Daily harness signal

Make control flow inspectable

July 27, 2026 · JST Two fresh findings State identity · delegation receipts
Two fresh failures point to one harness rule: bind control flow to inspectable identity and events, not advisory prose that a resumed session or hidden prompt can reinterpret.
01 · Fresh · source date 2026-07-23

Give every long run its own named scratch state

Use when: One checkout executes, pauses, or resumes more than one plan, especially when briefs, reviewer packages, and completion ledgers live outside Git.

Action: Derive a stable run directory from the canonical plan path plus a digest, for example .agent-runs/<plan-slug>-<digest>/ Put every brief, report, review artifact, and progress.md there. Make line one of the ledger plan=<canonical-path> digest=<digest> and, on resume, reject any mismatch. Delete only that run directory after final review passes.

Acceptance check: Start Plan A, mark task one complete, then start Plan B in the same checkout and resume A. Pass only if B begins with an empty ledger, A resumes at task two without Git archaeology, neither writes into the other directory, and finishing A leaves B intact.

Evidence: Superpowers 6.2.0 ships this structure after cross-plan contamination. Its release links 25 baseline and post-change evals; the versioned sdd-workspace, task-brief, and review-package scripts all require a plan file and resolve artifacts beneath its directory.

Caveat: The shipped helper keys by basename, not a path digest, so identically named plans in different folders can still collide. Git-ignored scratch also disappears under git clean -fdx; keep commit IDs in the ledger and recover durable truth from Git.

02 · Fresh · source date 2026-07-25

Require a subagent receipt, not a standing wish

Use when: A review, research, or verification step is correctness-critical and must run in a separate Claude Code subagent rather than silently falling back to the parent.

Action: Put the delegation in the live request, not only CLAUDE.md:

@"code-reviewer (agent)" review <range> against <criteria>; return findings with file:line evidence.

For automation, require an Agent tool event naming that worker before accepting its result.

Acceptance check: Run one harmless canary after every Claude Code or model upgrade. Pass only if the transcript records the intended Agent invocation, the child returns the sentinel, and the parent does not substitute its own review. Otherwise stop the workflow.

Evidence: Claude’s current docs say an @-mention ensures that subagent runs. Open reports #80988 and #81263 inspect 2.1.219/2.1.220 and reproduce an Opus 5 dynamic prompt that both discourages unrequested delegation and conflicts with a sibling background-job instruction.

Caveat: Anthropic has not confirmed the reports, feature flags may vary by account, and natural-language naming merely encourages delegation. Even @-mentions do not prove review quality; retain a deterministic verifier and inspect the child receipt.

Compact source notes

  1. Superpowers v6.2.0 (release date 2026-07-23; published 2026-07-24 UTC), plus versioned workspace, brief, and review-package scripts. Maintained release, runnable artifacts, failure history, and documented eval counts.
  2. Claude Code subagent documentation (retrieved 2026-07-27). Official current reference distinguishing natural-language requests from an @-mention that ensures invocation.
  3. Claude Code issue #80988 (opened 2026-07-24 UTC; updated 2026-07-25) and #81263 (opened 2026-07-25 UTC). Open, inspectable binary/flag-state reproductions; no Anthropic confirmation yet.
  4. Method: anchor lens—versioned scripts and reproductions; unity lens—one structural invariant, “critical transitions need identities and receipts.” The two acceptance canaries are the observations that would falsify transfer to a given harness.