Copilot History Ingest

Copilot History Ingest — Conversation Mining

You are extracting knowledge from the user's past GitHub Copilot CLI conversations and distilling it into the Obsidian wiki. Conversations are rich but messy — your job is to find the signal and compile it.

This skill can be invoked directly or via the wiki-history-ingest router (/wiki-history-ingest copilot).

Before You Start

1. Resolve config — follow the Config Resolution Protocol in llm-wiki/SKILL.md (walk up CWD for .env → ~/.obsidian-wiki/config → prompt setup). This gives OBSIDIAN_VAULT_PATH, COPILOT_HISTORY_PATH (defaults to ~/.copilot/session-state), and COPILOT_VSCODE_STORAGE_PATH (VS Code workspaceStorage; platform-specific — ask the user if absent) 2. Read .manifest.json at the vault root to check what's already been ingested 3. Read index.md at the vault root to know what the wiki already contains

Ingest Modes

Append Mode (default)

Check .manifest.json for each source file (events JSONL, transcript JSONL, checkpoint, session-store DB). Only process:

• Sessions not in the manifest (new sessions)
• Sessions whose updated_at is newer than their ingested_at in the manifest

This is usually what you want — the user ran a few new sessions and wants to capture the delta.

Full Mode

Process everything regardless of manifest. Use after a wiki-rebuild or if the user explicitly asks.

GitHub Copilot Data Layout

Copilot stores data in three locations. Scan all three.

Source 1: ~/.copilot/session-state/ (CLI sessions)

~/.copilot/session-state/
├── <session-uuid>/
│   ├── workspace.yaml           # Session metadata (id, cwd, summary_count, created_at, updated_at)
│   ├── vscode.metadata.json     # VS Code context (workspaceFolder, repositoryProperties, customTitle)
│   ├── events.jsonl             # Full event log — all turns, tool calls, reasoning
│   ├── session.db               # Per-session SQLite (todos/todo_deps only — skip for ingestion)
│   ├── index.md                 # Session summary written at session end
│   ├── checkpoints/             # Checkpoint JSON files (mid-session summaries)
│   │   └── <uuid>.json          # title, overview, history, work_done, technical_details,
│   │                            #   important_files, next_steps
│   ├── files/                   # Artifacts produced during session (plans, diagrams, etc.)
│   └── research/                # Research artifacts
└── ...

Source 2: ~/.copilot/session-store.db (Global SQLite)

The canonical cross-session database. This is the highest-value source: structured, queryable, and pre-summarised.

sessions       — id, cwd, repository, branch, summary, created_at, updated_at, host_type
turns          — session_id, turn_index, user_message, assistant_response, timestamp
checkpoints    — session_id, checkpoint_number, title, overview, history, work_done,
                 technical_details, important_files, next_steps, created_at
session_files  — session_id, file_path, tool_name, turn_index, first_seen_at
session_refs   — session_id, ref_type (commit/pr/issue), ref_value, turn_index, created_at
search_index   — FTS5 virtual table (content, session_id, source_type, source_id)

Source 3: VS Code Workspace Storage (<workspaceStorage>/<hash>/GitHub.copilot-chat/)

VS Code extension data, keyed by workspace hash. The path is platform-specific and must come from .env or user input.

<hash>/GitHub.copilot-chat/
├── transcripts/
│   └── <session-uuid>.jsonl     # Conversation transcripts (same JSONL format as events.jsonl)
├── memory-tool/
│   └── memories/
│       └── <base64-session-id>/ # Per-session saved artifacts (plan.md, etc.)
│           └── plan.md
└── codebase-external.sqlite     # Codebase index (skip — no conversation knowledge)

Key data sources ranked by value:

1. Checkpoints (session-store.db checkpoints table + per-session checkpoints/*.json) — Pre-distilled summaries with overview, work_done, technical_details, important_files, next_steps. Gold. 2. Session summaries (session-store.db sessions.summary + index.md) — One-paragraph synopsis per session. 3. Turns (session-store.db turns table + events.jsonl / transcript JSONL) — Full conversation. Rich but verbose. 4. Memory artifacts (memory-tool/memories/<id>/plan.md etc.) — Pre-written plans and structured notes the user saved explicitly. Worth importing verbatim (or lightly summarised). 5. File access patterns (session_files table + tool.execution_* events) — Which files the agent repeatedly touched — reveals high-value project files. 6. Session refs (session_refs table) — Commits, PRs, and issues linked to sessions. 7. `vscode.metadata.json` — Workspace folder path, branch, customTitle (user-set session label). Useful for grouping and naming.

