeted session mining."
user-invocable: true

Mine

Deep extraction. Turn raw sources into structured context.

Not a quick search (that's find). Not session timeline (that's history). Mine is the heavy operation — reading source material systematically, extracting everything of value, and routing it into the world.

---

When It Fires

• The human says "mine this", "extract from these", "what's buried in here"
• `alive:session-history` escalates — "these sessions look rich, want to deep-mine them?"
• The human points at a folder of transcripts, documents, or exports
• A bundle has raw sources that haven't been fully processed
• The human asks "what have I missed?" or "find everything about X in my sources"

---

What It Does

0. Pre-check: Source Location

Before mining, check whether the source file lives inside a bundle's `raw/` folder:

• **In a bundle** — proceed normally. Phase 4 tracking writes to that bundle's `context.manifest.yaml` `discovered:` field.
• **Loose file** (inbox, walnut root, or anywhere without a manifest) — enter **loose-mine mode**. Mining still works, but tracking state writes to `$WORLD_ROOT/.alive/_mining-log.json` instead of a bundle manifest. This prevents duplicate work if the same file is mined again before being captured into a bundle.

// .alive/_mining-log.json — created on first loose mine
{
  "mined": [
    {
      "path": "03_Inbox/Will x Ben Standup.md",
      "mined_at": "2026-04-16T10:30:00Z",
      "items_extracted": 14,
      "routed_to": "alive-os"
    }
  ]
}

If the file appears in `_mining-log.json` already, warn: "This file was mined on {date} ({N} items extracted). Mine again or skip?"

1. Assess Source Material

Scan what's available. Don't read everything — just understand the landscape.

> **squirrel scanning sources**
> client-calls/raw/ — 12 transcripts (3 unprocessed)
> research/raw/ — 8 documents (5 unprocessed)
> .alive/_squirrels/ — 23 sessions (8 with rich stash, 4 unmined transcripts)
>
> Total: 43 sources, 12 unprocessed, 4 unmined sessions

2. Build Extraction Plan

Before touching any content, propose a plan. What to look for, in what order, expected yield.

> **squirrel extraction plan**
> Source: client-calls/raw/ (12 transcripts)
>
> Targets:
> - People mentioned (create/update person walnuts)
> - Decisions made (route to walnut logs)
> - Tasks assigned (route to walnut tasks)
> - Domain knowledge (route to insights)
> - Recurring themes (flag as potential bundles)
>
> > Run this plan?
> 1. Yeah, mine it
> 2. Adjust targets
> 3. Mine specific files only

The plan adapts to what the human needs:

• "Mine the last 3 squirrels" -> session-focused extraction
• "Mine everything about glass-cathedral from February" -> topic-focused, date-filtered
• "Who keeps coming up that I haven't tracked?" -> people discovery mode
• "What patterns do you see?" -> theme extraction

3. Extract Systematically

Process sources one at a time or in batches. For each source:

• **Read** the full content (transcript, document, export)
• **Extract** against the plan targets
• **Route** extracted items:
• People -> stash for person walnut creation/update
• Decisions -> stash for log routing at save
• Tasks -> stash for task routing at save
• Knowledge -> stash as insight candidates
• Themes -> flag as potential new bundles
• **Update context.manifest.yaml** -> mark the source as processed in the bundle's `context.manifest.yaml` `discovered:` field

Extraction is bounded by source type:

| Source Type | Extract | |------------|---------| | Transcript | Decisions, action items, people + roles, key quotes, domain knowledge, commitments, deadlines | | Document | Key claims, data points, relevant sections, author context, links to other work | | Session transcript | Decisions + rationale, files touched, architectural choices, dead ends, open threads | | Export (ChatGPT, etc.) | Topics discussed, decisions made, knowledge synthesized, people referenced | | Email thread | Commitments, deadlines, people + relationships, action items, context updates |

4. Track Progress

The bundle's `context.manifest.yaml` tracks extraction state:

discovered:
  status: partial          # none | partial | complete
  last_mined: 2026-03-10
  processed:
    - path: raw/2026-02-15-call-with-jax.md
      extracted: [3 people, 2 decisions, 1 insight]
    - path: raw/2026-02-20-shielding-review.md
      extracted: [1 person, 4 decisions, 2 tasks]
  unprocessed:
    - raw/2026-03-01-vendor-followup.md
    - raw/2026-03-05-budget-review.md

This means the squirrel can resume mining across sessions — it knows exactly what's been done and what remains.

5. Discover Targets

The most valuable part. While extracting, actively watch for:

• **New people** — names that appear across multiple sources but have no person walnut. Surface them with context about who they are and how they relate.
• **Recurring subjects** — topics that keep coming up across sources. These might deserve their own bundle or even walnut.
• **Interests and tendencies** — patterns in what the human focuses on, returns to, or avoids. Not for judgment — for awareness.
• **Cross-walnut connections** — references in one walnut's sources to another walnut's domain. These are invisible links the system should know about.
• **Contradictions** — decisions in one source that conflict with decisions in another. Surface these gently. **Detection procedure:** for each extracted decision, grep the target walnut's `_kernel/insights.md` and last 10 entries of `_kernel/log.md` for the topic keyword. If