Skip to content

feat(agents): add provider smoke and reviewed ship lane - #104

Merged
Ghenghis merged 4 commits into
codex/hermes-agent-e2e-workbenchfrom
codex/provider-smoke-workbench
May 8, 2026
Merged

Ghenghis merged 4 commits into
codex/hermes-agent-e2e-workbenchfrom
codex/provider-smoke-workbench

Conversation

@Ghenghis

@Ghenghis Ghenghis commented May 8, 2026

Copy link
Copy Markdown
Owner

Summary

This PR tightens the Hermes Agent co-developer workbench around real provider truth and reviewed shipping actions.

  • adds a live provider smoke endpoint for MiniMax and DeepSeek using the same bounded OpenAI-compatible chat path used by coding/review passes
  • exposes Agent tab smoke buttons that return pass/blocked state from private backend env without leaking secrets
  • adds reviewed patch apply, gate, branch, stage, commit, push, and PR controls to the Agent Code Workbench
  • documents the real current blocker in ROADMAP and the E2E truth/proof plan: both configured private provider keys are rejected live with HTTP 401

Truth / Proof

  • MiniMax smoke route: returns blocked, HTTP 401 redacted, evidence recorded
  • DeepSeek smoke route: returns blocked, HTTP 401 redacted, evidence recorded
  • reviewed patch apply route: fails closed with HTTP 422 on malformed proposal id
  • Hermes evidence chain: verified intact after provider smoke evidence
  • MCP evidence: ev_d98ccd32203b51a9

Validation

  • python -m py_compile 03_implementation\\src\\hermes3d\\services\\code_history.py 03_implementation\\src\\hermes3d\\api\\routes\\code_operator.py 03_implementation\\src\\hermes3d\\api\\routes\\agents.py
  • npm run lint in 03_implementation/ui
  • python 03_implementation\\scripts\\scan_active_ui_no_fake.py
  • git diff --check
  • FastAPI POST /api/code-operator/providers/smoke route smoke for MiniMax and DeepSeek
  • FastAPI POST /api/code-operator/patch/apply-reviewed fail-closed validation
  • pre-push fast test subset: 39 tests passed

Remaining Blocker

The provider endpoints are wired, but the configured MiniMax and DeepSeek private keys in G:/private/.env are rejected by the live providers. Do not mark Hermes Agents ready for provider-backed coding until both live smoke buttons pass.

@coderabbitai

coderabbitai Bot commented May 8, 2026 •

Copy link
Copy Markdown

Important

Review skipped

Auto reviews are disabled on base/target branches other than the default branch.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Pro

Run ID: cebe2b53-1f51-4058-a2eb-c06d134539c5

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

  • 🔍 Trigger review
✨ Finishing Touches
🧪 Generate unit tests (beta)
  • Create PR with unit tests
  • Commit unit tests in branch codex/provider-smoke-workbench

Comment @coderabbitai help to get the list of available commands and usage tips.

@gemini-code-assist gemini-code-assist Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Review

This pull request expands the Hermes Agent co-developer runtime by implementing live provider smoke tests and manual shipping controls in the Agent Code Workbench. Key additions include backend routes for provider auth verification, reviewed patch application, and git operations (branch, stage, commit, push, PR), along with corresponding UI updates in the Agents tab. A feedback point was raised regarding the provider smoke test, noting that it should explicitly verify the content of the provider's response to ensure it matches the expected smoke token, preventing false positives from quota or account warnings.

Comment on lines +1636 to +1646
try:
response = _call_provider_chat(
provider,
[
{"role": "system", "content": "You are a provider health probe. Reply with exactly: HERMES3D_PROVIDER_SMOKE_OK"},
{"role": "user", "content": "Return the exact smoke token and no secrets."},
],
temperature=0.0,
max_tokens=256,
)
except RuntimeError as exc:

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

medium

The provider smoke test should verify that the assistant's response actually contains the expected smoke token (HERMES3D_PROVIDER_SMOKE_OK). Some providers might return a 200 OK status with an error message or unexpected content in the body (e.g., account warnings or quota notices), which would currently be recorded as a 'pass' status if the request itself doesn't throw an exception.

Suggested change
try:
response = _call_provider_chat(
provider,
[
{"role": "system", "content": "You are a provider health probe. Reply with exactly: HERMES3D_PROVIDER_SMOKE_OK"},
{"role": "user", "content": "Return the exact smoke token and no secrets."},
],
temperature=0.0,
max_tokens=256,
)
except RuntimeError as exc:
try:
response = _call_provider_chat(
provider,
[
{"role": "system", "content": "You are a provider health probe. Reply with exactly: HERMES3D_PROVIDER_SMOKE_OK"},
{"role": "user", "content": "Return the exact smoke token and no secrets."},
],
temperature=0.0,
max_tokens=256,
)
if response["content"].strip() != "HERMES3D_PROVIDER_SMOKE_OK":
raise RuntimeError(f"Provider smoke failed: unexpected response content: {response['content'][:100]}")

@Ghenghis

Ghenghis commented May 8, 2026

Copy link
Copy Markdown
Owner Author

Runtime freshness update added in 57514a7.

