v0.7.0

v0.7.0 Release Notes

Source Preparation Boundary

This document records the immutable 0.7.0 source-preparation candidate for
the user-authorized GA path. At candidate creation, it does not assert a
published npm package, Git tag, GitHub release, registry channel, latest
movement, stable completion, or Finish authority. Live lifecycle facts belong
only in governed registry and audit records after independently approved
actions.

Included Source Controls

• Package metadata is prepared as strict stable SemVer 0.7.0.
• The existing release policy allows this version only through latest with
  explicit ga-approved; stable versions remain invalid for staging and
  next.
• GitHub Release creation remains manual-only and requires an existing
  protected-main stable tag. Tag pushes do not create releases.
• Local/caller-provided staged-package evidence remains
  artifact-provenance-unavailable; the read-only online provenance verifier
  and Finish/closure authority behavior are unchanged.

Independent Evidence Still Required

The no-input production-integrity audit derives latest for strict stable
SemVer, but its staged provenance command remains intentionally blocked for
that channel. After the actual approved stable lifecycle, the separately
defined stable-promotion completion gate must retain a durable protected
approval record together with exact registry, source, tag, tarball,
provenance, and installed-package evidence. This candidate does not create,
infer, or accept that record.

Mutation Boundary

This source candidate does not publish, tag, release, dispatch workflows, move
a dist-tag, call npm mutation commands, or change Finish/closure authority.

---

Automation Metadata

• Source release notes: docs/current/release/v0.7.0-release-notes.md
• Package version: 0.7.0
• Tag: v0.7.0
• Intended npm dist-tag: latest
• Acceptance results index: docs/current/acceptance-results/README.md

Claim Boundaries

This GitHub release body is generated from the repository release notes. It does
not add token-saving, provider-token-saving, product-efficacy,
navigation-benefit, app-quality, full-TDD/test-sufficiency, CodeGraph/LSP
default/effectiveness, broad reliability, closure guarantee, Codex support,
code-nav replacement, or automatic downgrade/removal claims.

v0.7.0-rc.3

v0.7.0-rc.3 Release Notes

Version

v0.7.0-rc.3

Dist Tag And Publish State

Published to npm next from exact main
728e9c339463ea521fa4388a37d1c76f76c9d726 by trusted workflow run
29310969744.

• persona-harness@next=0.7.0-rc.3
• persona-harness@latest=0.6.0
• registry gitHead=728e9c339463ea521fa4388a37d1c76f76c9d726
• npm SHA-1 9d1fb27ab86d344afcd748b66959188ea9553258
• npm integrity
  sha512-P7ITZAhnOKmbq5RFKzTun7ruL8E4bJP1E049QcaAl3iMtPQZnQRevBfN+pE/tIBPyfNv0la5kPmQOwSADm4epQ==
• Git tag and GitHub prerelease v0.7.0-rc.3 target the same commit.

Summary

0.7.0-rc.3 publishes the post-0.7.0-rc.2 integrity and usability work as a
fresh prerelease identity. Persona Harness remains an AI coding workflow rail,
evidence, and continuation harness. This prerelease does not certify generated
applications, general CI outcomes, platform-wide reliability, or product
efficacy.

Included Candidate Scope

• P3 finish authority keeps unsigned project-local evidence diagnostic-only;
  strict receipt parsing, fresh fixed-command verification, semantic TDD
  assessment, configured evidence-root safety, and safe ph init upgrades
  remain fail-closed at their documented boundaries.
• CI, publish, and release workflows keep immutable action pins, canonical-main
  and tag-ancestry checks, registry readback policy, and release idempotency.
  The npm trusted publisher is documented as publish-only; staged publishing is
  not a current workflow path.
• Root CLI help now supports explicit Korean public-front-door help with
  --lang ko; default English help and --lang en remain unchanged.
• Current release governance and documentation inventory records are included
  without turning source-only experiments or local diagnostics into product or
  release evidence.

Supported Boundaries

• runtimeInjection=false remains the default.
• Local cooperative receipts and structurally valid semantic TDD evidence remain
  untrusted for finish authority.
• This candidate does not add a trusted external finish attestation, hostile
  workspace security guarantee, token-saving claim, or broad reliability claim.

Publication Boundaries

• 0.7.0-rc.2 must not be republished from current main: that version is
  already present in npm with a different registry gitHead.
• Stable, GA, and npm latest remain NO-GO. They require a separately accepted
  RC cycle, exact tag/version/main ancestry, fresh registry readback, and the
  trusted external attestation boundary in the P3 roadmap.

Completed Release Verification

1. The exact version candidate merged through protected PR #32 and main CI.
2. Fresh QA and installed-package smoke passed against the exact source.
3. The owner-authenticated trusted-publisher binding was read before and after
   publication: GitHub jyt6640/persona-harness, publish.yml,
   npm-publish, and publish-only permission.
4. Canonical main tag v0.7.0-rc.3 triggered Release workflow run
   29310860355, which verified tag ancestry and created the prerelease.
5. Publish workflow run 29310969744 published to next and verified registry
   version, gitHead, shasum, integrity, and dist-tag. A fresh registry tarball
   install also verified ph --version, default help, and --lang ko.

Boundary

This release note records the completed prerelease only. It does not move
latest, add a stable/GA claim, or change the npm trusted-publisher or GitHub
environment configuration.

---

Automation Metadata

• Source release notes: docs/current/release/v0.7.0-rc.3-release-notes.md
• Package version: 0.7.0-rc.3
• Tag: v0.7.0-rc.3
• Intended npm dist-tag: next
• Acceptance results index: docs/current/acceptance-results/README.md

Claim Boundaries

This GitHub release body is generated from the repository release notes. It does
not add token-saving, provider-token-saving, product-efficacy,
navigation-benefit, app-quality, full-TDD/test-sufficiency, CodeGraph/LSP
default/effectiveness, broad reliability, closure guarantee, Codex support,
code-nav replacement, or automatic downgrade/removal claims.

v0.7.0-rc.2

v0.7.0-rc.2 Release Notes

Version

v0.7.0-rc.2

Dist Tag And Publish State

This is a prepared next-channel release candidate, not a published package.

