feat: LLM-summarized wake-up briefing (auto-detect invoker, cheap defaults, configurable)

[pszymkowiak]

Problem

Current wake-up = static top-N memory bullets (~440 tokens). For projects with rich memory history (decisions, resolved errors, blockers, licensing stages, smoke test states), this is too shallow — agents have to fire icm_memory_recall to get useful context, adding latency and turn cost.

When the user asks "where are we?" / "what's missing?", the static bullets give a surface view; the real signal lives in the broader memory store.

Proposal

Replace (or augment) the static wake-up with an LLM-generated narrative briefing, compiled out-of-band whenever project memories change.

Flow

1. Memory write → enqueue summarization job for the affected project
2. Job reads all memories (project-scoped + relevant globals)
3. LLM produces a structured briefing:
   • State of work (in flight, blocked)
   • Recent decisions and their rationale
   • Known errors / resolutions to remember
   • Project-specific user preferences
4. Briefing cached at ~/.icm/cache/wake-up-{project}.md
5. Session start loads cached briefing — zero added latency

Auto-detect the invoker AI tool

ICM should reuse the CLI already authenticated on the user's machine. No extra API key to manage.

Detection priority: | Prio | Mechanism | Notes |
|---|---|---|
| 1 | ICM_INVOKER env var | Explicit override (settable by hooks) |
| 2 | Tool-specific env vars | CLAUDECODE, CURSOR_TRACE_ID, GEMINI_CLI, CODEX_HOME, etc. |
| 3 | MCP clientInfo.name | When ICM is invoked via MCP |
| 4 | Parent process walk | Fallback (/proc/$PPID/comm on Linux) |
| 5 | Config TOML default | fallback_provider = "claude" |

Provider mapping (cheap defaults):

Invoker	Command	Default model
Claude Code	claude -p --model <m>	claude-haiku-4-5
Codex CLI	codex exec --model <m>	cheap codex equivalent (TBD)
Gemini CLI	gemini -p --model <m>	gemini-2.5-flash
Ollama	HTTP API	qwen2.5:7b
API direct	Anthropic/Google/OpenAI SDK	provider-specific cheap default

Why cheap defaults: summarization is a simple task. No need for Opus/Sonnet tier. User pays nothing extra in most cases (their own CLI quota).

Configuration (3 levels)

~/.config/icm/config.toml:

[wakeup.summarizer]
provider = "auto"              # auto | claude | codex | gemini | ollama | api
model = ""                     # empty = provider default
fallback_provider = "claude"
max_tokens = 1500
cache_dir = "~/.icm/cache"
regen_on_memory_write = true

CLI override: icm wakeup --summarizer-provider ollama --summarizer-model qwen2.5:7b
TUI: provider/model picker in icm interactive menu.
Per-project: .icm/config.toml in repo root.

Suggested PR breakdown (3 sub-PRs targeting v0.11)

PR 1 — Foundation (~500 LOC)

• Provider trait
• Claude CLI provider (only)
• Auto-detect invoker (env vars + fallback, no MCP/pproc yet)
• Config schema extension
• Synchronous regen at session start (no cache yet)
• MVP works end-to-end for Claude Code users

PR 2 — Cache & async hooks (~400 LOC)

• File-based cache with hash-based invalidation
• Memory write hook → enqueue async regen
• Wake-up loads cache if present, falls back to static bullets
• Zero added latency at session start

PR 3 — Multi-provider + TUI (~500 LOC)

• Codex / Gemini / Ollama providers
• TUI picker for summarizer config
• CLI flags --summarizer-*
• Docs + integrations updates

Non-goals

• Replace icm_memory_recall — raw recall stays for deep dives
• Real-time regeneration on session start (must be pre-computed)
• Cross-project synthesis — scope is per-project briefings

Open questions

• Invalidation: hash of all memory rows vs explicit dirty flag?
• Should briefing include "delta since last session"?
• Token budget: 1500 default, hard cap at 3000?
• Fallback on LLM unreachable: stale cache vs static bullets vs error?
• Cost guardrail: rate-limit regen (e.g., max 1/minute per project)?

Acceptance criteria

• PR 1: Claude CLI provider + auto-detect + config + sync regen
• PR 2: cache + memory-write hook + async regen
• PR 3: Codex + Gemini + Ollama providers + TUI + docs
• Smoke test: full flow with each provider
• Fallback behavior verified (no LLM, stale cache, missing CLI)
• Docs cover config + provider setup + opt-out

Milestone

Target: v0.11 (no milestone created yet — to be assigned).

[pszymkowiak]

wshm · Automated triage by AI

🔍 Automated Triage

✨ Category	feature
🟡 Priority	medium
🎯 Confidence	96%

Summary

Well-structured feature request for LLM-summarized wake-up briefings with auto-detected provider, caching, and multi-provider support; includes detailed PR breakdown targeting v0.11.

---

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

[pszymkowiak]

Extension scope: also drive icm consolidate

The same Provider/auto-detect/config infrastructure should be reused for icm consolidate (currently a naïve summaries.join(" | ") concatenation in cmd_consolidate, which inflates output past input size — bench-quality.ts Suite 3 surfaces the issue: compression ratio ~105%, score 90/100).

What changes

Add a parallel config block alongside [wakeup.summarizer]:

[consolidate.summarizer]
provider = "auto"           # auto | claude | codex | gemini | ollama | none
model = ""                  # empty = provider default
fallback_provider = "claude"
max_tokens = 400            # smaller than wake-up briefing
trigger = "manual"          # manual | threshold | nightly

Behavior:

• provider = "none" → keep current lexical concat (default for first release, opt-in to LLM)
• provider = "auto" or explicit → call the same Provider trait introduced in PR 1, with a consolidation-specific prompt template
• New CLI flags: icm consolidate --topic X --summarizer-provider claude --summarizer-model haiku

Why bundle this here

• Same Provider trait, same auto-detect, same config plumbing — no infra duplication
• Same cost model (cheap default tier, reuse the user's existing CLI quota)
• Same fallback story (LLM unreachable → fall back to lexical concat instead of erroring)

Impact on PR breakdown

PR 1 (Foundation) gets ~80 extra LOC to wire cmd_consolidate through the new Provider. Wake-up stays the headline use case; consolidate becomes a free side-benefit of the same infrastructure.

Acceptance criteria additions

• PR 1: icm consolidate honours [consolidate.summarizer] config when set; default behavior unchanged (provider = "none" preserves lexical concat)
• PR 1: smoke test — bench-quality.ts --suite consolidation shows compression ratio ≤ 60% with provider = "claude" and concept coverage stays at 8/8

Non-goals (explicit)

• No automatic LLM-driven consolidation on every memory write — that's auto_consolidate_with_embedder's territory and stays purely lexical for cost reasons
• LLM consolidate stays opt-in (manual icm consolidate invocation or explicit config)