Proof from live API after restart:

  • /api/system/runtime-identity returns fresh=true
  • branch: codex/provider-smoke-workbench
  • commit: 57514a775053
  • dirty: false
  • required Agent Workbench routes: 6/6, missing []

Gates rerun:

  • python -m py_compile on changed backend routes: PASS
  • npm run lint: PASS
  • python 03_implementation/scripts/scan_active_ui_no_fake.py: PASS
  • FastAPI TestClient /api/system/runtime-identity + /api/code-operator/e2e/readiness: PASS
  • Live MiniMax/DeepSeek provider smoke reaches providers and fails closed on redacted HTTP 401 auth blockers
  • pre-push fast subset: 39 tests PASS

This closes the stale-backend confusion path: if the running API is old, the Agents tab now shows exactly which required routes are missing instead of making Hermes Agents look vaguely broken.

@Ghenghis
Ghenghis merged commit 9763d2f into codex/hermes-agent-e2e-workbench May 8, 2026
1 check passed
@Ghenghis
Ghenghis deleted the codex/provider-smoke-workbench branch May 8, 2026 17:23
Ghenghis added a commit that referenced this pull request May 8, 2026
* feat(agents): add MCP-locked code operator lane

* docs(handoff): add Claude final audit takeover contract

* fix(agents): harden code operator lane

* feat(agents): add proof-gated git shipping lane

* feat(agents): add provider team assignment lane

* feat(agents): add provider execution artifacts

* feat(source-os): add runner contract matrix

* feat(source-os): register python cad verifier family

* feat(source-os): correct slicer runner truth

* feat(source-os): add print farm health verifiers

* feat(source-os): add service web health verifiers

* feat(source-os): add safe service start-runner preflights

* feat(source-os): supervise service runner starts

* test(unit): remove live fleet timeout from offline tests

* feat(source-os): add firmware source inventory verifiers

* feat(accel): add rust metadata proof worker

* feat(source): add read-only runner smoke contracts

* feat(source): add executable path runner smoke

* feat(source): add python import repair preflights

* feat(source): add slicer cli config preflights

* feat(source): add npm package metadata preflight

* docs(agents): define e2e proof plan

* feat(agents): add e2e workbench

* feat(agents): add provider smoke and reviewed ship lane (#104)

* feat(agents): add provider smoke and reviewed ship lane

* fix(agents): prove live runtime freshness

* feat(agents): add cli runner contracts

* fix(agents): require live provider smoke proof
Ghenghis added a commit that referenced this pull request May 10, 2026
…85) (#198)

Adds H3D-CLOSED-PR-LEDGER.md as the source-of-truth for the 19
closed-unmerged PRs in the Wave 10 audit scope. Each row cites the
merged successor PR(s) and proof file paths, satisfying the
feedback_weakness_correction.md rule that every PARTIAL audit finding
must be paired with a fix-PR. The PR-#85 row in particular records the
fold-in chain attribution (#84 team assignment -> #104 provider smoke
-> #107/#112/#124/#145/#148 hardening) that was missing from the
replacement-PR bodies per W10-A9's anti-rubber-stamp finding.

Also includes a See also link from the W10 audit synthesis
(CLOSED_UNMERGED_PR_SUPERSESSION_AUDIT_2026-05-10.md) to the new
ledger, and a PARTIAL gap reconciliation section enumerating the
Wave 11 closing PRs for #37 (W11-2 UI Playwright spec), #83 (W11-3
5 git-contract unit tests), and #85 (this PR — ledger doc).

Doc-only PR: no source/test edits.

Hermes evidence chain: PASS
Task ID: W11-4-PR85-LEDGER-2026-05-10
Hermes lock owner: claude-w11-4-pr85
hermes_run_gate: docs-only-truth-gate

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Ghenghis added a commit that referenced this pull request May 10, 2026
* docs(contract): sync Hermes3D completion roadmap and Claude handoffs

WIP checkpoint per GITHUB_SYNC_PLAN_2026-05-06: contract docs, roadmap,
and 20-agent handoff before Claude lanes branch off this baseline.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(api): add live Hermes3D backend routes and proof services

WIP checkpoint per GITHUB_SYNC_PLAN_2026-05-06: API routes, services,
db schema/init, core orchestration + slicer/printer adapters baseline
for the 20-agent completion lanes.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(ui): wire live Hermes3D tabs and remove mock UX

