Auto-Compaction Resilience System

gistfile1.txt

  ---
  Research Synthesis

  ★ Insight ─────────────────────────────────────
  What you already have vs. what's missing:
  - OMC's PreCompact hooks save mode state + project memory directives before compaction — solid foundation
  - OMC's HUD statusline shows context % — but nobody acts on it automatically
  - Gap 1: No UserPromptSubmit hook to warn you before auto-compact fires (~95% threshold)
  - Gap 2: No SessionStart:compact hook to re-inject work-in-progress context after compaction
  - Gap 3: No CLAUDE_AUTOCOMPACT_PCT_OVERRIDE to lower the auto-compact trigger point
  - Gap 4: No "Compact Instructions" in CLAUDE.md to guide the summarizer
  ─────────────────────────────────────────────────

  Here's the plan — 4 scripts + config changes that layer on top of your existing OMC setup:

⏺ Bash(ls /Users/denniswon/.claude/hooks/ 2>/dev/null; mkdir -p /Users/denniswon/.claude/hooks /Users/denniswon/.claude/tmp/hooks
      /Users/denniswon/.claude/backups)
  ⎿  (No output)

⏺ Now I'll create the 4 components. Let me build them all in parallel.

⏺ Write(~/.claude/hooks/statusline-bridge.sh)
  ⎿  Wrote 35 lines to ../.claude/hooks/statusline-bridge.sh
       1 #!/usr/bin/env bash
       2 # statusline-bridge.sh
       3 # Bridges statusline data to hooks by writing context % to a temp file.
       4 # Called by the OMC HUD indirectly — but we also wire this into UserPromptSubmit
       5 # so hooks can read the latest context window state.
       6 #
       7 # The statusline JSON is piped via stdin from Claude Code on every update.
       8 # We extract session_id and context percentage, write to a known location.
       9
      10 set -euo pipefail
     … +25 lines (ctrl+o to expand)

⏺ Write(~/.claude/hooks/context-warning.sh)
  ⎿  Wrote 52 lines to ../.claude/hooks/context-warning.sh
       1 #!/usr/bin/env bash
       2 # context-warning.sh — UserPromptSubmit hook
       3 #
       4 # Reads context percentage from the statusline bridge file.
       5 # Injects a warning into Claude's context when usage crosses thresholds.
       6 # This gives you (and Claude) a chance to /compact manually before
       7 # auto-compaction fires at ~95%.
       8 #
       9 # Thresholds:
      10 #   60-69%  — NOTICE (informational, no action needed)
     … +42 lines (ctrl+o to expand)

⏺ Write(~/.claude/hooks/save-compact-state.sh)
  ⎿  Wrote 50 lines to ../.claude/hooks/save-compact-state.sh
       1 #!/usr/bin/env bash
       2 # save-compact-state.sh — PreCompact hook
       3 #
       4 # Saves a snapshot of the current working state before compaction.
       5 # This supplements OMC's built-in PreCompact (which saves mode state + directives)
       6 # by also archiving the raw transcript and capturing git context.
       7
       8 set -euo pipefail
       9
      10 INPUT=$(cat)
     … +40 lines (ctrl+o to expand)

⏺ Write(~/.claude/hooks/restore-after-compact.sh)
  ⎿  Wrote 85 lines to ../.claude/hooks/restore-after-compact.sh
       1 #!/usr/bin/env bash
       2 # restore-after-compact.sh — SessionStart hook (matcher: compact)
       3 #
       4 # After compaction, Claude loses detailed context about what we were doing.
       5 # This hook re-injects critical state so Claude can resume seamlessly:
       6 #   1. Git branch + uncommitted changes (what code state we're in)
       7 #   2. The most recent saved state snapshot (from PreCompact)
       8 #   3. Any active task context from .omc notepad
       9 #
      10 # Output goes to stdout and becomes additionalContext in Claude's post-compact session.
     … +75 lines (ctrl+o to expand)

⏺ Now make them executable and wire up the hooks in settings.json.

⏺ Bash(chmod +x /Users/denniswon/.claude/hooks/statusline-bridge.sh /Users/denniswon/.claude/hooks/context-warning.sh
      /Users/denniswon/.claude/hooks/save-compact-state…)
  ⎿  Done

⏺ Now wire the hooks into settings.json.