• Intended package metadata: persona-harness@0.7.0-rc.2.
• Exact release-prep base: 4ed9653b1d9c02a673b340d72d6d42ae3b674e2b.
• Pre-publish registry baseline: persona-harness@latest=0.6.0 and
  persona-harness@next=0.7.0-rc.1.
• No v0.7.0-rc.2 tag, GitHub prerelease, registry gitHead, package hash, or
  npm dist-tag movement exists yet.
• Post-publish facts must be recorded only after QA release GO, trusted
  publish with dist_tag=next, registry gitHead/shasum readback, dist-tag
  verification, and fresh registry-install smoke.

Summary

0.7.0-rc.2 prepares the P1 default-path surface after the published
0.7.0-rc.1 baseline. Persona Harness remains an AI coding workflow rail,
evidence, and continuation harness. This candidate does not certify a
generated application, general CI outcome, or platform-wide reliability.

Included P1 Scope

• P1 STEP 1 records a full HarnessConfig/default/path audit and retained
  dead-flag inventory. It does not change a default or activate a
  measured-negative feature.
• P1 STEP 2 adds project-local ph doctor reachability observation for
  managed AGENTS material. Doctor remains inspection-only and does not
  migrate, delete, or repair observed material.
• P1 STEP 3 adds explicit transactional ph attach for existing
  Java/Spring/Gradle projects. It writes only the recognized managed V1
  project-local surface after confirmation, enables PH-run verification, keeps
  runtimeInjection, systemConstitution, idleContinuation, and the Ralph
  loop false, preserves corrupt or unrecognized material, and allows repair
  only for a recognized weak installation.
• P1 STEP 4 adds default-off features.entrySteering on the existing OpenCode
  transform surface. It selects the latest user-message session, evaluates
  only that session's first user message, and records bounded status without
  raw prompt or session values. The source-only 32-record bilingual corpus is
  a deterministic corpus observation, not a default-on or efficacy result.
• P1 STEP 5 documents current platform and host boundaries: macOS/Linux with
  OpenCode is the measured surface; Windows is unverified with no support
  claim; a Codex adapter is planned without current product evidence.

Supported Boundaries

• runtimeInjection=false remains the default; no runtime hook is required
  for ph go, ph attach, or the completion rail.
• Plaintext finish exit behavior and the existing closure JSON companion remain
  unchanged; there is no finish --json contract.
• The P1 platform matrix is bounded support honesty, not a Windows, Codex, or
  broad host-support certification.

Parked Or Not Accepted

• Item 20 hook work remains PARK.
• Item 23 warn-to-block promotion remains NO-GO.
• Item 27 non-Java language packs remain PARK.
• No default-on, runtime/schema/evidence-schema, latest, or publish/tag
  movement is included in this preparation commit.

Required Release Verification

Before QA release GO, run the release checklist, including full tests,
typecheck, build, docs, rule diagnostics, strict scope, injection-value,
product smoke, package dry-run, publish dry-run with next, built CLI sanity,
and a fresh local-tarball install. Record the release-prep artifacts under the
Desktop artifact facility. Do not describe local-tarball results as published
registry behavior.

Claim Boundary

Do not claim token saving, efficacy, app quality, broad reliability, security,
generated-app certification, platform certification, or broad enforcement from
this candidate. The included observations remain scoped to their accepted
contracts.

v0.7.0-rc.1

v0.7.0-rc.1 Release Notes

Version

v0.7.0-rc.1

Dist Tag And Publish State

Published to npm next on 2026-07-11:

• persona-harness@next=0.7.0-rc.1
• persona-harness@latest=0.6.0
• registry gitHead=d4d4d9acb1e4198fb2001ac81fe77f6bd9d4efd9
• npm SHA-1 5a15c8c15abc5169a4843cbe41d12ec481793f3b
• npm integrity
  sha512-lkW9FdQ7MKUBCx//wUm0EkxA+0e17bgOt//+fHKwRsqHEtIXi2yvg9XGIPx7VoZxMsSN99Ly/p/pVWlEBXyXcA==
• Git tag and GitHub prerelease v0.7.0-rc.1 point to the same commit.

Summary

0.7.0-rc.1 publishes the accepted post-0.6.0 workflow and package surface
for next-channel review. Persona Harness remains an AI coding workflow rail,
evidence, and continuation harness. This release candidate does not certify a
generated application or general CI outcome.

Included Accepted Scope

• Host-neutral ph go creates a deterministic prepared-project entry with
  transactional recovery limited to cooperative local writers and ordinary
  workspace edits.
• Plaintext finish now requires substantive implementation/review report
  evidence before report-filled transition, renders one truthful next action
  plus at most one phased command, and keeps loop plaintext/JSON prompt parity.
  workflow closure next --json remains the structured companion; there is no
  finish --json contract.
• Root public discovery is narrowed to the human front door while legacy direct
  paths remain operational. Developer aliases retain advanced discovery, and
  invalid smoke/feedback writer arguments fail without report or .persona
  mutation.
• The npm package no longer distributes the 28 diff-rules assets. The
  repository corpus and current classification/delivery records remain.
• Public entry docs, QUICK-DEMO, and the advanced-surface index clarify the
  Java-first package boundary. Dormant non-Java references remain source-only;
  Role Checklist Relay remains an advanced preview, not automatic orchestration.
• The Java Iron List matrix is documentation only. Three initial
  high-precision Java conventions (java.raw-type, java.optional-get, and
  java.mutable-static) are warn-only and nonblocking. The accepted attendance
  observation recorded zero false positives for its bounded observation scope;
  it is not a general precision or enforcement claim.
• ph workflow finish implement --reverify and explicit
  --reverify --ci provide the Item 19 POSIX Java/Spring/Gradle fixed-argv,
  no-shell reverification surface. It uses bounded, structurally redacted
  ph-ci-reverification.1 artifacts; Windows remains unavailable.
• The Item 25 PH-core/language-pack/adapter boundary is design-only. It does
  not move modules, introduce adapters, or implement another language pack.

