#!/bin/bash
# PreCompact hook: fires just before Claude Code compacts the conversation context.
# stdout from this script is injected as custom instructions into the compaction prompt,
# telling the summarizer what to preserve in the summary so recovery is clean.

COMPACT_STATE_FILE="$HOME/.claude/compact-state.json"

# Write a state marker with timestamp so post-compact recovery can detect it
TIMESTAMP=$(date -u +"%Y-%m-%dT%H:%M:%SZ")
echo "{\"compact_at\": \"$TIMESTAMP\", \"recovered\": false}" > "$COMPACT_STATE_FILE"

# Output instructions for the summarizer — these get embedded in the compaction prompt
cat <<'INSTRUCTIONS'
NEURON AGENT — COMPACT PRESERVATION RULES:

This is a Neuron agent session. The agent has persistent memory via Neuron MCP.
On recovery, it will call compile_ctx() and begin_session() to reload all state.
The summary only needs to cover what Neuron doesn't already have.

CRITICAL — the summary MUST preserve ALL of the following:

1. ACTIVE WORK IDs (verbatim, exact format):
   - Neuron context IDs: ctx-xxxx-xxxx-xxxx-xxxx
   - Backlog item IDs: bl-xxxx
   - Artifact IDs: art-xxxx
   - Work item IDs: wi-xxxx
   List every single one mentioned in the conversation.

2. CURRENT TASK STATE:
   - Exact task name / description
   - Last completed step
   - Next step to execute (be specific)
   - Files actively being edited (full paths)
   - Any code/content that was being written but not yet saved

3. PENDING USER INSTRUCTIONS (verbatim):
   - Every instruction the user gave that has NOT yet been fully executed
   - User preferences stated this session
   - Things the user said they "never want" or "always want"

4. KEY DECISIONS (architectural, irreversible):
   - What was chosen and why
   - What was explicitly rejected and why

5. ERRORS & BLOCKERS:
   - Any errors encountered and their resolutions
   - Any unresolved blockers

6. CONVERSATION CONTEXT:
   - What the user has been asking about
   - Tone / mood of the session
   - Any frustrations or preferences expressed

FORMAT:
Begin the summary with:
  COMPACT RECOVERY POINT [ISO-TIMESTAMP]:

Then use these exact section headers in order:
  ## Active Work IDs
  ## Current Task
  ## Pending Instructions
  ## Key Decisions
  ## Errors & Fixes
  ## User Context

RECOVERY BEHAVIOR (instruct the post-compact agent):
- Do NOT say "I see we were working on..." or "Based on the previous conversation..."
- Do NOT acknowledge the compact at all
- Silently call mcp__neuron__compile_ctx() and mcp__neuron__begin_session() FIRST
- Then immediately resume the task at the exact next step, as if no interruption occurred
- The user should not know a compact happened unless they look for it

INSTRUCTIONS

exit 0