dld-implement: keep the review subagent running under harness subagent restrictions

[jimutt]

Problem

Claude Code 2.1.220 ships this in its system prompt, unconditionally:

Do not call the AgentTool unless the user requested it
Do not use workflows or deep-research unless the user requested it

Confirmed by grepping the shipped binary — it is not config-driven, not an Opus 5 behavior change, and there is no file to edit to turn it off.

Under that instruction, an agent running /dld-implement can read step 6's review subagent as an unrequested subagent call and skip it. Observed in a real run: the agent skipped the subagent, reviewed the diff inline instead, and disclosed the skip. Graceful, but the disclosure was luck rather than a rule, and inline review is the author checking their own work — the weaker check the separate agent exists to avoid.

Approach

Raising the imperative ("you MUST launch a subagent") pits skill text against a system-prompt directive and yields nondeterminism across runs. The directive is conditional — unless the user requested it — so the fix satisfies the condition instead:

• Step 6 states that enabling implement_review and running the skill is the user's request, and notes this is one review agent at a fixed point rather than an unprompted fan-out.
• Names the tool as Agent / AgentTool so skill text and directive use the same token.
• Adds a fallback: if the subagent genuinely cannot run, review inline as a deliberate re-read from disk, then state in the final report which path ran and why.

The reframing is still an inference the model has to accept, so it is a strong nudge rather than a guarantee. The fallback is what makes the outcome safe either way — worst case is an inline review explicitly labeled as such, instead of a silent skip.

Changes

• skills/dld-implement/SKILL.md and .claude/skills/dld-implement/SKILL.md — step 6, kept in sync
• docs/framework/project-configuration.md — new "Implement Review" section covering the flag, why the reviewer is a separate agent, and the harness note

Verification

• tessl tile lint passes
• 165/165 bats tests pass
• The two skill copies diff only on the expected script paths and frontmatter

🤖 Generated with Claude Code