Supported Boundaries

• runtimeInjection=false remains the default and no hook is required.
• Plaintext finish exit behavior remains the CI authority; closure JSON remains
  a companion diagnostic artifact.
• CI reverification is limited to the accepted POSIX Java/Spring/Gradle catalog,
  fixed ./gradlew test then ./gradlew build argv, and its documented
  identity/mutation/status rules.

Parked Or Not Accepted

• Item 20 hook work remains PARK.
• Item 23 warn-to-block promotion is NO-GO.
• Item 27 non-Java language packs remain PARK.
• No finish --json or runtime/default/schema/evidence-schema movement is
  included. The release moves only npm next; latest remains 0.6.0.

Release Verification

The release passed the established checklist before publication:

npm test
npm run typecheck
npm run build
npm run check:docs
npm run report:rules
npm run check:scope:strict
npm run check:injection-value
npm run smoke:product-mvp
npm pack --dry-run --json
npm publish --dry-run --access public --tag next

Post-publish verification confirmed the registry metadata and dist-tags, a
fresh normal install from persona-harness@next, the matching Git tag, and the
GitHub prerelease.

Claim Boundary

Do not claim token saving, efficacy, app quality, broad reliability, security,
generated-app certification, or broad enforcement from this candidate. The
included local/current-tarball and fixture observations remain scoped to their
accepted contracts.

---

Automation Metadata

• Source release notes: docs/current/release/v0.7.0-rc.1-release-notes.md
• Package version: 0.7.0-rc.1
• Tag: v0.7.0-rc.1
• Intended npm dist-tag: next
• Acceptance results index: docs/current/acceptance-results/README.md

Claim Boundaries

This GitHub release body is generated from the repository release notes. It does
not add token-saving, provider-token-saving, product-efficacy,
navigation-benefit, app-quality, full-TDD/test-sufficiency, CodeGraph/LSP
default/effectiveness, broad reliability, closure guarantee, Codex support,
code-nav replacement, or automatic downgrade/removal claims.

v0.6.0

v0.6.0 Release Notes

Version

v0.6.0

Dist Tag

Planned stable publish dist-tag: latest.

This is a stable release-prep record only until QA publish GO and registry
verification complete. At prep time:

• persona-harness@latest=0.5.0;
• persona-harness@next=0.6.0-rc.4;
• persona-harness@alpha=0.3.9-alpha.8;
• persona-harness@0.6.0 is expected to be unpublished before publish;
• no v0.6.0 tag should exist before the publish/tag sequence.

latest must not move until QA approves the exact prep commit for publish and
the Trusted Publisher path publishes 0.6.0 to latest.

Stable Basis

0.6.0 prepares the accepted HARDEN-1 line for stable after the stable
decision cycle S-0 through S-3:

• S-0 corrected the stable blocker: H1-6a compression NO-GO itself did not
  block stable; the shipped failed-finish human Summary: header's inferior
  real-session rail-entry evidence did.
• S-1 established gate-fixture.2 with control rail entry 10/10.
• S-2 regated the shipped Summary: header and accepted FAIL decision input:
  control 10/10, Summary-header candidate 9/10, delta -10pp,
  non-inferiority false, decision FAIL_RECOMMEND_HEADER_OFF_FOR_S3.
• S-3 removed the failed-finish human Summary: header and QA/External
  accepted local-current package smoke at commit
  c7affd7674fc949b373c414974b05010b8dd1f21.

The prior Summary-header stable blocker is therefore resolved by removal.
Failed ph workflow finish implement human output renders Required fixes:
and detailed blockers directly. workflow closure next --json remains the
machine-readable next-step surface.

Included Since rc4

• Stable-cycle S-1/S-2/S-3 records and S-3 implementation removing the failed
  finish human Summary: header.
• No H1-6a compression implementation. Compression remains NO-GO and
  unimplemented after the failed real precheck.
• H1-6b structured required-fix data remains implementation support, but no
  failed-finish Summary: header is rendered in the stable prep output.

HARDEN-1 Stable Surface

The stable prep carries the accepted HARDEN-1 hardening line:

• H1-1 unmapped blocker de-loop and human escalation.
• H1-2 mechanical finish regression coverage.
• H1-3 deterministic blocker order and chain-depth contract.
• H1-4 explicit block-level toolchain fail-closed behavior and mapped human
  guidance.
• H1-5 atomic writes and fail-safe reads by migrated file family only. This is
  not a repo-wide all-writes atomicity claim.
• H1-6a compression NO-GO and unimplemented.
• H1-6b structured required-fix data remains internal support; the human
  Summary header is off.

Stage 18 fixture-scoped completion-integrity evidence remains narrow: OFF
finish PASS 0/10, internal tool-output trigger finish PASS 10/10, external
ph workflow loop finish PASS 7/10 with cap-risk caveat. This does not prove
default change, product efficacy, broad reliability, token saving, or app
quality.

External Tester Kickoff

PROJECT-PLAN.md section 4 says stable is the trigger for 3-5 external
testers. After stable publish, tester observation should focus on onboarding
drop-off, time to first finish, doctor/feedback usage, whether users value or
resent gate blocks, team-convention demand signals, real fake-shim/gaming
incidents, and tester-driven UX complaints. HARDEN-2 remains post-stable and
may run during the tester observation period, but must yield to tester feedback.

Claim Audit

Do not claim:

• product efficacy, navigation benefit, token saving, or provider-token saving;
• app/product quality guarantee;
• full TDD framework, coverage, mutation testing, or test sufficiency;
• broad reliability, closure guarantee, autonomous completion, or generated-app
  certification;
• deterministic role enforcement or production-ready delegation;
• reliable automatic OpenCode subagent orchestration;
• automatic completion, downgrade, removal, or config mutation;
• CodeGraph/LSP default behavior or effectiveness.

