bring-context

name

bring-context

description

Pull context from a previous entire-tracked session into the current chat. Lists sessions in the current repo (or any path passed as arg), lets the user pick when more than one matches, writes a 6-section compaction summary to entire_context_<id>.md in cwd, and feeds the same summary into chat. Use when the user says "bring context", "import context", "pull from session", or invokes /bring-context.

Bring Context

Pull context from a previous entire-tracked session into this chat as both a persistent file and live agent context.

STOP — Rules

1. The picker is interactive. When 2+ candidate sessions exist, list them, then STOP your turn and wait for the user to pick.
2. Read transcripts only after the pick. Listing reads only small JSON metadata; only the picked session's full transcript is parsed.
3. Never auto-answer Unanswered Questions. If the imported session ended on an open question, present it and halt.
4. Do NOT invoke the entire CLI. Read JSON files and transcripts directly. The skill works without entire on PATH.
5. Never modify the source session's repo. Only write to the current cwd.

Inputs

• /bring-context — sessions in <cwd>/.git/entire-sessions/
• /bring-context <repo-path> — sessions in <repo-path>/.git/entire-sessions/
• /bring-context <session.json> — skip picker, use this exact session
• /bring-context <hex-prefix> — match unique prefix in cwd's session list

Flow

1. Resolve source

• No arg → source = <cwd>/.git/entire-sessions/
• Arg is a .json file (path exists, ends in .json) → use directly, jump to step 4
• Arg is a directory → source = <arg>/.git/entire-sessions/
• Arg matches ^[a-f0-9]+$ (and is not a path) → search <cwd>/.git/entire-sessions/ for a unique prefix match; if 1 match, use it; if 0 or 2+, halt with the matches printed
• If source doesn't exist → halt: No entire sessions at <path>.

2. Identify the active session (to exclude)

Glob <cwd>/.git/entire-sessions/*.json. The one with phase == "active" is the currently running session. Record its session_id. If none found, exclude nothing.

3. List candidates

Glob <source>/*.json. For each, parse and read only:

• session_id, agent_type, model_name, last_prompt, last_interaction_time, files_touched, base_commit, transcript_path

Filter out the active session_id. Sort by last_interaction_time DESC.

Branch on count:

• 0 → halt: No importable sessions in <source>.
• 1 → auto-pick, jump to step 4
• 2+ → print numbered list (format below), STOP your turn, wait for the user to reply with a number or session ID

Found <N> sessions in <source>
(excluding current session <short-id>)

  1. <3hex>  <agent_type>  <model_name>  <relative time>  "<last_prompt[:80]>"  (<N> files)
  2. <3hex>  ...
  ...

Which one? (1-N, or paste a session ID)

If 20+ sessions, show only the 20 most recent and add: Showing 20 of <total> — pass a session ID to pick directly.

4. Read transcript

From the picked session's JSON, get transcript_path. Extract the relevant lines:

grep -E '"type":"(message|function_call|user|assistant)"' "<transcript_path>" | cut -c1-2000

If the output exceeds ~500 lines, fall back to head + tail to capture original task and final state:

grep -E '"type":"(message|function_call|user|assistant)"' "<transcript_path>" | tail -100 | cut -c1-2000
grep -E '"type":"(message|function_call|user|assistant)"' "<transcript_path>" | head -20 | cut -c1-2000

If the transcript file is missing, halt: Transcript missing at <path> — likely recorded on a different machine.

Do NOT show the raw grep output to the user. Process it internally to compose the summary in step 6.

5. Stale-context check

The check must run against the source repo, not cwd. The source repo path comes from the session JSON's worktree_path field (when importing from a different repo, this points there; when importing from the current repo, it equals cwd).

git -C "<worktree_path>" rev-parse HEAD

Compare to base_commit from the picked session JSON. If they differ at all (even by one commit), record this for inclusion in the file and chat output.

If worktree_path is missing or unreadable (e.g. moved repo), skip the check rather than guessing.

6. Compose the summary

From the grep output (step 4) plus the session JSON metadata, produce 6 sections. Be concise but complete — err toward keeping detail that prevents duplicate work or repeated mistakes.

1. Task Overview — user's core request, success criteria, stated constraints
2. Current State — completed work, files modified, decisions made, artifacts produced
3. Important Discoveries — technical constraints, rationale, errors hit and resolutions, failed approaches and why
4. Next Steps — specific remaining actions, blockers, priority ordering
5. Context to Preserve — user preferences, domain details, commitments
6. Unanswered Question — only if the last user-assistant exchange ended on a question the user never answered. Capture exactly as asked.

7. Build the filename

short_id = first 3 hex chars of session_id.

filename = <cwd>/entire_context_<short_id>.md.

If it exists, version-bump: try entire_context_<short_id>_v2.md, _v3.md, ... until a free name is found. Never overwrite.

8. Write the file

> Imported from session `<full session_id>` (<agent_type> · <model_name>) — repo `<basename of worktree_path>` @ commit `<base_commit[:7]>` — <ISO timestamp>
> Source transcript: `<transcript_path>`

# Context from session <short_id>

[If stale: > ⚠️ HEAD has moved since this session (base `<base[:7]>`, current `<HEAD[:7]>`). File references may be out of date.]

## Task Overview
...

## Current State
...

## Important Discoveries
...

## Next Steps
...

## Context to Preserve
...

## Unanswered Question
(only if applicable)

---

*Files touched: <comma-separated, from session JSON's `files_touched` field>*
*Reconstruct: `entire explain --session <full session_id>`*

9. Print summary to chat

Output the same 6-section summary in chat (so the agent has it in live context) and end with: Wrote <filename>.

If stale, prefix with the same ⚠️ warning line.

10. Continue

• If summary contains an Unanswered Question → present it to the user and halt. The user is the decision-maker; do not pick a default.
• Otherwise → start the next step from the summary. Do not ask for permission.

Notes

• 3-char short IDs collide easily (4096 possibilities). Version-bump guarantees the file is never overwritten; the full UUID in the header disambiguates between collisions.
• The skill's only side effect is writing one .md file in cwd. It never modifies git, never installs anything, never touches the source repo.
• If the grep recipe returns nothing (different agent's transcript shape), fall back to summarizing from the session JSON's last_prompt + files_touched only and note reduced fidelity in the file.