launch_wave: write the wave to a file and print a listing that fits in one tool result #247

Jira: AISDLC-453 (under AISDLC-187, pipeline hardening). This issue is item 1 of that story; item 2 (a longer wait-for-wave block) waits on the CI runner's Bash budget and is not part of this issue.

Observed

Every production batch of 13 items or more since 2026-10-02 makes the orchestrator improvise: the launch_wave result does not fit in one tool result, so the model greps its own transcript files for KEY= / DATA_FILE= / PROMPT_PATH=, and on 2026-10-14 it listed artifacts/rfe-reviews/ and the staging directory by hand while a wave was pending. Batches of 9 or fewer never do this.

Why it happens

scripts/pipeline_state.py next-action prints the whole wave as YAML (the launch_wave branch: output = {"action": "launch_wave", ..., "agents": agents} dumped with _BlockDumper). Every agent entry carries its full vars text: the type's launch block (type_registry.launch_vars(<type>, "auto-fix"), 65 lines / about 2.4 KB for rfe, 69 lines for initiative) followed by the agent's own two or three lines, rendered by _render_vars. ASSESS launches two agents per item (scorer and feasibility), so a 25-item wave prints about 50 copies of the same launch block, roughly 125 KB. The harness truncates a result that large to a file, and the model goes looking for the values in it.

What to do

1. Write the wave to a file and print a listing that fits. In the launch_wave branch of next-action:
   • write the complete agent list, one block per agent (subagent_type if present, prompt_file, the full vars), to tmp/pipeline-launch-wave.yaml (same naming family as tmp/pipeline-wave-ids.txt and tmp/pipeline-wave-launch.txt; overwrite per wave);
   • print action: launch_wave, phase, message, agents_file: tmp/pipeline-launch-wave.yaml, agent_count: N, then common_vars (the launch block, once, {ID} left unsubstituted) and agents: with each entry's subagent_type, prompt_file and only its per-agent lines (the lines _render_vars adds after the launch block, with {ID} substituted). The listing for a 50-agent wave must stay well under 30,000 characters; pin an upper bound in a test.
2. Read-back through the state helper. Add python3 scripts/state.py read-launch tmp/pipeline-launch-wave.yaml [--agent N] (prints the whole file, or one agent's full block) in scripts/state.py, next to read / read-ids. It exists so the model never reads the transcript: after a compaction it re-reads the file.
3. Say so where the model reads it. .claude/skills/rfe-auto-fix/SKILL.md, the **launch_wave**: step and the "Example launch_wave output" section: the prompt is common_vars with {ID} replaced by the agent's KEY value, then the agent's own lines, then "\n\nRead <prompt_file> and follow all instructions exactly."; a lost or compacted listing is recovered with the read-launch command, not by searching files. Keep the DISPATCH_LOOP banner in scripts/pipeline_state.py (cmd_dispatch_context, printed after a compaction) in step with the skill text. rfe-review does not use launch_wave and is untouched.

Constraints

• The production command stream is pinned: tests/test_pr1_transparent_edits.py and tests/test_typed_prompts.py::test_every_legacy_script_invocation_survives. Never edit the frozen corpus under tests/fixtures/legacy-prompts/; a changed or new script invocation in a skill body is recorded with _allow_command(...) in tests/test_typed_prompts.py, the way fix(#241): clear a feasibility flag a passing revision addressed #242 did for check_revised.py.
• rfe-auto-fix/SKILL.md body (everything after the frontmatter) must stay under 16,000 characters (skillsaw's 4,000-token budget, strict in CI; it is at about 7,200 today).
• No new literal key-prefix predicates in scripts/ (scripts/lint_prefix_predicates.py).
• tests/test_pipeline_state.py asserts launch_wave results in many places (around lines 2381–2649 and 3846–4279): keep action, phase, message as they are; extend the assertions to the file and the compact listing for both types.
• Verification: python3 -m pytest -q tests/test_pipeline_state.py tests/test_pr1_transparent_edits.py tests/test_typed_prompts.py tests/test_state.py and python3 scripts/lint_prefix_predicates.py. Do not run the whole suite or make lint.

Done when

• A 25-item ASSESS wave (50 agents) produces one launch_wave result under the bound, the file holds every agent's full block, and state.py read-launch returns it, for rfe and initiative (tests).
• The skill text and the dispatch banner describe the file form and the read-back command, and no other wording in the production command stream changed.