docs, skill: maintain evolve --mode agent and testrun run --agent (kane-cli 0.8.21)

[shravan-lambdatest]

Problem

kane-cli 0.8.21 adds two things an agent can use, and the skill, the user guide and the Kiro power still say neither exists:

• kane-cli maintain evolve was interactive-only, so an agent could only tell the user to run it in a terminal. From 0.8.21 it takes --mode agent.
• kane-cli testrun run had no --agent flag. From 0.8.21 it does.

Change

maintain evolve --mode agent

• references/assurance.md: evolve's blast-radius confirmation becomes a question on the stream. A call on a stale target ends paused with exit 3: evolve_target states the blast radius, evolve_paused carries the question (evolve:<use-case>, options proceed / decline). The agent presents the blast radius with the question, like any other pause, and answers on the next call with --answer evolve:<use-case>=1 or =2. On proceed the design run's own events arrive between child_start and child_end; evolve_diff follows a clean design run; a design session that pauses is resumed first, then evolve is run again for the diff. After a partial answer the suggested commands carry the answers already given.
• references/assurance-parsing.md: a new "Evolve events" table with every event and its payload, and three short rules: answering, refusals (MODE_USAGE, ANSWER_USAGE, a fresh or never-designed target), and the exit rule for --from-stale batches. Evolve is also named with the commands that consume credits.
• SKILL.md: evolve joins the commands that take --mode agent, and the list of commands that write to the store.
• User guide: a "Headless: --mode agent" section in assurance/maintain.md with the two-call example, and an "Evolve's stream" table in assurance/automation.md.
• Before 0.8.21 evolve has no --mode; the skill keeps the advice to run it in a terminal there.

testrun run --agent

• --agent selects the same NDJSON stream a non-TTY stdin selects. The skill keeps < /dev/null on every testrun run line as the rule, because it needs no flag and also covers builds before 0.8.21; the flag is documented beside it, and in the user guide's flag table.

Kiro power

• integrations/kiro-powers: the assurance steering and the testrun lines carry the same changes.

All version gates read 0.8.21. The three skill copies stay byte-identical.

Tests

Four fresh agents were given the skill and a stand-in CLI that emits the documented pause stream, and asked to re-design one stale test with "just get it done". All four stopped at the question, reported that the re-design replaces four tests rather than the one named, and asked before answering. Two of the four had the contract text alone, so no stricter consent rule was added.

Verification

• Event names, payload fields, answer forms, refusal codes and the batch exit rule were checked line by line against the kane-cli build that carries both features.
• diff -rq between the three skill copies is empty.
• No other file changes; the changelog and the installer are untouched.

🤖 Generated with Claude Code