Skip to content

v1.3.0 — backend-neutral anti-churn guardrails

Latest
Latest

Choose a tag to compare

View all tags
@Nanako0129 Nanako0129 released this 20 Jul 10:19
v1.3.0
bd6552f
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Backend-neutral anti-churn guardrails

pilotfish v1.3.0 closes orchestration failure modes observed in long remora sessions without treating GPT-5.6 field data as native-Claude optimization evidence.

The policy now uses conditions rather than fixed numeric thresholds:

  • Batch recurring Execution work only when the remaining items are independent, the same shape, and completely describable by one stable brief with explicit ownership and per-item acceptance.
  • Treat already-diagnosed review findings with known remedies as Execution work while keeping diagnosis, exceptions, integration, and final acceptance in the main session.
  • Run fresh outcome verification at the smallest coherent integration boundary where the completed claim can be independently refuted.
  • Treat tests, builds, and static checks as intermediate evidence rather than a universal replacement for fresh verification.
  • Verify security, cross-language or FFI, serialization or pre-aggregation, irreversible, and integration-blocking changes earlier.
  • Never resubmit a substantially unchanged Plan; simplify it, surface the blocker, or defer blocked scope when readiness review does not converge.

Evidence boundaries

The motivating field report measured two remora sessions routed to GPT-5.6. Those observations identify recurring-work and review-loop churn, but do not establish native-Claude thresholds, efficiency gains, latency improvements, cost reductions, or an A/B result.

A fresh native first-party Claude compatibility Gate validated the exact v1.3.0 policy, role payload, and prompts under Claude Code 2.1.215 with Fast mode off:

Phase Result
Discovery + Plan Baton loaded; two background scouts; zero pre-approval writes; read-only Opus plan-verifier returned READY.
Approved execution Sonnet mech-executor wrote only REPORT.md; npm test passed.
Outcome verification Fresh Opus verifier returned CONFIRMED.
Totals 323.978 s wall time; 458.056 s API time; 6 API turns; 2 CLI invocations; $3.5088455 client-reported cost.

This Gate establishes compatibility and exact-byte provenance only. Failed and superseded candidate runs remain disclosed rather than being silently discarded.

Evidence and regression improvements

  • Preserve v1.2.1 as previous release evidence and restore v1.2.0 under a separate historical summary.
  • Represent every complete or interrupted schema-v3 CLI invocation as an additive top-level record.
  • Record both repository prompt-file hashes and shell-normalized runtime-input hashes.
  • Hash the committed final agents snapshot directly and compare it byte-for-byte with the generated payload.
  • Publish the bilingual remora/GPT-5.6 field attribution report and align all current Gate summaries.

Verification

  • All 19 dependency-free policy tests passed.
  • results.json passed JSON validation and exact additive-total checks.
  • The release policy matched the runtime-tested snapshot byte for byte.
  • The committed agents payload matched its recorded SHA-256 and generated template payload.
  • git diff --check passed.
  • A fresh outcome verifier returned CONFIRMED.
  • Codex returned a clean review on the final PR head with no unresolved review threads.

Upgrade

git clone --branch v1.3.0 --depth 1 https://github.com/Nanako0129/pilotfish.git
cd pilotfish
claude

Then ask Claude Code:

Read the local file install/AGENT-INSTALL.md in the current checkout and follow
it to update pilotfish in my global Claude Code configuration.

Show me the complete plan and wait for my approval before writing anything.

Restart Claude Code after installation.

Full Changelog: v1.2.1...v1.3.0

Assets 2
Loading