Step 1: Survey and Compute Delta

Scan all three data locations and compare against .manifest.json:

# --- Source 1: per-session directories ---
# Find all session directories (each has workspace.yaml)
ls ~/.copilot/session-state/

# For each session, read workspace.yaml for id/cwd/updated_at
# and vscode.metadata.json for customTitle / repositoryProperties

# --- Source 2: global database ---
# Query session-store.db with sqlite3 (or Python sqlite3)
SELECT s.id, s.cwd, s.repository, s.branch, s.summary, s.updated_at,
       COUNT(DISTINCT t.turn_index) AS turn_count,
       COUNT(DISTINCT c.id)         AS checkpoint_count
FROM sessions s
LEFT JOIN turns t ON t.session_id = s.id
LEFT JOIN checkpoints c ON c.session_id = s.id
GROUP BY s.id
ORDER BY s.updated_at DESC;

# --- Source 3: VS Code workspace storage ---
# For each <hash> directory under workspaceStorage, check for GitHub.copilot-chat/
# Find transcript files
ls <workspaceStorage>/<hash>/GitHub.copilot-chat/transcripts/

Build a unified inventory — one entry per session UUID — and classify:

• New — not in manifest → needs ingesting
• Modified — in manifest but updated_at is newer → needs re-ingesting
• Unchanged — in manifest and not modified → skip in append mode

Report to the user: "Found X sessions in session-state, Y in session-store.db, Z VS Code transcript files. Checkpoints: A. Delta: B new, C modified."

Step 2: Ingest Checkpoints and Summaries First

Checkpoints are already distilled — process them before touching raw turns.

From session-store.db:

SELECT s.id, s.cwd, s.repository, s.branch, s.summary,
       c.checkpoint_number, c.title, c.overview, c.work_done,
       c.technical_details, c.important_files, c.next_steps,
       c.created_at
FROM checkpoints c
JOIN sessions s ON c.session_id = s.id
ORDER BY s.updated_at DESC, c.checkpoint_number ASC;

From per-session checkpoints/*.json:

Each checkpoint file has: title, overview, history, work_done, technical_details, important_files, next_steps.

Read index.md (if present) as a session-level summary — it's typically written at session end and is already concise.

What to extract:

• overview → high-level description of what the session accomplished
• work_done → concrete tasks completed (good for skills / project pages)
• technical_details → implementation specifics (good for concepts pages)
• important_files → high-value files in the project (good for project pages)
• next_steps → open threads (good for linking to ongoing project work)

Step 3: Parse Session Turns

Read turns from session-store.db (preferred — already parsed) or from events.jsonl / transcript JSONL.

From session-store.db:

SELECT turn_index, user_message, assistant_response, timestamp
FROM turns
WHERE session_id = '<uuid>'
ORDER BY turn_index ASC;

From events.jsonl / transcript JSONL:

Each file is one session. Each line is a JSON event. See references/copilot-data-format.md for the full schema.

Relevant event types:

Extraction strategy for `assistant.message`:

• data.content is the assistant's text response — extract this
• data.reasoningText is internal reasoning — skip (it's the unpacked reasoningOpaque field)
• data.toolRequests lists tool calls — skim tool names and arguments for file access patterns
• Skip type: "tool.execution_end" entirely

Step 3b: Process Memory Artifacts

For each session that has a memory-tool/memories/<base64-id>/ directory in VS Code workspace storage, read any markdown files saved there (typically plan.md). These are documents the user explicitly saved — treat them as high-quality, user-authored content.

Decode the base64 directory name to get the session UUID:

import base64
session_id = base64.b64decode(dir_name).decode('utf-8')

Memory artifacts map to project skills/ or concepts/ pages, depending on content type.

Step 3c: Extract File and Ref Patterns

From session-store.db:

-- Most-touched files per project
SELECT repository, file_path, COUNT(*) AS touch_count
FROM session_files
GROUP BY repository, file_path
ORDER BY touch_count DESC;

-- Linked commits/PRs/issues per session
SELECT session_id, ref_type, ref_value, turn_index
FROM session_refs
ORDER BY session_id, turn_index;

File access patterns reveal which files are architecturally important — note them on project pages.