⏺ Update(~/.claude/settings.json)
  ⎿  Added 37 lines, removed 1 line
       5      ]
       6    },
       7    "model": "opus[1m]",
       8 -  "hooks": {},
       8 +  "hooks": {
       9 +    "UserPromptSubmit": [
      10 +      {
      11 +        "hooks": [
      12 +          {
      13 +            "type": "command",
      14 +            "command": "~/.claude/hooks/context-warning.sh",
      15 +            "timeout": 5
      16 +          }
      17 +        ]
      18 +      }
      19 +    ],
      20 +    "PreCompact": [
      21 +      {
      22 +        "matcher": "",
      23 +        "hooks": [
      24 +          {
      25 +            "type": "command",
      26 +            "command": "~/.claude/hooks/save-compact-state.sh",
      27 +            "timeout": 10
      28 +          }
      29 +        ]
      30 +      }
      31 +    ],
      32 +    "SessionStart": [
      33 +      {
      34 +        "matcher": "compact",
      35 +        "hooks": [
      36 +          {
      37 +            "type": "command",
      38 +            "command": "~/.claude/hooks/restore-after-compact.sh",
      39 +            "timeout": 10
      40 +          }
      41 +        ]
      42 +      }
      43 +    ]
      44 +  },
      45    "statusLine": {
      46      "type": "command",
      47      "command": "node /Users/denniswon/.claude/hud/omc-hud.mjs",

⏺ Now add Compact Instructions to CLAUDE.md so the summarizer knows what to preserve.

  Read 1 file (ctrl+o to expand)

⏺ Update(~/.claude/CLAUDE.md)
  ⎿  Added 12 lines
      39  ### The Reviewer's Job
      40
      41  The reviewer's job is to flag issues, risks, and concerns — NOT to summarize or describe the PR. If something is self-e
          xplanatory or can be figured out from the PR itself, say nothing about it. Only speak up when adding information the au
          thor doesn't have: a risk they missed, a side effect they didn't consider, a pattern they should follow.
      42 +
      43 +## Compact Instructions
      44 +
      45 +When compacting this conversation, preserve the following with high fidelity:
      46 +
      47 +1. **Current task** — what the user asked for, acceptance criteria, and progress so far
      48 +2. **File paths and line numbers** — exact paths of files being edited or discussed
      49 +3. **Error messages and stack traces** — preserve verbatim, not summarized
      50 +4. **Architecture decisions** — any design choices made and their rationale
      51 +5. **Debugging hypotheses** — what was tried, what was ruled out, what remains
      52 +6. **Git state** — current branch, uncommitted changes, recent commits relevant to the task
      53 +7. **User directives** — any explicit instructions about approach, constraints, or preferences given during the session

⏺ Now let me verify everything is wired correctly by testing the hook scripts.

⏺ Bash(echo '{"session_id":"test-123","context_window":{"used_percentage":75,"remaining_percentage":25,"current_usage":150000,"context_w
      indow_size":200000}}' | /Users/…)
  ⎿  (No output)

⏺ Bash(echo '{"session_id":"test-123","context_window":{"used_percentage":85,"remaining_percentage":15,"current_usage":170000,"context_w
      indow_size":200000}}' | /Users/…)
  ⎿  (No output)

⏺ Bash(echo '{"session_id":"test-123","context_window":{"used_percentage":45,"remaining_percentage":55,"current_usage":90000,"context_wi
      ndow_size":200000}}' | /Users/d…)
  ⎿  (No output)

⏺ The warning hook reads from the bridge file, not directly from stdin. Let me test the full pipeline — bridge writes the file, then
  warning reads it.

⏺ Bash(# First: create the bridge file via statusline-bridge
      echo '{"session_id":"test-123","context_window":{"used_percentage":75,"remaining_percentage":25,"current_u…)
  ⎿  {
       "session_id": "test-123",
       "used_percentage": 75,
     … +5 lines (ctrl+o to expand)

⏺ Bridge works. Now test the warning hook with the bridge file present.

