auto extraction doesn't work on opencode

[mirsella]

hello !

following #14 PR #28 implemented it. but it doesn't work for me at least.
i dont have the mcp. only the icm.js opencode plugin, and no memories are ever recorded.
icm 0.10.5

how can i debug this issue ?

thanks !

[pszymkowiak]

wshm · Automated triage by AI

🔍 Automated Triage

🐛 Category	bug
🟡 Priority	medium
🎯 Confidence	72%

Summary

Auto memory extraction not working with opencode plugin (icm.js); user reports no memories are recorded despite feature being implemented in PR #28. Needs debugging guidance or investigation.

---

Triaged automatically by wshm · This is an automated analysis, not a human review.

[pszymkowiak]

Fixed in #60 — switched OpenCode plugin to execFileSync + stdin, added binary check at load.

[pszymkowiak]

@mirsella Thanks for reporting! Two issues were found and fixed in v0.10.6:

Bug 1: Shell escaping (fixed in #60)

The plugin used execSync with shell quoting that broke on text containing quotes/special chars. Switched to execFileSync + stdin — no more shell escaping issues.

Bug 2: Extraction threshold too aggressive

The extraction engine requires a keyword score ≥ 3.0 to store a fact. Many typical tool outputs (file listings, diffs, etc.) don't hit this threshold, so nothing gets stored. Additionally, extraction only runs every 15th tool call (EXTRACT_EVERY = 15), so most outputs are skipped entirely.

How to verify after upgrading to v0.10.6

1. Update ICM: curl -fsSL https://raw.githubusercontent.com/rtk-ai/icm/main/install.sh | sh

2. Re-init: icm init --mode hook (reinstalls the fixed plugin)

3. Test extraction manually:

   echo "Fixed the database connection timeout by increasing pool size to 20" | icm extract -p myproject --dry-run

   This should output "Would extract 1 facts".

4. Check if anything was stored after a session:

   icm topics
   icm list --all

If extraction still doesn't fire, the issue may be that the tool.execute.after hook in your OpenCode version doesn't pass output as a string. You can verify by adding console.error("[icm] output type:", typeof output, "len:", output?.length) to the plugin at ~/.config/opencode/plugins/icm.js.

[mirsella]

thanks a lot will test !

[mirsella]

installed from git/source. facts are getting stored !

❯ icm stats
Memories:  352
Topics:    2
Avg weight: 0.958
Oldest:    2026-04-06 14:20
Newest:    2026-04-06 14:36

but now i think the auto insertion (layer 3) doesn't work. i dont have this log showing but i have the successful startup log:

console.error(
            `[icm] recalled ${ctx.split("\n").length} lines of context`,
          );

but especially the agent says it doesn't have any memories on session start, and even using the mcp tool doesn't work https://opncd.ai/share/D2bV84dz

Every recall returned:

• No memories found.

for the mcp maybe its the model wrongly using the mcp ?

manually:

✦2 ❯ icm recall-project
Project: waykeepers
Here is context from previous analysis of this project. Use it to answer efficiently without re-reading files.

- queued_players += 1;
- 181:     counts_by_mode: &mut HashMap<GameMode, GamemodePopulationCount>,
- 180: fn add_queue_population_counts(
- 193: fn add_in_game_population_counts(
- Line 1061: fn select_inviter_group_indices(
- Line 1102: fn append_matching_solo_indices(
- 195:         fn compatible_versions_require_exact_prerelease_match() {
- 203:         fn deployment_matching_uses_major_compatibility_for_release_versions() {
- 171:         fn request(version: &str) -> MatchmakingRequest {
- 182:         fn compatible_versions_match_same_major_and_build_metadata() {

---

so recall-* commands at least works