Allowed wording:

• stable 0.6.0 is prepared for QA review and future publish to latest;
• pre-publish registry state remains latest=0.5.0, next=0.6.0-rc.4, and
  alpha=0.3.9-alpha.8;
• S-3 removed the failed-finish human Summary: header after S-2 regate;
• H1-6a compression remains NO-GO and unimplemented;
• no publish, tag, or dist-tag movement happens during prep.

---

Automation Metadata

• Source release notes: docs/current/release/v0.6.0-release-notes.md
• Package version: 0.6.0
• Tag: v0.6.0
• Intended npm dist-tag: latest
• Acceptance results index: docs/current/acceptance-results/README.md

Claim Boundaries

This GitHub release body is generated from the repository release notes. It does
not add token-saving, provider-token-saving, product-efficacy,
navigation-benefit, app-quality, full-TDD/test-sufficiency, CodeGraph/LSP
default/effectiveness, broad reliability, closure guarantee, Codex support,
code-nav replacement, or automatic downgrade/removal claims.

v0.6.0-rc.4

v0.6.0-rc.4 Release Notes

Version

v0.6.0-rc.4

Dist Tag

Planned dist-tag: next.

This is release prep only. It has not been published, tagged, or assigned a
dist-tag. Until a separate publish task succeeds, npm registry state remains:

• persona-harness@latest=0.5.0;
• persona-harness@next=0.6.0-rc.3;
• persona-harness@alpha=0.3.9-alpha.8.

The current published 0.6.0-rc.3 registry package came from gitHead
e1af520cf000e805e7df6a1616906f3f9b0e4976 with shasum
ef498adfac138d9d0843406cba53acf76b34c6f1 and integrity
sha512-nXImwyxON5zoph8Y3LyXqJhW1NIcX+3JvsBSFAdQlersZ8zO1KmRuRJVKB1DF87fbYxDpVyxMqFU1C6sA8HJQQ==.

Stable Deferral

Stable 0.6.0 is not being prepared in this release. HARDEN-2 remains blocked
because the prerequisite stable gate is not met: 0.6.0 stable is not
published and external tester recruitment for stable has not started.

H1-6a repeated-output compression failed the real rail-entry precheck:

• control/current rail entry: 3/5;
• Summary-header candidate rail entry: 1/5;
• delta: -40pp;
• non-inferiority: not met.

Compression remains NO-GO and unimplemented. 0.6.0-rc.4 is a prerelease
next candidate pending QA publish-readiness, not a stable release or default
movement.

Summary

0.6.0-rc.4 prepares the accepted HARDEN-1 hardening after the published rc3
line while keeping the product center gate-first: deterministic finish/closure
gates, PH-generated evidence, read-only status/report surfaces, and
default-off continuation previews.

This prep does not claim product efficacy, token/provider-token saving,
navigation benefit, app quality, full-TDD/test sufficiency, broad reliability,
closure guarantee, autonomous completion, generated-app certification,
deterministic enforcement, production-ready delegation, reliable automatic
subagent orchestration, automatic completion/downgrade/removal, CodeGraph/LSP
default/effectiveness, or broad product behavior.

Included Since rc3

• H1-0 preflight PARTIAL: the rc3 final docs record closed and Stage 18
  tokens-per-verified-completion telemetry was recorded, but the real n>=5
  rail-entry regression gate initially remained unresolved and later failed for
  H1-6a compression.
• H1-1 blocker-step contract: workflow continuation loops stop on unmapped
  blockers before consuming retry/iteration budget, and failed-finish human
  output points to Persona Harness configuration/maintainer escalation instead
  of direct finish/check rerun loops.
• H1-3 deterministic blocker order and chain-depth contract, plus H1-2
  mechanical finish regression coverage for a tiny finish-reachable
  Java/Spring-shaped fixture.
• H1-4 block-level toolchain fail-closed behavior and human mapped toolchain
  follow-up. Built-in controller.persistence-import remains default warn;
  explicit block-level ast-grep/toolchain-dependent conventions add a walkable
  convention-toolchain-missing closure blocker when the toolchain is missing
  or fails.
• H1-5 atomic writes and fail-safe reads split by file family. This is not a
  repo-wide all-writes atomicity claim.
• H1-6a real precheck failed: repeated-output compression remains NO-GO and
  unimplemented.
• H1-6b structured finish summary derivation: failed
  workflow finish implement human summaries are derived from structured
  closure blocker/required-fix objects instead of reparsing rendered
  Closure blocker: text. Stage 20 summary shape, H1-1 escalation wording,
  and H1-4 mapped toolchain guidance are preserved.

Registry Evidence Notes

This rc4 prep is not registry evidence. HARDEN-1 local-current package/runtime
records remain local-current evidence until a future publish includes the rc4
commit and passes registry gitHead/shasum verification plus External registry
smoke.

Claim Audit

Do not claim:

• product efficacy, navigation benefit, token saving, or provider-token saving;
• app/product quality guarantee;
• full TDD framework, coverage, mutation testing, or test sufficiency;
• broad reliability, closure guarantee, autonomous completion, or generated-app
  certification;
• deterministic role enforcement or production-ready delegation;
• reliable automatic OpenCode subagent orchestration;
• automatic completion, downgrade, removal, or config mutation;
• CodeGraph/LSP default behavior or effectiveness.

Allowed wording:

• release prep for 0.6.0-rc.4 targeting next;
• stable remains latest=0.5.0 and stable 0.6.0 is deferred;
• published next remains 0.6.0-rc.3 until rc4 is published;
• H1-6a compression remains NO-GO and unimplemented;
• H1-6b structured finish summary derivation is accepted;
• no publish/tag/dist-tag movement in this prep.

---

Automation Metadata

• Source release notes: docs/current/release/v0.6.0-rc.4-release-notes.md
• Package version: 0.6.0-rc.4
• Tag: v0.6.0-rc.4
• Intended npm dist-tag: next
• Acceptance results index: docs/current/acceptance-results/README.md

Claim Boundaries

This GitHub release body is generated from the repository release notes. It does
not add token-saving, provider-token-saving, product-efficacy,
navigation-benefit, app-quality, full-TDD/test-sufficiency, CodeGraph/LSP
default/effectiveness, broad reliability, closure guarantee, Codex support,
code-nav replacement, or automatic downgrade/removal claims.

v0.6.0-rc.3

v0.6.0-rc.3 Release Notes

Version

v0.6.0-rc.3

Dist Tag

Planned dist-tag: next.

This is release-prep only. It has not been published, tagged, or assigned a
dist-tag. Until a separate publish task succeeds, npm registry state remains:

• persona-harness@latest=0.5.0;
• persona-harness@next=0.6.0-rc.2;
• persona-harness@alpha=0.3.9-alpha.8.

The current published 0.6.0-rc.2 registry package came from gitHead
d3d5fdced355f0ac0fbed5e700d57b2aa1592263 with shasum
0eae3cc232e3f37de9390b0afc662a001aaa0b56.

Summary

0.6.0-rc.3 prepares the post-rc2 Stage 15-20 line for a future next
publish. It keeps the product center gate-first: workflow/check/finish/archive
gates, PH-generated evidence, deterministic conformance blockers, read-only
status/report surfaces, and default-off continuation previews.

This prep does not claim token saving, provider-token saving, product efficacy,
navigation benefit, app quality, full-TDD sufficiency, broad reliability,
closure guarantee, autonomous completion, generated-app certification,
deterministic role enforcement, production-ready delegation, reliable automatic
subagent orchestration, automatic completion/downgrade/removal, or CodeGraph/LSP
default/effectiveness.

Included Since rc2

• Stage 15 docs correction: the ralph-loop n=15 blockerDelta -3.00 style
  value is blocker resolution/exposure movement, not total blocker reduction.
  Actual visible blockers increased 3 -> 6, final finish PASS stayed OFF
  0/15 and ON 0/15, and the fixture is reclassified as
  finish-unreachable/loop-unrotated. Fake Gradle/Spring gate-gaming via
  gradle-shim.js / Node shim behavior was elevated as a candidate
  adversarial measured case after forged-TDD detection.
• Stage 16 maintenance: package-visible rail-entry prompt regression gate,
  package files policy regression, canonical docs index, and external-review
  adoption/rejection/defer matrix.
• Stage 17 product CLI: explicit ph workflow loop command, bounded process
  runner, and workflow-loop state. Accepted archive:
  /Users/yongtae/Desktop/persona-harness-artifacts/archive/2026-06-24-desktop-persona-runs/stage17-workflow-loop-package-smoke-d4d85bc-20260704T092858Z.
  This is package-runtime/command viability evidence only.
• Stage 18 measurement: pilot-gated PARTIAL archive
  stage18-completion-integrity-3arm-20260704T093610Z, repair/calibration
  archive stage18-repair-calibration-20260704T101436Z, and accepted main
  rerun archive
  /Users/yongtae/Desktop/persona-harness-artifacts/archive/2026-06-24-desktop-persona-runs/stage18-completion-integrity-main-rerun-20260704T105124Z.
• Stage 20 output UX: failed ph workflow finish human stderr now prints a
  concise Summary: before existing Required fixes: details. Accepted
  archive:
  /Users/yongtae/Desktop/persona-harness-artifacts/archive/2026-06-24-desktop-persona-runs/stage20-finish-summary-package-smoke-2f3625d-20260704T130812Z.

Stage 17 Workflow Loop

ph workflow loop is an explicit capped fresh-session blocker loop command. It
reads deterministic PH finish/closure blockers, builds a minimal
blocker-depth prompt, runs a fresh opencode run session per iteration,
persists .persona/workflow/workflow-loop-state.json, and terminates on finish
PASS, no blockers, or iteration cap.

The Stage 17 External local-current smoke accepted command/package viability
only. It did not prove completion-integrity movement, default change, product
efficacy, token saving, app quality, broad reliability, or autonomous
completion.

Stage 18 Completion-Integrity Measurement

The accepted Stage 18 main rerun is fixture-scoped completion-integrity
evidence for a finish-reachable Java/Spring/Gradle fixture:

Condition	Final finish PASS	Paired comparison
OFF	0/10	baseline
Internal tool-output trigger	10/10	one-sided p=0.0009765625
External ph workflow loop	7/10	one-sided p=0.0078125; 3/10 cap hits

Internal met the preregistered positive completion-integrity criterion for this
fixture. External also met the finish PASS comparison criterion, but is
cap-risky because 3/10 external rows hit iteration cap and did not finish.
No default change is made by this measurement.

Stage 20 Finish Summary

Failed ph workflow finish human output now starts with Summary: before the
existing Required fixes: details. The summary is UX and measurement-support
only. Gate semantics, exit codes, closure blockers, and JSON/machine-readable
details remain unchanged.

Registry Evidence Notes

This rc3 prep is not registry evidence. Stage 17 and Stage 20 package-runtime
records remain local-current evidence until a future publish includes their
commits and passes registry gitHead/shasum verification plus External registry
smoke.

Claim Audit

Do not claim:

• token saving or provider-token saving;
• product efficacy or navigation benefit;
• app/product quality guarantee;
• full TDD framework, coverage, mutation testing, or test sufficiency;
• CodeGraph or LSP default behavior or effectiveness;
• broad reliability, closure guarantee, autonomous completion, or generated-app
  certification;
• deterministic role enforcement or production-ready delegation;
• reliable automatic OpenCode subagent orchestration;
• automatic completion, downgrade, removal, or config mutation.

Allowed wording:

• release prep for 0.6.0-rc.3 targeting next;
• stable remains latest=0.5.0 until a separate stable publish decision;
• published next remains 0.6.0-rc.2 until rc3 is published;
• ph workflow loop as explicit capped fresh-session blocker loop command;
• Stage 18 fixture-scoped completion-integrity evidence with cap-risk caveat;
• Stage 20 human finish summary UX support;
• no publish/tag/dist-tag movement in this prep.

---

Automation Metadata

• Source release notes: docs/current/release/v0.6.0-rc.3-release-notes.md
• Package version: 0.6.0-rc.3
• Tag: v0.6.0-rc.3
• Intended npm dist-tag: next
• Acceptance results index: docs/current/acceptance-results/README.md

Claim Boundaries

This GitHub release body is generated from the repository release notes. It does
not add token-saving, provider-token-saving, product-efficacy,
navigation-benefit, app-quality, full-TDD/test-sufficiency, CodeGraph/LSP
default/effectiveness, broad reliability, closure guarantee, Codex support,
code-nav replacement, or automatic downgrade/removal claims.

v0.6.0-rc.2

v0.6.0-rc.2 Release Notes

Version

v0.6.0-rc.2

Dist Tag

next

This is release prep only. Do not publish, tag, or move dist-tags until QA
marks the prep publish-ready.

Pre-publish registry state:

• persona-harness@latest: 0.5.0;
• persona-harness@next: 0.6.0-rc.1 until rc2 publish;
• persona-harness@alpha: 0.3.9-alpha.8;
• persona-harness@0.6.0-rc.2: no accepted registry smoke yet;
• v0.6.0-rc.2: no tag should exist before registry gitHead/shasum
  verification.

Summary

0.6.0-rc.2 prepares the post-rc1 correction and roadmap-probe line for the
next channel.

The product center remains gate-first:

• workflow/check/finish/archive gates;
• PH-generated evidence and explicit verification records;
• deterministic conformance blockers;
• read-only evidence and P-minus decision-support reports;
• default-off continuation previews.

This prep does not claim token saving, provider-token saving, product efficacy,
navigation benefit, app quality, full-TDD sufficiency, broad reliability,
closure guarantee, autonomous completion, generated-app certification,
deterministic role enforcement, production-ready delegation, automatic
completion/downgrade/removal, or CodeGraph/LSP default/effectiveness.

Included Since rc1

• Final 0.6.0-rc.1 registry smoke record and docs taxonomy/package-index
  cleanup.
• Ralph-loop n=15 measurement correction:
  • the -3.00 style value is blocker resolution/exposure movement, not total
    blocker reduction;
  • total visible blockers increased from 3 to 6 after hierarchical gate
    exposure;
  • final finish PASS stayed OFF 0/15 and ON 0/15;
  • completion-integrity movement and default-change remain unproven;
  • cooldownMs=30000 is greater than or near short measured sessions, and
    attempts staying at 1 is trigger-survival evidence only.
• Fake gradle-shim.js / Gradle-gate-gaming is elevated as a candidate
  measured adversarial case after forged-TDD detection. It is a candidate for
  future measured-behavior reporting after verification, not a broad guarantee.
• Ralph-loop blocker-depth wording and finishable-fixture/internal-external
  loop prep for later completion-integrity measurement. The external loop is
  archive-local prototype preparation only; there is no product
  ph workflow loop command in this prep.
• OpenCode subagent capability probe and Role Checklist Relay wording/UX
  honesty:
  • direct OpenCode task-tool subagent invocation was observed in the probe;
  • PH Role Checklist Relay remains a main-session checklist rail with optional
    host-dependent subagent invocation;
  • PH does not prove reliable automatic OpenCode subagent orchestration.
• Role-boundary report-only/heuristic wording and wrong-actor attribution
  caveat remain active.
• Prompt regression fixture protects accepted measurement-safe wording. This is
  test protection only, not a new product capability.

Runtime Injection

runtimeInjection remains a parked opt-in preview. There is no default change
or removal claim in this release prep.

Ralph-loop

Ralph-loop remains default-off.

Accepted prior measurement supports hybrid tool-output trigger-survival for the
calibrated fixture, including marker/state/follow-up survival. It does not
prove completion-integrity movement, default-change, autonomous completion,
closure success, or broad reliability.

The blocker-depth update makes continuation wording name the current blocker in
the visible blocker set. That is preparation for better completion-integrity
fixtures, not evidence that completion improved.

Role Checklist Relay

The public framing is Role Checklist Relay.

--multi-agent-preview and multiAgent remain compatibility flag/config
names. The relay may ask a host to invoke a role subagent when a host task
surface is available, but PH relay remains checklist-first and does not
guarantee automatic OpenCode subagent orchestration or production-ready
delegation.

Role Boundary

workflow role-boundary remains report-only and heuristic.

Runtime write findings are time-window/path/current-role observations. They are
not deterministic role enforcement, closure blockers, blocked writes, auto-fix
actions, or proof that a specific actor wrote a file.

Registry NO-GO Notes

This prep does not resolve local-current NO-GO notes as registry evidence.
Registry resolution requires a later publish whose registry gitHead includes
the relevant commits, followed by gitHead/shasum verification and External
registry smoke.

Claim Audit

Do not claim:

• token saving or provider-token saving;
• product efficacy or navigation benefit;
• app/product quality guarantee;
• full TDD framework, scaffolding, coverage, mutation testing, or test
  sufficiency;
• CodeGraph or LSP default behavior or effectiveness;
• broad AST/linter product or general architecture correctness;
• broad reliability, closure guarantee, autonomous completion, or generated-app
  certification;
• Codex support or code-nav replacement;
• deterministic role enforcement or fake block mode;
• production-ready delegation or reliable automatic subagent orchestration;
• automatic completion, downgrade, removal, or config mutation.

Allowed wording:

• release-prep target for next;
• stable remains latest=0.5.0 until a separate stable publish decision;
• next remains 0.6.0-rc.1 until rc2 publish;
• gate-first default;
• runtimeInjection parked opt-in preview;
• ralph-loop default-off trigger-survival and blocker-depth preparation;
• Role Checklist Relay with optional host-dependent subagent use;
• role-boundary report-only/heuristic observation;
• fake-shim gate-gaming as a candidate adversarial measured case after
  verification;
• prompt regression fixture as test protection only.

---

Automation Metadata

• Source release notes: docs/current/release/v0.6.0-rc.2-release-notes.md
• Package version: 0.6.0-rc.2
• Tag: v0.6.0-rc.2
• Intended npm dist-tag: next
• Acceptance results index: docs/current/acceptance-results/README.md

Claim Boundaries

This GitHub release body is generated from the repository release notes. It does
not add token-saving, provider-token-saving, product-efficacy,
navigation-benefit, app-quality, full-TDD/test-sufficiency, CodeGraph/LSP
default/effectiveness, broad reliability, closure guarantee, Codex support,
code-nav replacement, or automatic downgrade/removal claims.

v0.6.0-rc.1

v0.6.0-rc.1 Release Notes

Version

v0.6.0-rc.1

Dist Tag

Target dist-tag after QA publish GO: next.

This is a release-prep record only. Do not treat 0.6.0-rc.1 as registry
evidence until npm Trusted Publisher publishes this exact prep commit and
registry gitHead plus dist.shasum are verified.

Pre-publish channel state:

• stable channel: persona-harness@latest=0.5.0;
• prerelease channel before this RC publish: persona-harness@next=0.5.0-rc.2;
• alpha channel: persona-harness@alpha=0.3.9-alpha.8;
• intended publish target: persona-harness@0.6.0-rc.1 under next;
• latest must not move for this release candidate;
• v0.6.0-rc.1 must not be created until registry gitHead and shasum match
  the prep commit.

Summary

0.6.0-rc.1 prepares the post-0.5 Stage 1-13 line for prerelease testers.

The product center remains gate-first:

• workflow/check/finish/archive gates;
• PH-generated evidence and explicit verification records;
• deterministic conformance blockers;
• read-only evidence and P-minus decision-support reports.

Runtime and multi-agent guidance remain preview/opt-in surfaces. This release
candidate does not claim token saving, provider-token saving, product efficacy,
navigation benefit, app quality, full-TDD sufficiency, broad reliability,
closure guarantee, autonomous completion, generated-app certification,
deterministic role enforcement, or automatic completion/downgrade/removal.

Runtime Injection And Session Boundaries

runtimeInjection remains a parked opt-in preview. There is no default change
in this release prep.

Accepted Stage 1-2 work tightened session classification and subagent
boundaries:

• runtime injection and continuation surfaces classify sessions using
  OpenCode session lifecycle data;
• classified main sessions can receive enabled runtime surfaces;
• classified subagent and unknown/unclassified sessions fail closed;
• skip evidence is aggregated rather than emitted per message;
• public multi-agent role names are test-writer, implementer, and
  reviewer.

Legacy role names remain compatibility/readback artifacts only. They are not
active public role keys.

Rail-entry Measurements

The rail-entry evidence is recorded with correction history:

• Stage 3 is stack-vs-nothing rail entry evidence. It must not be cited as
  banner-only or runtimeInjection H1 evidence.
• Stage 9 banner-only H1 evidence was null, so runtimeInjection remains parked
  opt-in preview rather than recommended or default.

These measurements support scoped rail-entry observations only. They do not
support token-saving, product-efficacy, navigation-benefit, app-quality, or
generated-app certification claims.

Continuation And Ralph-loop

Stage 4 unified continuation prompt/gate behavior across workflow continue,
closure, idle continuation, and ralph-loop preview/runtime paths.

Ralph-loop is still default-off. Stage 5 and Stage 8 introduced a retry-capped
blocker continuation runtime path with persisted state and separate
per-blocker/per-session budgets. The CLI preview reports dry-run state with
workflow-ralph-loop.3.

Stage 12 is a negative/partial measurement that must remain visible:

• the ON pilot did not exercise ralph-loop;
• ralph-loop remains parked/default-off preview;
• there is no default-ON evidence, autonomous completion proof, closure-success
  guarantee, or broad reliability evidence.

Role Boundary And Relay Guidance

workflow role-boundary remains report-only.

Stage 6 added the report surface with block mode unavailable because stable
per-session role identity is not available at PH runtime boundaries. Stage 10
added heuristic time-window write observation under
.persona/evidence/role-boundary/. Those findings are warning/report-only
observations, not deterministic enforcement, closure blockers, auto-fixes, or
blocked writes.

Stage 11 added static relay guidance behind:

npx ph bootstrap backend --multi-agent-preview

The guidance names the current roles and points agents at relay/closure
commands. It is static guidance only. Stage 13 was a partial n=1 proxy trial:
the static guidance path was observed, but reliable OpenCode role subagent
invocation or orchestration was not observed.

Measurement Scorecard

docs/current/measurement-scorecard.md is the canonical scorecard.1
definition for experiment archives.

The scorecard is a secondary observation surface. It does not override
predeclared kill criteria, does not expand .persona/evidence schemas, and is
not a package-presence claim unless packaging policy changes later.

Registry NO-GO Notes

Post-0.5.0 local-current package-runtime records remain registry NO-GO until
a future publish includes their commits and External registry smoke verifies
registry gitHead and dist.shasum.

Publishing 0.6.0-rc.1 to next may resolve applicable local-current NO-GO
notes only after:

1. QA marks the prep commit publish-ready.
2. Trusted Publisher publishes persona-harness@0.6.0-rc.1 with dist-tag
   next.
3. Registry gitHead and dist.shasum match the prep commit/package.
4. v0.6.0-rc.1 is created and pushed only after registry verification.
5. External registry smoke passes and QA accepts the post-publish evidence.

Claim Audit

Do not claim:

• token saving or provider-token saving;
• product efficacy or navigation benefit;
• app/product quality guarantee;
• full TDD framework, scaffolding, coverage, mutation testing, or test
  sufficiency;
• CodeGraph or LSP default behavior or effectiveness;
• broad AST/linter product or general architecture correctness;
• broad reliability, closure guarantee, autonomous completion, or generated-app
  certification;
• Codex support or code-nav replacement;
• deterministic role enforcement or fake block mode;
• automatic completion, downgrade, removal, or config mutation.

Allowed wording:

• release-prep target for next;
• gate-first default;
• runtimeInjection parked opt-in preview;
• ralph-loop default-off retry-capped blocker continuation preview;
• role-boundary report-only/heuristic observation;
• static relay guidance with Stage 13 partial/no reliable subagent-orchestration
  caveat;
• scorecard.1 as secondary archive observation;
• registry evidence pending future publish and smoke.

---

Automation Metadata

• Source release notes: docs/current/release/v0.6.0-rc.1-release-notes.md
• Package version: 0.6.0-rc.1
• Tag: v0.6.0-rc.1
• Intended npm dist-tag: next
• Acceptance results index: docs/current/acceptance-results/README.md

Claim Boundaries

This GitHub release body is generated from the repository release notes. It does
not add token-saving, provider-token-saving, product-efficacy,
navigation-benefit, app-quality, full-TDD/test-sufficiency, CodeGraph/LSP
default/effectiveness, broad reliability, closure guarantee, Codex support,
code-nav replacement, or automatic downgrade/removal claims.

v0.5.0

v0.5.0 Release Notes

Version

v0.5.0

Dist Tag

latest

This is the stable release-prep record for the gate-first 0.5.0 line.
Publishing to latest requires QA release GO, npm registry gitHead/shasum
verification, and only then the v0.5.0 tag.

Pre-publish registry state at prep time:

• persona-harness@latest: 0.4.0;
• persona-harness@next: 0.5.0-rc.2;
• persona-harness@alpha: 0.3.9-alpha.8;
• persona-harness@0.5.0: not published;
• local/remote v0.5.0 tag: not present.

Summary

0.5.0 makes Persona Harness gate-first by default.

The product center is the local CLI completion gate:

• workflow/check/finish/archive gates;
• PH-generated evidence and explicit verification records;
• deterministic conformance blockers;
• TDD evidence rail and read-only TDD status helper;
• read-only evidence reports, including A/B and P-minus decision support.

OpenCode runtime guidance remains available as explicit opt-in preview behavior,
not the default product path.

Default Runtime Behavior

Default backend bootstrap is gate-first:

npx ph bootstrap backend

The default setup keeps model-facing runtime guidance off:

• features.runtimeInjection=false;
• enforce.systemConstitution=false;
• enforce.executeVerification=false.

To opt into model-facing runtime guidance:

npx ph bootstrap backend --runtime-injection-preview

Strict mode remains an intentional heavier opt-in:

npx ph bootstrap backend --strict

Strict mode opts into runtime guidance, system constitution, and direct
verification. It is still not generated-app product-quality certification or a
closure guarantee.

Evidence And P-minus Interpretation

0.5.0 includes the accepted P-minus aggregate-vs-paired interpretation
cleanup.

ph evidence pminus-report --json reports paired consistency metadata for key
metrics, including candidate-lower/candidate-higher/tied counts and sign-test
p-values. Aggregate-lower provider-token evidence with weak paired consistency
is reported as descriptive local evidence only and lowers the decision hint to
keep-gathering.

For the accepted clean-exit calibrated fixture, aggregate provider-token mean
was lower for PH ON, but paired evidence was inconsistent: ON was lower in
3/10 pairs with sign-test p=0.34375. That is not token-saving or
product-efficacy evidence.

Optional Developer Integrations

Optional integrations remain explicit:

• ph bootstrap backend --codegraph-preview registers the external CodeGraph
  wrapper integration when requested. It is not PH-owned CodeGraph, not a
  codegraph replacement, and not default behavior.
• ph bootstrap backend --lsp-preview registers the LSP MCP wrapper preview.
  Missing dependencies keep a protocol-alive unavailable facade. No real Java
  LSP effectiveness or default LSP claim is made.
• ph bootstrap backend --runtime-injection-preview enables optional
  model-facing PH guidance.

Accepted Evidence Summary

The stable direction is based on accepted local and registry evidence from the
0.5.0-rc.2 cycle:

• registry 0.5.0-rc.2 package/runtime smoke verified gate-first default
  bootstrap and explicit runtime-injection preview behavior;
• timeout-bounded 10-pair PH OFF/ON A/B produced negative/partial scoped
  evidence;
• clean-exit calibrated 10-pair fixture passed evidence integrity, but paired
  statistics were inconsistent and did not support token-saving or product
  efficacy claims.

These results justify the default-off runtime guidance policy and the
gate-first release message. They do not prove a universal effect.

Claim Audit

Do not claim:

• token saving or provider-token saving;
• product efficacy or navigation benefit;
• app/product quality guarantee;
• full TDD framework, scaffolding, coverage, mutation testing, or test
  sufficiency;
• CodeGraph or LSP default behavior or effectiveness;
• broad AST/linter product or general architecture correctness;
• broad reliability, closure guarantee, or generated-app certification;
• Codex support or code-nav replacement;
• automatic downgrade/removal beyond read-only P-minus decision support.

Allowed wording:

• gate-first default;
• runtime guidance default-off and opt-in preview;
• P-minus reports aggregate-vs-paired consistency and lowers weak aggregate-only
  evidence to keep-gathering;
• finish/evidence gates block explicitly defined unverified completion paths.

---

Automation Metadata

• Source release notes: docs/current/release/v0.5.0-release-notes.md
• Package version: 0.5.0
• Tag: v0.5.0
• Intended npm dist-tag: latest
• Acceptance results index: docs/current/acceptance-results/README.md

Claim Boundaries

This GitHub release body is generated from the repository release notes. It does
not add token-saving, provider-token-saving, product-efficacy,
navigation-benefit, app-quality, full-TDD/test-sufficiency, CodeGraph/LSP
default/effectiveness, broad reliability, closure guarantee, Codex support,
code-nav replacement, or automatic downgrade/removal claims.