deep-thinking

Deep thinking (chain of thought)

Externalize a written reasoning chain before irreversible or expensive actions, and a short after-action when outcomes falsify assumptions. The goal is not ceremony — it is to catch shortcut reasoning (e.g. nohup in a dying agent shell) before it burns a run.

This is thinking on the page, not a silent checklist. Prefer prose that shows why, then compress for the user-facing reply.

When to run (gates)

Invoke before acting when any of these are true:

• Kill / restart / rebind a long-lived process (inference, MCP, Docker, tunnels)
• Launch a multi-minute/hour eval, train, or data wipe (rm -rf output/…)
• Choose a new start pattern over a known-good one already used this session
• Symptoms could be process-lifecycle vs app-crash vs model-quality
• User asks to “think hard,” “CoT,” “deep think,” or explain a decision
• About to claim a root cause from a single datapoint

Skip for trivial edits, pure Q&A, or when the user already dictated the exact command.

How to use

1. Write the chain (tool scratch, todo note, or a short internal block) using the template below — fill every section; mark unknowns as unknowns.
2. Act only after the Decision section names the pick and why alternatives lost.
3. After-action when the result surprises you or burns >~5 minutes — update assumptions and leave one sticky takeaway (optionally memory_ingest if vault tools are available).

Do not dump the full chain into every user reply unless they asked for it. User-facing: 2–6 sentences of the conclusion + the sticky takeaway. Keep the full chain in the agent trail.

---

Chain template (fill in order)

Copy and complete:

### CoT — <short decision title>

#### 1. Goal
What success looks like for *this* step (not the whole project).

#### 2. World state (observed, not hoped)
- Processes / ports / terminals still alive:
- What was started how (Cursor background terminal vs nohup vs systemd vs Docker):
- What dies if *this* shell exits:
- Artifacts / run ids / last known-good scores:
- Clock / duration expectations:

#### 3. Evidence so far
Bullet facts with sources (log line, exit code, healthz, pack_errors).
Separate **signal** from **interpretation**.

#### 4. Assumptions (explicit)
| # | Assumption | If false, what breaks? | How to falsify quickly? |
|---|------------|------------------------|-------------------------|
| A1 | … | … | … |

Challenge at least one assumption that feels “obvious.”

#### 5. Hypotheses (competing)
H1: …
H2: …
H3: …
What would we observe if each were true?

#### 6. Options
For each option: steps, durability, blast radius, time cost, reuse of known-good patterns.

| Option | Durability | Blast radius | Time | Notes |
|--------|------------|--------------|------|-------|
| O1 … | | | | |
| O2 … | | | | |

#### 7. Decision
Pick: Ox
Because: …
Rejected: … because …
Risk I am accepting: …
Rollback: …

#### 8. First verification
The smallest probe that proves the decision held (healthz, one smoke chat,
one doc, listener still up after parent shell ends).

#### 9. Stop conditions
Abort / rethink if: …