Inspect

Global Review

Dispatch a code-reviewer subagent to review the whole change as one body of work, not task-by-task. Where tn:assemble reviewed each task in isolation (spec compliance + code quality), inspect catches problems that only surface when you look at the change end-to-end: cross-task consistency, architectural integrity, and security.

Core Principle

REVIEW THE WHOLE, NOT THE PARTS.
Each task passed review alone — but do they fit together?
The reviewer is a safety net, not a rubber stamp.

Why inspect ≠ assemble's per-task review

	tn:assemble review	tn:inspect
Scope	one task	the entire change, all tasks together
Questions	did this task match its spec?	do the tasks fit? is the whole sound?
Catches	per-task defects	emergent issues: inconsistency, arch drift, security holes

Inspect is not a re-run of assemble's reviews. It looks at what those reviews could not see — the gaps between tasks.

Steps

Phase 1: Request Review

Step 1: Get Git SHAs

Use the merge-base of the change, not HEAD~1 (a change often spans many commits, so HEAD~1 misses everything but the last):

BASE_SHA=$(git merge-base HEAD origin/main)  # whole change from branch point
HEAD_SHA=$(git rev-parse HEAD)

Step 2: Dispatch Code-Reviewer Subagent

Dispatch a reviewer subagent using the template at ./code-reviewer-prompt.md. The template is host-neutral — see its header for how to dispatch on Claude Code (Agent tool) vs Codex (subagent). Fill the placeholders, then send the full prompt block to the reviewer.

Placeholders to fill:

• WHAT_WAS_IMPLEMENTED — What the whole change built (the proposal's "what")
• PLAN_OR_REQUIREMENTS — Reference proposal.md + the full tasks.md (all tasks, not one)
• BASE_SHA — Merge-base of the change
• HEAD_SHA — Current HEAD
• DESCRIPTION — Brief summary of the change as a whole
• CODING_CONVENTIONS — From techneering/config.yaml coding_conventions section (if exists)
• QUALITY_STANDARDS — From techneering/config.yaml quality_standards section (if exists)

Step 3: Receive Feedback

The reviewer returns:

• Strengths: What was done well
• Issues: Critical / Important / Minor
• Assessment: Approved / Needs fixes

Phase 2: Process Feedback

The Response Pattern:

WHEN receiving code review feedback:

1. READ: Complete feedback without reacting
2. UNDERSTAND: Restate requirement in own words (or ask)
3. VERIFY: Check against codebase reality
4. EVALUATE: Technically sound for THIS codebase?
5. RESPOND: Technical acknowledgment or reasoned pushback
6. IMPLEMENT: One item at a time, test each

Forbidden Responses:

• "You're absolutely right!" / "Great point!" / "Excellent feedback!"
• "Let me implement that now" (before verification)
• Agreeing without understanding the issue

Correct Responses:

• Restate the technical requirement
• Ask clarifying questions
• Push back with technical reasoning if wrong
• Just start working (actions > words)

Step 4: Act on Feedback

• Fix Critical issues immediately
• Fix Important issues before proceeding
• Note Minor issues for later
• Push back if reviewer is wrong (with reasoning)

Red Flags

Thought	Reality
"assemble already reviewed each task"	Per-task review can't see cross-task gaps — that's inspect's whole point
"Review passed, can skip it"	Review is a quality gate, not a formality
"The reviewer is always right"	The reviewer may lack full context — must verify independently
"Merge directly without review"	Unreviewed code is a source of technical debt
"Fix Critical issues later"	Critical = must fix immediately, no exceptions
"Feedback too long, won't read"	Every item must be read, understood, verified, responded to

Common Rationalizations

Excuse	Reality
"assemble's reviews covered it"	assemble reviewed each task alone; inspect reviews the assembled whole
"Tight deadline, skip review"	Time saved by skipping returns doubled in bug costs
"Just a small change, no review needed"	Small changes introduce big bugs all the time
"Already ran tests"	Tests verify behavior, review verifies design and integration

Quick Reference

Dimension	Key point
Scope	The whole change (merge-base → HEAD), not one task
Focus	Cross-task consistency · architecture · security
Difference from assemble	assemble = per-task; inspect = whole-change
Two phases	Request review → receive and process feedback
Feedback levels	Critical (must fix) / Important (should fix) / Minor (optional)
Processing pattern	read → understand → verify → evaluate → respond → implement
Push back allowed	Breaks functionality / lacks context / YAGNI / technically wrong

Next Step

When done:

• Review passed (Approved) → invoke tn:ship for branch integration decision
• Review found Critical issues → fix and re-submit for review
• Review found architecture issues → roll back to tn:forge to redo affected tasks
• User requests stop → end the current flow

Guardrails

• Review runs after tn:audit passes and before tn:ship
• inspect is a whole-change review — do not collapse it into a per-task re-check
• Never skip review for any reason
• Never blindly accept all feedback — verify each point independently
• Never merge without review approval
• Use merge-base, not HEAD~1, so the review covers the entire change
• Push back when: suggestion breaks functionality / reviewer lacks context / violates YAGNI / technically incorrect
• Acknowledge correct feedback with technical descriptions, not praise
• Fix one issue at a time, test each fix