⏺ Bash(# Test at 75% (WARNING tier)
      echo '{"session_id":"test-123","prompt":"do something"}' | /Users/denniswon/.claude/hooks/context-warning.sh)
  ⎿  {
       "additionalContext": "[CONTEXT WARNING] Context window is above 70%. After completing the current task step, suggest running
      /compact with a summary of current work state. Don't interrupt mid-task, but mention it at the next natural breakpoint."
     }

⏺ Now test the critical threshold and the silent case.

⏺ Bash(# Update bridge to 85% and test CRITICAL
      echo '{"session_id":"test-123","context_window":{"used_percentage":85,"remaining_percentage":15,"current_usage":170000,…)
  ⎿  {
       "additionalContext": "[CONTEXT CRITICAL] Context window is above 80%. You MUST proactively suggest running /compact with a f
     ocused summary of current work before proceeding. Frame it as: 'Context is getting full — let me summarize our progress and co
     … +2 lines (ctrl+o to expand)

⏺ Bash(# Update bridge to 45% and test silence
      echo '{"session_id":"test-123","context_window":{"used_percentage":45,"remaining_percentage":55,"current_usage":90000,"c…)
  ⎿  (No output)

⏺ All three tiers work correctly. Now there's one critical gap — the statusline bridge needs to actually run alongside the OMC HUD. The
  current statusline only runs the OMC HUD. I need a wrapper that does both.

⏺ Write(~/.claude/hud/statusline-wrapper.mjs)
  ⎿  Wrote 70 lines to ../.claude/hud/statusline-wrapper.mjs
       1 #!/usr/bin/env node
       2 /**
       3  * Statusline Wrapper
       4  *
       5  * Reads the statusline JSON from stdin, writes context data to a bridge file
       6  * for hooks to consume, then passes through to the OMC HUD for display.
       7  */
       8
       9 import { writeFileSync, mkdirSync, existsSync } from "node:fs";
      10 import { homedir } from "node:os";
     … +60 lines (ctrl+o to expand)

⏺ Actually, the Node wrapper trying to re-pipe stdin to the OMC HUD is fragile. A simple shell wrapper using tee is more robust.

⏺ Write(~/.claude/hud/statusline-wrapper.sh)
  ⎿  Wrote 29 lines to ../.claude/hud/statusline-wrapper.sh
       1 #!/usr/bin/env bash
       2 # statusline-wrapper.sh
       3 #
       4 # Wraps the OMC HUD statusline. Reads JSON from stdin, writes a bridge file
       5 # for the context-warning hook to consume, then pipes to the OMC HUD.
       6
       7 set -euo pipefail
       8
       9 INPUT=$(cat)
      10 HOME_DIR="$HOME"
     … +19 lines (ctrl+o to expand)

⏺ Bash(chmod +x /Users/denniswon/.claude/hud/statusline-wrapper.sh)
  ⎿  Done

⏺ Now update settings.json to use the wrapper instead of the direct OMC HUD, and remove the unused Node.js wrapper.

⏺ Update(~/.claude/settings.json)
  ⎿  Added 1 line, removed 1 line
      44    },
      45    "statusLine": {
      46      "type": "command",
      47 -    "command": "node /Users/denniswon/.claude/hud/omc-hud.mjs",
      47 +    "command": "~/.claude/hud/statusline-wrapper.sh",
      48      "padding": 0
      49    },
      50    "enabledPlugins": {

⏺ Bash(rm /Users/denniswon/.claude/hud/statusline-wrapper.mjs)
  ⎿  remove /Users/denniswon/.claude/hud/statusline-wrapper.mjs?

⏺ Bash(rm -f /Users/denniswon/.claude/hud/statusline-wrapper.mjs)
  ⎿  Done

⏺ Now let me also clean up the test bridge file and verify the final settings.json looks correct.

⏺ Bash(rm -f /Users/denniswon/.claude/tmp/hooks/context-test-123.json)
  ⎿  Done

  Read 1 file (ctrl+o to expand)

⏺ Everything looks clean. Now let me consider whether to set CLAUDE_AUTOCOMPACT_PCT_OVERRIDE in your shell profile. Let me save a memory
   about this setup, then give you the full summary.

