wesleymatosdev/ce-pickup

ce-pickup 是做什麼的？

Extracts ce-handoff/v1 documents from past Hermes sessions using a local model. Zero frontier token cost. Output is standard ce-handoff/v1 Markdown, so /ce-handoff resume <path> works immediately.

Use this instead of running /ce-handoff on a cold session that would burn subscription quota just to catch up.

When to use

• User wants to resume a session they left unfinished without spending frontier tokens re-reading it
• You need a handoff from one or more past sessions but the context window is already loaded
• The user asks for a "pickup" or "catch-up" for an old session
• Running /ce-handoff on a large cold session would be expensive

Quick start

bash ~/.hermes/scripts/ce-pickup.sh SESSION_ID[,SESSION_ID,...]

Outputs to /tmp/compound-engineering-$(id -u)/ce-pickup/ by default.

Then resume from any output file:

/ce-handoff resume /tmp/compound-engineering-<uid>/ce-pickup/<slug>.md

Arguments

• Arg 1 (required): comma-separated session IDs (from session_search or Hermes browse)
• Arg 2 (optional): Ollama model name — default hf.co/ornith-ai/Ornith-1.5-35B-A3B-GGUF:Q5_K_M
• Arg 3 (optional): output directory override

Env overrides (for tests and alternate hosts): CE_PICKUP_DB (default ~/.hermes/state.db), CE_PICKUP_OLLAMA_URL (default http://localhost:11434/api/chat), CE_PICKUP_MAX_CHARS (default 24000).

Model routing

Default: Ornith-1.5-35B-A3B-GGUF:Q5_K_M — verified good extraction quality, runs locally.

If Ornith is not loaded, any capable local Ollama model works. Do NOT use a frontier model (that defeats the purpose).

Check what is available: ollama list

How it works

1. Reads ~/.hermes/state.db via sqlite3 — no Python, no TCC prompts
2. Extracts user + assistant messages ordered by timestamp
3. Truncates transcript to 24 000 chars (~6k tokens) — enough for meaningful extraction
4. Calls Ollama /api/chat via curl with a structured extraction prompt
5. Writes ce-handoff/v1 frontmatter + model body to the output dir
6. Filename is a slug of the session title; numeric suffix on collision

Output format

Every output file is valid ce-handoff/v1 with:

• artifact_contract: "ce-handoff/v1" frontmatter
• session_id, cwd, branch, repository where available
• Extraction covering: objective, completed work, current state, decisions/constraints, blockers, failed approaches, references, next steps

Context-ceiling watchdog (companion)

scripts/context-ceiling-watchdog.sh is a deterministic, no-agent monitor for live session context. It estimates the current context of each live session (active=1 message content ÷ 4 chars/token — never use sessions.input_tokens, which is cumulative billing volume including context re-reads) and enforces:

• >= 50k REMIND — one-line nudge, repeats hourly
• >= 70k ALERT — checkpoint warning, repeats every 30m
• >= 80k EXTRACT — runs ce-pickup locally and reports the ce-handoff/v1 path; after extraction, only re-alerts if the session climbs another ~10k (CRIT)

It is silent (empty stdout) unless a threshold fires, so it slots directly into a no-agent cron job with change detection. State lives in ~/.cache/context-ceiling-watchdog/ (flat TSV — no LLM memory, no recursive continuity). It never writes to Hermes config or runtime state; state.db is opened with sqlite3 -readonly.

Test it without a real model: bash scripts/mock-ollama.sh 21143 & then bash scripts/test-watchdog.sh (fixture DB + scaled thresholds, 20 checks).

Pitfalls

TCC dialogs from Python (macOS Sequoia+): Running Python as a subprocess of an app with broad entitlements can trigger TCC prompts for iCloud Drive or Documents — even when only reading ~/.hermes/state.db. The shell script uses sqlite3 + curl + jq + awk only and does NOT trigger TCC. Never use a Python-based implementation of this workflow unless the user explicitly accepts TCC prompts.

Transcript truncation: 24k chars covers most sessions. Very long sessions (200+ messages with large tool outputs) get tail-truncated. If the handoff feels thin, split into narrower queries.

Thinking models + num_predict: Ollama's num_predict caps total generated tokens INCLUDING message.thinking. A thinking model (Ornith is one) can burn the entire budget in the thinking channel and return empty or tail-cut message.content — which the script reads exclusively. Symptom: "empty response from Ollama" or a handoff that stops mid-sentence, looking like a model-quality problem when it is a payload bug. The script sends think:false (with a retry-without-the-flag fallback for models that reject it). Scope note from Wesley: that flag is correct for this mechanical extraction job ONLY — Ornith's thinking is an asset in agentic engineering use, and an extraction run says nothing about agentic capability. Do not generalize a bad extraction into a verdict on the model.

Ollama must be running: Check with ollama list before invoking.

Output is /tmp: OS may reclaim on reboot. Copy to a stable path if resuming across reboots.

Script

The canonical implementation is scripts/ce-pickup.sh in this skill — call it by its skill path. A host-side copy historically lived at ~/.hermes/scripts/ce-pickup.sh (documented as a symlink); on machines where that copy exists but is NOT a symlink, it is stale — trust the skill dir.

See also: ce-handoff skill for the resume workflow once the handoff doc exists.