Artifacts & Continuity

The packs are the cathedral. They take a small amount of caller input, run a deterministic pipeline, and write a file you can open next week.

Where packs write

~/.ollama-intern/artifacts/
├── incident/
│   ├── 2026-04-16-5-am-paging-regression.md
│   └── 2026-04-16-5-am-paging-regression.json
├── repo/
│   ├── 2026-04-16-sprite-foundry.md
│   └── 2026-04-16-sprite-foundry.json
└── change/
    ├── 2026-04-16-pr-412.md
    └── 2026-04-16-pr-412.json

Identity is {pack, slug}. The JSON sidecar carries structured fields; the markdown carries the reader-facing layout. Renderers are code, not prompts — shape is deterministic.

Reading the artifact shape

Every pack markdown has the same anatomy:

# <Pack type> — <title>
slug: <date>-<slug>
weak: <bool> · evidence_count: <n>

## Evidence
- e1: <path>:<lines> (<hint>)
- e2: ...

## <Pack-specific body>
(e.g. Incident: symptoms, timeline, suspected cause)

## Next checks
- investigative only — never "apply this fix"

## Read next
- source paths worth opening next

If weak: true, the intern flags that evidence is thin and notes coverage gaps. This is a feature — weak evidence surfacing beats smooth confabulation.

The continuity tier

Seven tools that render without calling a model:

• artifact_list — metadata-only index. Filter by pack, date, slug_glob. Use when you want to know what’s on the desk.
• artifact_read — typed read by {pack, slug} or {json_path}. Returns both markdown text and parsed JSON.
• artifact_diff — structured same-pack comparison. Cross-pack diff is refused loudly — the payloads are shaped differently and should not be collapsed. A weak flip (was false, now true or vice versa) is surfaced prominently.
• artifact_export_to_path — handoff. Writes an existing markdown artifact, plus a provenance header (source pack, slug, timestamp), to a caller-declared allowed_roots. Refuses existing files unless overwrite: true.
• artifact_incident_note_snippet — operator-note fragment for pasting into a Slack thread or a ticket.
• artifact_onboarding_section_snippet — handbook fragment for pasting into a docs repo.
• artifact_release_note_snippet — DRAFT release-note fragment derived from change_pack.

Path safety rules

• .. is rejected before normalize — join() collapse does not save you from traversal, and we don’t rely on it.
• Export gates on allowed_roots — the caller must explicitly name the directories writes are allowed into.
• Overwrite is opt-in: overwrite: true must be set, or the call fails loud.
• Protected-path writes (memory/, .claude/, docs/canon/, etc.) require confirm_write: true on draft.

Identity precedence

{pack, slug} is primary. Paths are secondary. Collisions fail loud — the intern never silently overwrites a filing cabinet.

No model calls here

The artifact tier reads, diffs, exports, and renders snippets — all from stored content. No Ollama calls happen in this tier. This is the line between “intern producing output” and “intern filing output.” They are separate systems.

Snippet workflows

The three snippet helpers are the bridge from “there is a pack on disk” to “here is a handful of lines I can paste.” Each is deterministic — same artifact in, same snippet out. No model call, no re-render.

artifact_incident_note_snippet

You ran incident_pack on a 5 AM paging regression. Later, you need a compact markdown block to paste into a Slack thread or a ticket.

{
  "tool": "ollama_artifact_incident_note_snippet",
  "arguments": {
    "slug": "2026-04-16-sprite-pipeline-5-am-paging-regression"
  }
}

Returns:

{
  "result": {
    "markdown": "### Incident: sprite pipeline 5 AM paging regression\n- symptom: worker-3 OOM killed, ollama reports evicted=true\n- suspected cause: OLLAMA_MAX_LOADED_MODELS too high vs loaded size\n- next checks:\n  - residency.evicted across last 24h\n  - OLLAMA_MAX_LOADED_MODELS vs loaded size\n- slug: 2026-04-16-sprite-pipeline-5-am-paging-regression"
  }
}

Paste the markdown field into a ticket. The shape is fixed — reviewers who see one of these know the shape and can skim.

artifact_onboarding_section_snippet

You ran repo_pack against F:/AI/sprite-foundry/. Now you’re writing the onboarding doc for a new contributor and want a handbook-shaped section lifted from the pack.

{
  "tool": "ollama_artifact_onboarding_section_snippet",
  "arguments": { "slug": "2026-04-16-sprite-foundry" }
}

Returns a markdown fragment: ## <repo name>, a one-paragraph orientation, and a “start reading here” list with file paths. Paste into your handbook.

artifact_release_note_snippet

You ran change_pack on PR #412. Now you’re drafting the release note.

{
  "tool": "ollama_artifact_release_note_snippet",
  "arguments": { "slug": "2026-04-16-pr-412" }
}

Returns a DRAFT release-note fragment — kind, scope, summary, a bullet list of user-visible changes. Marked DRAFT because change_pack is investigative; a human should edit before shipping.

artifact_prune

New in v2.1.0. Age-based deletion of pack artifacts. Dry-run is the default — deletion requires explicit dry_run: false.

Safety example (always start with dry-run)

{
  "tool": "ollama_artifact_prune",
  "arguments": {
    "older_than_days": 90,
    "pack": "incident"
    // dry_run defaults to true
  }
}

Returns:

{
  "result": {
    "dry_run": true,
    "would_delete": [
      { "pack": "incident", "slug": "2026-01-04-vram-probe-flake",
        "age_days": 107 },
      { "pack": "incident", "slug": "2026-01-11-corpus-index-stall",
        "age_days": 100 }
    ],
    "total": 2
  }
}

Review the list. If it’s correct, re-run with dry_run: false:

{
  "tool": "ollama_artifact_prune",
  "arguments": {
    "older_than_days": 90,
    "pack": "incident",
    "dry_run": false
  }
}

Returns { deleted: [...], total: N }. The deletion is unrecoverable — rerun dry-run if you’re uncertain.

The tool honors the same path-safety rules as the rest of the artifact tier: it only operates under ~/.ollama-intern/artifacts/, .. is refused, pack-filter is required to be one of incident | repo | change. It cannot delete anything outside that tree.