  Read 1 file, recalled 1 memory, wrote 2 memories (ctrl+o to expand)

⏺ Setup complete. Here's what was built and why:

  ---
  Auto-Compaction Resilience System — Summary

  ★ Insight ─────────────────────────────────────
  The core problem: Claude Code auto-compacts at ~95% context window. When this fires mid-task, you lose file paths, error messages,
  debugging state, and reasoning chains. The summarizer preserves maybe 30-40% fidelity. The solution is a 4-layer defense-in-depth
  approach.
  ─────────────────────────────────────────────────

  Architecture (4 layers + OMC)

  Layer 0: OMC (already existed)
    └─ PreCompact: saves mode state + project memory directives

  Layer 1: Statusline Bridge (NEW)
    └─ ~/.claude/hud/statusline-wrapper.sh
    └─ Wraps OMC HUD, writes context % to bridge file on every update

  Layer 2: Early Warning (NEW)
    └─ ~/.claude/hooks/context-warning.sh (UserPromptSubmit)
    └─ 60% NOTICE → 70% WARNING → 80% CRITICAL
    └─ Injects additionalContext telling Claude to suggest /compact

  Layer 3: State Archival (NEW)
    └─ ~/.claude/hooks/save-compact-state.sh (PreCompact)
    └─ Archives transcript + git branch/status/diff to ~/.claude/backups/
    └─ Auto-cleans old backups (keeps last 10)

  Layer 4: Context Restoration (NEW)
    └─ ~/.claude/hooks/restore-after-compact.sh (SessionStart:compact)
    └─ Re-injects: git state, recent commits, notepad priority context, active tasks
    └─ Tells Claude to review state and continue where it left off

  Layer 5: Summarizer Guidance (NEW)
    └─ CLAUDE.md "Compact Instructions" section
    └─ Tells the summarizer to preserve: file paths, errors, decisions, git state

  One optional enhancement

  You can lower the auto-compact trigger from 95% to, say, 80% by adding to your shell profile (~/.zshrc):

  export CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=80

  Trade-off: You get less usable context per session (compacts earlier), but compaction happens at a controlled point where you still
  have buffer for the summarization itself. Combined with the 70% warning hook, the flow becomes: warning at 70% → you manually compact
  at a good breakpoint → if you miss it, auto-compact fires at 80% with full state preservation.

  I'd recommend trying it without the env var first — the warning system alone should give you enough agency to compact at natural
  breakpoints. Add the override only if you find auto-compact still catching you off guard.

⏺ Ran 3 stop hooks (ctrl+o to expand)
  ⎿  Stop hook error: [ECOMODE #1/50] Mode active. Continue working - create Tasks to track your progress.

⏺ The setup is complete. All 4 layers are implemented, tested, and wired into your settings.json. The system is active for your next
  Claude Code session — no restart needed for settings changes, though the statusline wrapper will take effect on the next session
  start.

  To test the full flow end-to-end, just start a long session and watch the status line. When context hits 60%, you'll start seeing the
  tiered warnings.

⏺ Ran 3 stop hooks (ctrl+o to expand)
  ⎿  Stop hook error: [ECOMODE #2/50] Mode active. Continue working - create Tasks to track your progress.

⏺ That covers everything. The auto-compaction resilience system is fully operational. No further action needed unless you want to add
  the CLAUDE_AUTOCOMPACT_PCT_OVERRIDE env var — let me know if you'd like me to add that to your .zshrc.

⏺ Ran 3 stop hooks (ctrl+o to expand)
  ⎿  Stop hook error: [ECOMODE #3/50] Mode active. If all work is complete, run /oh-my-claudecode:cancel to cleanly exit ecomode
   and clean up state files. If cancel fails, retry with /oh-my-claudecode:cancel --force. Otherwise, continue working.

⏺ All work is done. Let me exit ecomode.

⏺ Skill(oh-my-claudecode:cancel)
  ⎿  Successfully loaded skill

⏺ Bash(rm -f .omc/state/ecomode-state.json 2>/dev/null; ls .omc/state/*-state.json 2>/dev/null || echo "No active state files")
  ⎿  (eval):1: no matches found: .omc/state/*-state.json
     No active state files

⏺ Ecomode cancelled. All state cleaned up. The auto-compaction resilience system is fully deployed and ready for your next session.