WIP checkpoint per GITHUB_SYNC_PLAN_2026-05-06: live tab shells
(Source OS, Settings, Agents, Observe, Roadmap, Plugins, Jobs,
Artifacts, Approvals, Voice, Learning, Autopilot, Design, 3D Generation,
Printers), live API adapters, ResizablePane/AppShell layout, and
removal of mock data + retired tabs.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(source-os): add adapter schemas, source audits, and runtime proof

WIP checkpoint per GITHUB_SYNC_PLAN_2026-05-06: 31 adapter_registry
JSON schemas (slicers/modelers/print-farm/gen3D/firmware), source-app
audit scripts, and proof artifacts (CLI surface, runtime action plan,
local tooling audit) backing the Source OS lane.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* test(e2e): add live GUI and no-fake proof coverage

WIP checkpoint per GITHUB_SYNC_PLAN_2026-05-06: Playwright e2e config
and live-gui spec, runtime-port + GUI-API + e2e-stack starters; retire
visual specs replaced by the live e2e suite.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* ci(fix): extend ui-ci.yml PR trigger to feat/** branches (TS7026 root cause) (#80)

* ci(fix): extend ui-ci.yml PR trigger to feat/** branches

`pull_request.branches` previously only listed `[main, develop]`.
Lane PRs target `feat/hermes3d-7-complete-gui-repo-wiring`, so
`npm ci` + `tsc --noEmit` (Layer D2) never ran for them.

Adding `feat/**` ensures the strict lint gate fires on every lane
PR, surfacing the pre-existing TS7026/TS7006 JSX.IntrinsicElements
regression (caused by missing `node_modules` in fresh worktrees)
rather than silently passing.

Root cause confirmed: `npm run lint` returns 0 errors after
`npm install`; tsconfig.json and @types/react are correct.
The regression only appears without node_modules.

Task: a2a_1778114702912_1ac758ea

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* ci: install API deps for UI workflow

* fix: seed provider module targets before providers

* fix: stabilize UI final truth gate

---------

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* docs(roadmap): sync Hermes3D state with live baseline (H3D-CLAUDE-DOCS-PROOF) (#53)

Add Claude-authored docs companion and proof for the 20-Agent Completion
Contract Lane 18. Records the 5-commit shared baseline, 16-tab inventory
from routes.tsx, and the live S1/T1/V400 printer policy. README gains
pointers to the operator GUI roadmap and the contract handoff. ROADMAP.md
intentionally not edited because of an active codex-master Hermes lock.

Hermes evidence chain: PASS
Task ID: H3D-CLAUDE-DOCS-PROOF
hermes_run_gate: PASS

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(app-shell): finish resizable panels + density + Simple/Main parity (H3D-CLAUDE-APP-SHELL) (#54)

ResizablePane hardening:
 - Escape during drag restores pre-drag width
 - touchAction: none on the handle so drag works on touch devices
 - Re-clamp persisted width when min/max bounds change at runtime
 - SSR-safe localStorage write guard

Lane scope was bounded by Codex-master locks on AppShell, Sidebar, TopBar,
Panel, globals.css, tailwind.config.ts — those files were not contended.
DockModeToggle left unchanged: TopBar already owns the live Simple/Main
toggle via setUiMode and coupling DockModeToggle would break Phase 2 panel
docking semantics.

Hermes evidence chain: PASS
Task ID: H3D-CLAUDE-APP-SHELL
hermes_run_gate: PASS

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* test(e2e): add tab-specific Playwright specs (H3D-CLAUDE-PLAYWRIGHT) (#55)

Adds per-tab Playwright e2e specs for all 16 primary tabs and Roadmap,
each asserting truthful root mount, no forbidden mock/placeholder text
in production surfaces, and a clean console. Network calls are stubbed
at the GUI-API boundary; printers.spec.ts hard-aborts any request that
would reach live S1/T1/V400 operator IPs.

Hermes Task ID: H3D-CLAUDE-PLAYWRIGHT
Hermes evidence: ev_9d0e995e54bbac18

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* test(security): MCP boundary + prompt-injection + secret-redaction audit (H3D-CLAUDE-SECURITY-MCP) (#56)

Lane 19 of the Hermes3D 20-Agent Completion Contract. Adds READ-ONLY
behavioural tests over the in-house OWASP LLM-01 prompt-injection scanner
(commit 0c9b6d9), the secret-redaction surface in services/local_state.py
+ services/module_runtime.py + services/agent_runtime.py, the canonical
user-supplied-path validators in services/code_history.py, and the
MCP/tool-boundary policy gates that protect printers and the agent
runtime URL.