Session refs link Copilot sessions to git history — useful for connecting wiki knowledge to concrete code changes.

Step 4: Cluster by Topic

Don't create one wiki page per session. Instead:

• Group extracted knowledge by topic across sessions
• A single session about "debugging auth + setting up CI" → two separate topics
• Three sessions across different days about "React performance" → one merged topic
• cwd / repository give you a natural first-level grouping; vscode.metadata.json's customTitle gives a human-readable session label

Step 5: Distill into Wiki Pages

Each Copilot project maps to a project directory in the vault. Derive the project name from cwd or repository:

C:\Users\name\git\my-project   → my-project
/Users/name/code/another-app   → another-app

Prefer repository (e.g., owner/repo) from session-store.db over raw cwd when available.

Project-specific vs. global knowledge

For each project with content, create or update the project overview page at projects/<name>/<name>.md — named after the project, not `_project.md`. Obsidian's graph view uses the filename as the node label, so _project.md makes every project show up as _project in the graph. Naming it <name>.md gives each project a distinct, readable node name.

Important: Distill the _knowledge_, not the conversation. Don't write "In a session on March 15, the user asked about X." Write the knowledge itself, with the session as a source attribution.

Write a `summary:` frontmatter field on every new/updated page — 1–2 sentences, ≤200 chars, answering "what is this page about?" for a reader who hasn't opened it. wiki-query's cheap retrieval path reads this field to avoid opening page bodies.

Add confidence and lifecycle fields to every new page's frontmatter:

base_confidence: 0.42
lifecycle: draft
lifecycle_changed: <ISO date today>

Leave lifecycle unchanged on update.

Mark provenance per the convention in llm-wiki (Provenance Markers section):

• Checkpoints and index.md are pre-distilled by the system — treat checkpoint-derived claims as extracted (the system wrote them from observed actions).
• Memory artifacts are user-authored — treat as extracted.
• Conversation turn distillation is mostly inferred. You're synthesizing a coherent claim from many turns. Apply ^[inferred] liberally to synthesized patterns, generalizations across sessions, and "what the user really meant" interpretations.
• Use ^[ambiguous] when the user changed direction mid-session or when the session ended unresolved.
• Write a provenance: frontmatter block on every new/updated page summarizing the rough mix.

Step 6: Update Manifest, Journal, and Special Files

Update .manifest.json

For each session processed, add/update its entry with:

• ingested_at, session_id, updated_at
• source_type: one of "copilot_session", "copilot_checkpoint", "copilot_transcript", "copilot_memory_artifact"
• project: the decoded project name
• pages_created and pages_updated lists

Also update the projects section of the manifest:

{
  "project-name": {
    "repository": "owner/repo",
    "cwd": "C:\\Users\\name\\git\\project-name",
    "vault_path": "projects/project-name",
    "last_ingested": "TIMESTAMP",
    "sessions_ingested": 5,
    "sessions_total": 8,
    "checkpoints_ingested": 12,
    "memory_artifacts_ingested": 3
  }
}

Create journal entry + update special files

Update index.md and log.md per the standard process:

- [TIMESTAMP] COPILOT_HISTORY_INGEST projects=N sessions=M checkpoints=C pages_updated=X pages_created=Y mode=append|full

`hot.md` — Read $OBSIDIAN_VAULT_PATH/hot.md (create from the template in wiki-ingest if missing). Update Recent Activity with a one-line summary — e.g. "Ingested 5 Copilot sessions across 2 projects; surfaced patterns in API design and testing strategy." Keep the last 3 operations. Update Active Threads if any ongoing project is now better understood. Update updated timestamp.

Privacy

• Distill and synthesize — don't copy raw conversation text verbatim
• Skip anything that looks like secrets, API keys, passwords, tokens
• data.reasoningOpaque / data.reasoningText in assistant events is internal reasoning — skip entirely, never copy to wiki
• If you encounter personal/sensitive content, ask the user before including it
• The user's conversations may reference other people — be thoughtful about what goes in the wiki

Reference

See references/copilot-data-format.md for detailed data structure documentation.

QMD Refresh After Vault Writes

QMD is a search index, not the source of truth. If $QMD_WIKI_COLLECTION is empty or unset, skip this step. Run it only after this skill has written or rewritten vault markdown. If QMD refresh fails, do not roll back the vault changes; report the QMD status separately.

