context-mesh

Files (7)

• README.md
• SKILL.md
• ARCHITECTURE.md · docs/
• EXAMPLES.md · docs/
• OPERATIONS-AND-TESTING.md · docs/
• REFERENCE.md · docs/
• WORKFLOW.md · docs/

SKILL.md (first lines)

# Context Mesh

Context Mesh is the read-mostly portfolio layer above `context-offload`.

- `context-offload` owns each base's durable state and persona.
- `context-mesh` discovers those stores, indexes their relationships and timelines, adds zmem nodes, and
  produces both a human dashboard and a compact startup digest.

## Fixed locations

- Project: `C:\devdesk\Copilot\CopilotSessions\context-network`
- Generator: `C:\devdesk\Copilot\CopilotSessions\context-network\scripts\build_context_mesh.py`
- Startup digest: `C:\devdesk\Copilot\CopilotSessions\context-network\output\CONTEXT-MESH.md`
- Structured index: `C:\devdesk\Copilot\CopilotSessions\context-network\output\context-index.json`
- Dashboard: `C:\devdesk\Copilot\CopilotSessions\context-network\output\context-mesh.html`

## Access contract

1. Detect the current base by comparing the working directory with `contexts[].basePath` in
   `context-index.json`. Use the longest matching path.
2. The current base may read and write only its own `.agent-context`.
3. Every other context store is read-only. Never edit a sibling store while operating as another base.
4. While an unregistered current base is being initialized, all existing context stores are read-only.
   Create only the new current base's `.agent-context`, then refresh the mesh before normal work.
5. The generator is safe from any base: it reads stores and writes only the Context Mesh `output` directory.
6. This is an instructional boundary, not an operating-system security boundary. Never claim otherwise.

## Session hydration

At the beginning of a session:

1. Read `output\CONTEXT-MESH.md` completely.
2. Resolve the current base from `output\context-index.json`.
3. If a base matches, invoke/follow `context-offload`: read that base's `STATE.md` and `MEMORY.md`, skim
   relevant docs, and run `zmem read --query "<current topic>"`.
4. Use the mesh relationships to identify related sibling contexts.
5. Read sibling stores only when relevant; treat them strictly as reference material.
6. If no base matches, immediately follow `context-offload` to register the current working base, create its
   four-file store and persona entry, refresh Context Mesh, and then restart hydration from the new store.
7. Read the exact `Last context offload` marker in `STATE.md`. A missing, invalid, or at-least-12-hours-old
   marker is stale and requires a full offload before context-dependent work. Repeat after each 12-hour
   interval in a long-running chat.

## Refresh

Run after any `.agent-context` milestone or zmem offload:

```powershell
python C:\devdesk\Copilot\CopilotSessions\context-network\scripts\build_context_mesh.py --check
```

The build must report at least one context and must produce all three output files. If it fails, report the
failure plainly; do not leave a success-shaped fallback.

## Dashboard behavior

The HTML is dependency-free and works offline. It provides:

- Context cards with freshness, identity, focus, threads, next steps, open questions, memory and timeline.
- Search and freshness filters.
- Artifact Finder search by full path or filename. Normalize browser copy suffixes, rank original
  creation/generation above later inspection or mention evidence, and show chat name, session ID, working
  directory, timestamps, observed paths, concise supporting excerpts, and nearby prompt context.
- A conservative default graph containing context nodes and only high-confidence structural or explicitly
  declared handoff relationships.
- Optional detail levels for context-memory and zmem nodes, plus an evidence filter that can reveal
  low-confidence incidental references.
- Node and connection inspection. Every connection explanation includes its confidence, relationship class,
  and exact persisted file/line excerpt; path/name presence alone is never presented as strong evidence.
- A one-shot chronological graph animation with replay and stop behavior. Date provenance must remain honest:
  contexts use earliest dated journal evidence, zmem uses creation timestamps, inherited/unknown dates are
  labelled as such, and animation never implies a more precise creation time than the sources provide.
- Stable, distinct context-node circumference colors with zoom-independent stroke width so major contexts
  remain visually identifiable from a fitted or zoomed-out graph.
- Family-aware layout: connected context components occupy central anchors; isolated contexts and unscoped
  memory sit farther out. Pin-after-drag on a context relocates that family anchor and centers the viewport;
  unpin-all resets custom family anchors.

To open it on Windows:

```powershell
Start-Process C:\devdesk\Copilot\CopilotSessions\context-network\output\context-mesh.html
```

## Data-source honesty

- Source every claim from the indexed files or zmem export.
- Treat Artifact Finder results as evidence-ranked provenance. Label the first result “Likely origin” only
  when creation, generation, or modification evidence exists; never present a filename match as certainty.
- Read local session-state files only for indexing. Keep excerpts concise, redact common secret assignments,
  and never embed complete transcripts.
- Mark stale states as stale; never infer a fresh operational status.
- Copilot internal memory is not scraped unless a stable documented export becomes available.
- Never write credentials, secrets, tokens, or sensitive incident payloads into generated summaries.

From Zanark/skill-library, Apache-2.0.

More from Zanark/skill-library