Claude Code runbook — greeting capability gate

Role: This plan is for Claude Code in an interactive terminal, invoked by the Cursor orchestrator. Not for Cursor execute unless the human overrides. Parent plan: sophia-bot-2026-10-30-greeting-capability-gate Task: TICKET-108744 (+108597) Worktree: /home/asksuite/Documentos/.agent-worktrees/sophia-bot/ticket-108744 Branch: fix/TICKET-108744-greeting-day-use-offer (no upstream)

Orchestrator contract

Open interactive terminal (human can type mid-run).
cd into the worktree above.
Launch Claude Code once per task (or one session that does tasks in order — human chooses).
Paste the Task brief for the current task only.
Wait for the Output block (format below). Do not accept long essays as done.
Orchestrator reviews diffs + output → next task or review skill.
No commit. No push. Human asks separately.

Suggested launch (adjust to local Claude Code CLI):

cd /home/asksuite/Documentos/.agent-worktrees/sophia-bot/ticket-108744
# interactive — leave TTY attached; do not pipe away stdin
claude

If the CLI needs a one-shot prompt flag, still keep the session interactive so the human can intervene.

Hard rules for Claude Code

Read parent plan acceptance; implement one task at a time (TDD red→green).
Repo: only this worktree. Do not touch other repos.
Follow AGENTS.md (incl. prompt rules: general gates, no incident answer-keys).
No new code comments. No drive-by refactors.
End every task with the Output block only (plus errors if any).

Output block (mandatory — keep short)

STATUS: DONE | BLOCKED | PARTIAL
TASK: <N>
FILES: <comma-separated paths changed>
VERIFY: <exact pytest command> → <pass|fail|skipped> (<n passed>/<n>)
DIFF_SUMMARY: <one line>
NEXT: <ready for task N+1 | need human: …>
BLOCKERS: <none | …>

No prose outside this block unless STATUS=BLOCKED (then ≤5 lines).

Task briefs (paste one)

Brief — Task 1 (Critic)

Parent: ask-handoff plan sophia-bot-2026-10-30-greeting-capability-gate Task 1. TDD first: create tests/test_unit/test_response_critic_capability_gate.py, then edit src/agents/reason/reason_core/llm/response_critic.py. Gate: pure help/greeting exempt; naming specific product/service/amenity as offerable requires grounding; tool-backed quotation/form offers stay exempt; no “don’t say day use” ban. Verify: pipenv run pytest tests/test_unit/test_response_critic_capability_gate.py -q Emit Output block.

Brief — Task 2 (Reason unit)

Parent Task 2. TDD: tests/test_unit/test_reason_unit_capability_gate_prompt.py using _capture_system_prompt pattern from tests/test_unit/test_graph_configuration.py. Edit src/agents/reason/reason_unit/graph.py: drop day use prices/rates from primary-contact bullet; add general greeting/offer gate (grounded products / tool-backed only). Verify: pipenv run pytest tests/test_unit/test_reason_unit_capability_gate_prompt.py -q Emit Output block.

Brief — Task 3 (Reason chain)

Parent Task 3. Same as Task 2 for src/agents/reason/reason_chain/graph.py + tests/test_unit/test_reason_chain_capability_gate_prompt.py. Verify: pipenv run pytest tests/test_unit/test_reason_chain_capability_gate_prompt.py -q Emit Output block.

Brief — Task 4 (Integration smoke)

Parent Task 4. Create tests/test_integration/test_reason_greeting_capability_gate.py: bare PT greeting must not advertise Day Use without sources. Skip cleanly if no LLM env. Verify: pipenv run pytest tests/test_integration/test_reason_greeting_capability_gate.py -q Emit Output block.

Brief — Task 5 (Sweep)

Parent Task 5. Run focused unit sweep from parent plan; fix only regressions from Tasks 1–4. Verify: command from parent Task 5. Emit Output block.

After all tasks

Orchestrator runs leanflow review (designated model). Human validates. Commit only on explicit ask (TICKET-108744 - fix: …, note 108597 in PR/body).

Claude Code runbook — greeting capability gate (subagent)