Use $QMD_CLI if set; otherwise use qmd.

${QMD_CLI:-qmd} update

If the output says vectors are needed or embeddings may be stale, run:

${QMD_CLI:-qmd} embed

Verify the collection with either:

${QMD_CLI:-qmd} ls "$QMD_WIKI_COLLECTION"

or, when a specific page path is known:

${QMD_CLI:-qmd} get "qmd://$QMD_WIKI_COLLECTION/<page>.md" -l 5

Record one of:

• QMD refreshed: update + embed + verified
• QMD refreshed: update only + verified
• QMD skipped: QMD_WIKI_COLLECTION unset
• QMD skipped: qmd CLI unavailable
• QMD failed: <short error summary>

GitHub Copilot CLI Data Format — Detailed Reference

Session-State Directory

~/.copilot/session-state/ contains one directory per session the user has run with GitHub Copilot CLI. Each directory is named with a UUID.

workspace.yaml

Minimal session metadata file, always present:

id: <session-uuid>
cwd: /path/to/project
summary_count: 3
created_at: 2026-04-02T14:28:13.304Z
updated_at: 2026-04-29T12:00:00.000Z

summary_count reflects how many checkpoints were written. Sessions with summary_count: 0 were either very short or completed without checkpointing — check events.jsonl for content anyway.

vscode.metadata.json

VS Code context, written when the session is associated with a VS Code workspace:

{
  "workspaceFolder": {
    "folderPath": "c:\\Users\\name\\git\\my-project",
    "timestamp": 1773245818098
  },
  "writtenToDisc": true,
  "repositoryProperties": {
    "repositoryPath": "c:\\Users\\name\\git\\my-project",
    "branchName": "feature/my-branch",
    "baseBranchName": "origin/main"
  },
  "customTitle": "User-written session label or system-set title"
}

customTitle is the most human-readable session label — use it as a heading when creating session-derived wiki content. May be absent on older sessions.

events.jsonl

The full event log for one session. Each line is a JSON object representing one event in the session.

Event: session.start

{
  "type": "session.start",
  "data": {
    "sessionId": "09371a50-9a50-484a-8743-5c696de1623a",
    "version": 1,
    "producer": "copilot-agent",
    "copilotVersion": "0.0.420",
    "startTime": "2026-03-02T15:10:04.678Z",
    "context": {
      "cwd": "C:\\Users\\name\\git\\my-project",
      "gitRoot": "C:\\Users\\name\\git\\my-project",
      "branch": "master"
    }
  },
  "id": "<event-uuid>",
  "timestamp": "2026-03-02T15:10:04.817Z",
  "parentId": null
}

data.context.cwd and data.context.branch establish the project context. Always read session.start first.

Event: user.message

{
  "type": "user.message",
  "data": {
    "content": "review my staged but uncommitted changes for issues",
    "transformedContent": "<current_datetime>...</current_datetime>\n\nreview my staged...",
    "attachments": [],
    "interactionId": "9352571e-a0b9-4774-8ecb-40bc58f86e94"
  },
  "id": "<event-uuid>",
  "timestamp": "2026-03-02T15:10:45.058Z",
  "parentId": "<parent-event-uuid>"
}

Use data.content (not data.transformedContent) — the transformed version includes injected system context that's noise for wiki purposes.

Event: assistant.message

{
  "type": "assistant.message",
  "data": {
    "messageId": "<uuid>",
    "content": "I'll review the staged changes in those three files.",
    "toolRequests": [
      {
        "toolCallId": "tooluse_...",
        "name": "report_intent",
        "arguments": { "intent": "Reviewing staged changes" },
        "type": "function"
      },
      {
        "toolCallId": "tooluse_...",
        "name": "powershell",
        "arguments": {
          "command": "git --no-pager diff --cached --stat",
          "description": "Show staged diff"
        },
        "type": "function"
      }
    ],
    "interactionId": "9352571e-a0b9-4774-8ecb-40bc58f86e94",
    "reasoningOpaque": "<base64-encrypted-reasoning>",
    "reasoningText": "The user wants me to review staged git changes..."
  },
  "id": "<event-uuid>",
  "timestamp": "2026-03-02T15:10:50.235Z",
  "parentId": "<parent-event-uuid>"
}

Extraction strategy:

• Extract data.content — the assistant's visible text response
• data.toolRequests — skim tool names and description arguments for file/command patterns; ignore report_intent calls
• Skip `data.reasoningOpaque` entirely — encrypted/encoded internal reasoning
• Skip `data.reasoningText` entirely — decrypted reasoning; internal only, never user-visible

Event: assistant.turn_start

{
  "type": "assistant.turn_start",
  "data": { "turnId": "0", "interactionId": "..." },
  "id": "...",
  "timestamp": "..."
}

Marks the start of an assistant turn. Useful for turn boundary detection; no content to extract.

