Skip to content

Apply Opus 5.5 long-run guidance: named stops, gate reports, evidence rules, parallel [P] - #22

Merged
forkrul merged 1 commit into
masterfrom
ccr-3635ed7e-ja7qow
Oct 4, 2026
Merged

forkrul merged 1 commit into
masterfrom
ccr-3635ed7e-ja7qow

Conversation

@forkrul

@forkrul forkrul commented Oct 3, 2026

Copy link
Copy Markdown
Owner

Applies the maintainer-approved lessons from Getting the most out of Opus 5.5 (2026-09-22) to the pipeline skills.

Stops (article: "Tell it which stops you want")

  • quench: amber means "check the failure message yourself and log it". It no longer says "stop here and confirm", which could read as asking the user on every task. Tasks run back to back, and status notes go in the same message as the next action. Quench halts only at its gate, a refusal, a Golden Rule spec edit, or before anything destructive.
  • temper / hone: run the next round straight away; a round boundary is not a gate.
  • smithy: new "Between gates, keep going" rule.
  • CLAUDE.md: the same stop rule for work on this repo. It stops before force-pushes, tag pushes (a tag push publishes a Release) and deleting branches, tags or releases.

Gate reports (article: "Read what it needs from you first")

  • Every stage gate, and smithy's handoff, now leads with Needs your call. Each stage names its own items: waivers, freeze exceptions, rejected or open findings, assumptions, bypasses. The handoff then shows Changed and Found. "Nothing" is printed too.

Evidence (article: "check its evidence", "how to show it fails", "mark what it couldn't confirm")

  • quench: an agent's report is a claim, not evidence. Quench re-runs reported amber and green itself before they count.
  • temper: a blocking finding must quote the question an implementer would have to ask. One without a question is downgraded to NIT.
  • hone: a blocking finding must show how it fails (a failing test, a repro input, or a command and its output). That repro seeds the test that guards the fix.
  • New UNCONFIRMED class in both loops. The judge confirms, kills, or carries each one into the next round as a targeted check. A finding the session can't check at all is logged as open and listed at the gate. A round that carries one forward is not clean.

Parallel [P] tasks (article: "split big work across subagents")

  • Quench may run a phase's [P] tasks side by side, each in its own worktree with the full cycle, and re-verifies each result. Tasks merge one at a time, with the suite and gates re-run after each merge. tasks.md is ticked and quench-log.md written on the feature branch at merge, never in the worktree, so the shared files don't conflict. A conflict, or a suite that goes red only after a merge, means the tag was wrong: the task is logged as anvil feedback and redone sequentially. Anvil's [P] definition is tightened to match.

Docs and example

  • README describes the gate report, UNCONFIRMED and hone's repros.
  • The example trail's review logs now carry an Unconfirmed: count. Hone round 1 shows the blocking finding's repro and a carried timing suspicion, which round 2's targeted check kills. Ratings are unchanged (C → A+ → A++).
  • CHANGELOG: new entries under [Unreleased] → Added (nothing has been released yet).

Verification

  • tests/smoke.sh passes (bash 5.2, Linux).
  • Ran CI's frontmatter lint, alias-symlink check and example-trail check locally; all pass.

🤖 Generated with Claude Code

https://claude.ai/code/session_01NgVoHT9xneQu3hPZSabSGv


Generated by Claude Code

… rules, parallel [P]

From "Getting the most out of Opus 5.5" (claude.dev, 2026-09-22):

- Stops: quench's amber is a self-check, not a pause for the user; tasks
  and temper/hone rounds run back to back, status rides with the next
  action, and work halts only at a gate, a refusal, or before anything
  destructive. CLAUDE.md gains the same stop rule for work on this repo.
- Reports: every gate leads with "Needs your call" (waivers, freeze
  exceptions, rejected/open findings, assumptions, bypasses), then
  Changed and Found.
- Evidence: quench re-runs agent-reported amber/green itself; temper's
  blocking findings quote the implementer's question, hone's carry a
  repro; suspicions are UNCONFIRMED, get a targeted check next round,
  and a round that carries one forward is not clean.
- Parallel: [P] tasks run the full cycle in separate worktrees and merge
  one at a time with the suite re-run after each merge.

The example trail's review logs carry the new Unconfirmed count, and
hone's round 1 shows a carried suspicion killed by round 2's check.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NgVoHT9xneQu3hPZSabSGv
@forkrul
forkrul merged commit 1a6c8b8 into master Oct 4, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants