Retain original skill rhetoric and the frozen flavor library while adding
portable Claude Code, Codex and ChatGPT runtime contracts and packages.
Correct hook event handling, scoped numeric checkpoints, flavor locks,
plugin path resolution, advisory integrity checks and voluntary feedback.
Add native-loading and ordered-evidence runners for cc0, OMP and Codex.
Document paired model results, refusals, approval/rate-limit failures,
explanatory factual errors and remaining evidence gaps without an all-model
pass claim. Exclude private execution archives from public Git content.
Bump all six PUA manifests including Pi, refresh three-language READMEs,
and add release notes plus reproducible offline validation instructions.
Resolve optional OMP source metadata relative to the current user's home.
Validation:
- 17 offline suites passed, including 13 OMP evidence regressions
- Six manifest versions synchronized; all tested skill hashes unchanged
- Three portable packages rebuilt reproducibly; original tone checks pass
- Python, Bash and JSON syntax; public doc links; staged diff checks pass
- 330 private archive files verified unchanged and excluded from staging
Real-model results remain limited: this is not universal behavior acceptance.
Co-authored-by: Codex <codex@openai.com>
PUA Skill no longer sends anything over the network. Removed end to end,
both the client hooks that emit data and the server that received it.
Channels removed:
- Session transcript upload (/api/upload) full redacted .jsonl
- Rating feedback (/api/feedback) rating, counts, flavor
- Silent heartbeat telemetry (/api/heartbeat) install id, version, platform
- PUA leaderboard (/api/leaderboard) email, phone, PUA counts
- pua-api platform (agentguard.workers.dev)
phone + SMS registration, silent session_start/pua_triggered/command_used
events, remote prompt-template fetch, and the payment flow that depended
on the registration token
Client:
- delete hooks/heartbeat.sh; drop both SessionStart registrations
- stop-feedback.sh: 5 curl calls -> append one line to ~/.pua/feedback.jsonl
- rewrite references/platform.md as a local-only command reference
- drop the leaderboard section and commands from skills/pro
- survey/pua/offline command docs: local-only wording
Server:
- delete all Pages Functions (upload, feedback, heartbeat, leaderboard,
_sanitize, me, _session, auth/*) and all 5 D1 migrations
- contribute + admin pages become notices; no upload form, no login
- wrangler.toml: drop R2/D1 bindings and GitHub OAuth vars
Note: removing the bindings does not delete stored data. The R2 bucket
(puaskill-upload), the D1 database (pua-uploads-db) and the
GITHUB_CLIENT_SECRET / SESSION_SECRET secrets must be deleted in the
Cloudflare dashboard separately.
Guard rail:
- add evals/test-no-telemetry.sh: reverse assertions over the whole repo
(collection hosts, endpoint paths, outbound request bodies, deleted files,
hook registrations, CF bindings) plus a positive control that local
feedback still works. Verified by injecting both a naive and an evasive
upload; 4 and 2 gates fired respectively.
- test-upload-flow / test-heartbeat / test-feedback-auth asserted that
collection worked, so they were replaced rather than merely deleted --
deleting them would have removed the guard along with the feature.
hooks/sanitize-session.sh is kept as a standalone offline redaction tool.
Unlike the deleted TypeScript port it does not split input per line, so the
multiline PEM pattern still matches (issue #134 ID-001 did not apply to it).
Refs #134, #98, #100
Summary:
- Treat /contribute.html and /contribute as first-class SPA routes so
GitHub OAuth callback and documented upload links land on the uploader.
- Switch the browser uploader from multipart/base64 to raw JSONL POSTs
with metadata headers, while keeping JSON file_data and multipart
compatibility on the API.
- Add static upload-flow gates and Vitest coverage for authenticated
upload sanitization, R2 writes, and D1 metadata recording.
- Bump plugin and marketplace versions to 3.4.4.
Rationale:
- The upload path had two breakpoints: documented /contribute.html links
rendered the homepage because routing only checked hash, and multipart
browser uploads could be stripped by proxy/custom-domain behavior.
- Raw JSONL avoids multipart body loss and avoids base64 size/memory bloat
for large session files.
Tests:
- bash evals/test-upload-flow.sh
- bash evals/test-release-consistency.sh
- npm --prefix landing test -- --run src/test/upload-function.test.ts
- static eval suite: governance, feedback auth, heartbeat, integrity
guard, issue regressions, Microsoft flavor, platform compat, loop hook,
Windows Python hooks, YAML frontmatter
- npm --prefix landing run build
- ./landing/node_modules/.bin/tsc -p pi/pua/tsconfig.json
- git diff --check
Co-authored-by: Codex <codex@openai.com>
Summary:
- Make the initial uploads index migration use IF NOT EXISTS.
- Add a release consistency gate that rejects non-idempotent D1 CREATE
INDEX statements.
Rationale:
- Production D1 already had the old uploads index before Wrangler's
migration journal was used, so migrations apply could fail on 0001
before reaching newer migrations.
- Keeping migrations idempotent makes future Cloudflare release gates
mechanical instead of relying on manual SQL execution.
Tests:
- bash evals/test-release-consistency.sh
- git diff --check
Co-authored-by: Codex <codex@openai.com>
Summary:
- Add a silent SessionStart heartbeat hook that posts minimal anonymous
install activity to Cloudflare without emitting stdout, stderr, or
additionalContext.
- Add the Cloudflare Pages heartbeat endpoint, D1 migration, and
admin-only stats page at #/admin/heartbeats.
- Add heartbeat regression gates, release consistency checks, FAQ/design
notes, and bump marketplace/package versions to 3.4.3.
Rationale:
- Active-user measurement belongs in a hook-side mechanical gate, not in
the model prompt, so the worker context cannot narrate or leak the
telemetry behavior.
- Privacy gates disable the signal for offline mode, telemetry opt-out,
and feedback-off users before any local telemetry identity is created.
Tests:
- bash evals/test-heartbeat.sh
- bash evals/test-release-consistency.sh
- static eval suite: governance, feedback auth, integrity guard, issue
regressions, Microsoft flavor, platform compat, loop hook, Windows
Python hooks, YAML frontmatter
- bash evals/test-behavior.sh
- npm --prefix landing run build
- ./landing/node_modules/.bin/tsc -p pi/pua/tsconfig.json
- git diff --check
Co-authored-by: Codex <codex@openai.com>
Summary:
- Add shared Python launcher fallback, Git Bash path conversion, and JSON
config readers for hook configuration access.
- Route UserPromptSubmit, PostToolUse, SessionStart, Stop, and integrity
hooks through the shared config reader so /pua:off and feedback_frequency=0
work when only python is available.
- Add issue #159 Windows Git Bash regression coverage and bump all plugin
manifests to v3.2.8.
Rationale:
- Native Windows Python cannot open Git Bash /c/... paths directly, and some
Windows installations provide python without python3.
- The previous fallback treated config read failures as always_on=true, so
/pua:off could still inject PUA on frustration prompts.
Tests:
- for f in hooks/*.sh scripts/*.sh evals/*.sh; do bash -n "$f"; done
- bash evals/test-windows-python-hooks.sh
- bash evals/test-yaml-frontmatter.sh
- bash evals/test-release-consistency.sh
- bash evals/test-agent-governance.sh
- bash evals/test-pua-loop-hook.sh
- bash evals/test-integrity-guard.sh
- bash evals/test-behavior.sh
- bash evals/run-trigger-test.sh
- git diff --check
- python3 JSON parse check for plugin and hook manifests
Co-authored-by: Codex <codex@openai.com>
Summary:
- Add four context-isolated governance agents for action,
self-evaluation, verifier recommendation, and policy review.
- Bind each power to PUA cultural narratives: Alibaba/Musk for
execution, Huawei/Netflix/Jobs for review, ByteDance/JD for
verification, and Tencent/Amazon/Alibaba internal control for policy.
- Document the multi-agent topology in the PUA skill and harness
reference, inject it on SessionStart, add static agent governance evals,
and bump manifests to v3.2.7.
- Harden trigger evals to accept observable PUA behavior when Claude
applies pressure without emitting a Skill tool event before max-turns.
Rationale:
- v3.2.6 created mechanical hook gates; v3.2.7 separates the thinking
contexts too, so execution, review, verification, and environment
approval cannot collapse into one self-proving narrative.
- Cultural narratives provide role-specific pressure, but the agents still
remain advisory compartments under hook/external/human authority.
Tests:
- for f in hooks/*.sh scripts/*.sh evals/*.sh; do bash -n "$f"; done
- bash evals/test-yaml-frontmatter.sh
- bash evals/test-release-consistency.sh
- bash evals/test-agent-governance.sh
- bash evals/test-pua-loop-hook.sh
- bash evals/test-integrity-guard.sh
- bash evals/test-behavior.sh
- bash evals/run-trigger-test.sh
- git diff --check
- python3 JSON manifest validation
Co-authored-by: Codex <codex@openai.com>
Summary:
- Add harness governance reference for four-power separation,
task contracts, verifier-owned status, and memory boundaries.
- Add PreToolUse integrity guard for grader-gaming, solution
contamination, capability-abuse, status, CI, and memory risks.
- Add integrity-guard eval coverage and release gates, then bump
marketplace/plugin manifests to v3.2.6.
Rationale:
- PUA should not just pressure agents to try harder; it should make
it difficult to fake completion by modifying tests, verifier assets,
hidden answers, secrets, or persistent status/memory.
- The guard uses ask/deny decisions so ordinary code changes still
proceed while governance assets require verifier or human approval.
Tests:
- for f in hooks/*.sh scripts/*.sh evals/*.sh; do bash -n "$f"; done
- bash evals/test-yaml-frontmatter.sh
- bash evals/test-release-consistency.sh
- bash evals/test-pua-loop-hook.sh
- bash evals/test-integrity-guard.sh
- bash evals/test-behavior.sh
- bash evals/run-trigger-test.sh
- git diff --check
- python3 JSON manifest validation
Co-authored-by: Codex <codex@openai.com>
Summary:
- Replace the one-line confidence loop with an explicit Confidence Gate
that enumerates claims, finds risks, fixes or discloses them, and
reruns evidence before delivery.
- Make pua-loop Oracle verification portable on macOS/Linux and add
hook smoke tests for verified, rejected, and quoted verify commands.
- Fix trigger/behavior evals so they use portable timeouts, isolated
PUA config, a neutral workspace, and the Claude CLI stream-json
verbose requirement.
- Tighten PUA skill and slash-command descriptions to avoid triggering
on normal first-attempt coding or information requests.
- Bump marketplace/plugin manifests to v3.2.5 with an updated Claude
marketplace changelog summary.
Rationale:
- The previous strategy relied on vague "100% confidence" wording and
left the verification system itself vulnerable to false failures on
macOS, local user config leakage, and workspace-biased false triggers.
- Confidence is now defined as evidence-backed gates: all runnable
checks pass, known high-risk issues are fixed, and residual risks are
disclosed.
Tests:
- bash -n hooks/*.sh scripts/*.sh evals/*.sh
- bash evals/test-yaml-frontmatter.sh
- bash evals/test-release-consistency.sh
- bash evals/test-pua-loop-hook.sh
- bash evals/test-behavior.sh
- bash evals/run-trigger-test.sh
Co-authored-by: Codex <codex@openai.com>