Event: tool.execution_start

{
  "type": "tool.execution_start",
  "data": {
    "toolCallId": "tooluse_...",
    "toolName": "powershell",
    "arguments": { "command": "dotnet build ...", "description": "Build project" }
  },
  "id": "...",
  "timestamp": "..."
}

Reveals what tools (file reads, commands, searches) were invoked. File-related tools (view, edit, create) with their paths are worth noting for the session_files equivalent when reading events directly.

Event: tool.execution_end

Contains the raw tool output. Usually noise — skip unless diagnosing errors.

checkpoints/<uuid>.json

Mid-session progress summaries, written automatically as the session progresses:

{
  "title": "Implementing auth module",
  "overview": "Working on JWT authentication for the API...",
  "history": "1. Analyzed existing auth code\n2. Created IAuthService...",
  "work_done": "- Created IAuthService interface\n- Implemented JwtAuthService",
  "technical_details": "Uses RS256 signing. Token expiry configurable via settings...",
  "important_files": "- src/Auth/IAuthService.cs\n- src/Auth/JwtAuthService.cs",
  "next_steps": "- Wire up to DI container\n- Add refresh token support"
}

This is the highest-value structured content in the per-session directory — equivalent to Claude's memory files.

index.md

Session-end summary written as a markdown file. Typically 1–3 paragraphs summarizing what was accomplished. Content varies by session length and complexity. Read this before opening events.jsonl to decide if the session is worth deep-processing.

---

Global Session Store (session-store.db)

SQLite database at ~/.copilot/session-store.db. The canonical cross-session record.

Schema

sessions

turns

Note: user_message here is the pre-transformation content — use this, not transformedContent from events.jsonl.

checkpoints

session_files

⚠️ No id column — use COUNT(DISTINCT sf.file_path) not COUNT(DISTINCT sf.id).

Aggregate by file_path across sessions to identify architecturally important files.

session_refs

search_index (FTS5)

Full-text search index. Use for keyword discovery when surveying a large history:

SELECT content, session_id, source_type
FROM search_index
WHERE search_index MATCH 'auth OR authentication OR login'
LIMIT 20;

source_type values: "turn", "checkpoint_overview", "checkpoint_history", "checkpoint_work_done", "checkpoint_technical", "checkpoint_files", "checkpoint_next_steps", "workspace_artifact".

---

VS Code Workspace Storage

Location

The workspaceStorage directory is platform-specific:

Each <hash>/ subdirectory corresponds to a specific workspace (VS Code folder). The hash is derived from the workspace path — there is no human-readable mapping, so enumerate all <hash>/GitHub.copilot-chat/ directories and use the transcripts/ JSONL files' session.start events to identify which project each belongs to.

Transcript JSONL (transcripts/<uuid>.jsonl)

Identical format to events.jsonl from Source 1. Parse using the same event type handlers. The session.start event's data.context.cwd tells you which project this belongs to.

Memory Artifacts (memory-tool/memories/<base64-session-id>/)

Directory name is the session UUID encoded as base64. Files inside are markdown documents explicitly saved by the user or system during the session — typically plan.md containing the session plan.

Decode the directory name to link it to a session:

import base64
# Pad to multiple of 4 before decoding
session_id = base64.b64decode(dir_name + '==').decode('utf-8')

---

Processing Order

For maximum efficiency and signal-to-noise:

1. `session-store.db` checkpoints — Fastest, highest signal. Query all at once. 2. `session-store.db` sessions.summary — One-paragraph synopsis per session. 3. *Per-session `checkpoints/.json + index.md** — For sessions not yet in session-store.db or for additional detail. 4. **Memory artifacts** (memory-tool/memories/) — User-authored, high quality. 5. **session-store.db turns** — Full conversation, process selectively by topic. 6. **events.jsonl / transcript JSONL** — Only if session-store.db is absent or incomplete. 7. **session_files / session_refs`** — For file pattern and git linkage metadata.