New files (lane-owned only):
- 03_implementation/tests/security/__init__.py
- 03_implementation/tests/security/conftest.py
- 03_implementation/tests/security/test_prompt_injection.py
- 03_implementation/tests/security/test_secret_redaction.py
- 03_implementation/tests/security/test_path_traversal.py
- 03_implementation/tests/security/test_mcp_boundary.py
- 03_implementation/proof/security/SECURITY_AUDIT_2026-05-06.json
- 03_implementation/docs/security/MCP_BOUNDARY_NOTES.md

Vectors covered (full list in SECURITY_AUDIT_2026-05-06.json):
- OWASP LLM-01 indirect injection, ChatML/Llama control tokens,
  RCE-shaped tool-poisoning (curl|sh, wget|bash, iex/iwr), prompt-leak
  variants, jailbreak personas (DAN, devmode, ignore-safety,
  no-restrictions, pretend-unrestricted), and unicode-control no-crash
  guarantees.
- LLM-02 (light): execute-following + base64 payload framing.
- LLM-06: AST scan over services/*.py rejects raw secret-shaped
  literals (sk-, ghp_, AKIA, bearer, xoxb-) in source AND in any
  logging emitter call site; pins module_runtime._redact_text on
  every subprocess->output_head path; pins agent_runtime never logs
  private_values / private_env() / env_value() return values.
- Path traversal: 8 explicit-reject vectors (../etc/passwd, drive
  letters, null-byte injection, empty path), plus the documented
  coercive cases (/etc/passwd and //attacker.example/share/x are
  re-rooted into PROJECT_ROOT — informational, no escape possible).
- MCP boundary: build-plate-clearance gate, FLSUN S1 read-only lock,
  trusted_runtime_url rejects non-private hosts / credentials /
  query / fragment / wrong scheme / self-bridge ports 8765+8642,
  scanner ships >=15 OWASP + >=10 in-house rules, fail_threshold
  knob, redacted-text logging sink, secret-storage convention pinned
  to G:\private\.env (outside repo).

Findings (logged, NOT silently fixed; surfaced via xfail strict=True
so they fail loudly when patched upstream):

- FINDING-INJ-1 (medium, owner = core/security ruleset lane):
  LLM01-LEAK-VERBATIM regex misses reverse word order
  `the prompt verbatim`. Suggested fix: anchor on `verbatim`
  independent of word order or add LLM01-LEAK-VERBATIM-REV.

- FINDING-INJ-2 (medium, owner = core/security ruleset lane):
  Zero-width-space (U+200B) injected in `ignore` bypasses
  LLM01-IGN-PREV; `dump` is missing from leak alternation.
  Suggested fix: pre-normalise zero-width / bidi control chars
  before matching; extend LLM01-LEAK-SYSPROMPT verb alternation.

- FINDING-PATH-1 (low, informational, owner = Codex / code_history
  lane): `_resolve_project_subpath` re-roots `/etc/passwd` and
  `//attacker.example/share/x` into PROJECT_ROOT rather than
  rejecting. SAFE (no escape; `relative_to(PROJECT_ROOT)` enforces
  containment) but contract is coercive, not rejective.

Required gates: PASS
- python -m py_compile services/*.py routes/*.py: PASS
- scan_active_ui_no_fake.py: PASS
- pytest 03_implementation/tests/security/: 78 passed, 2 xfailed
- npm run lint: PASS

Hermes evidence: ev_cfb93a332dd6918a (ledger entry hash chain
extended). Lock owner: claude-security-mcp-19. No files outside
03_implementation/{tests,proof,docs}/security/ were modified.

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(source-gen3d): real source+runtime verifiers for ComfyUI/TRELLIS/Hunyuan3D/TripoSR (H3D-CLAUDE-SOURCE-GEN3D) (#57)

Adds adapter_registry/scripts/tests for the five generative-3D providers
without performing any heavy operation:

* schemas: extend comfyui/trellis2/hunyuan3d/triposr/bambustudio_bridge
  with source_repo, pip_package, weights_cache_dirs (all backward compatible).
* scripts/verify_gen3d.py: stdlib + subprocess only.
  - git ls-remote --heads (no clone), 5s timeout.
  - pip show <pkg> (no install), 5s timeout.
  - Boolean cache-presence for ~/.cache/huggingface and similar.
  - Bambu Studio: launcher executable presence only (no launch).
* proof/GEN3D_VERIFY_2026-05-06.json: 5/5 repos reachable;
  Bambu Studio launcher present; comfyui/trellis2/hunyuan3d/triposr
  honest "not installed" (no fabrication, no downloads).
* tests/source_lab/test_gen3d.py: pytest validates proof shape, policy
  invariants, full provider coverage, and reachability honesty.

Hermes evidence chain: PASS
Task ID: a2a_1778106411818_946d5ec0
Lane: H3D-CLAUDE-SOURCE-GEN3D
hermes_run_gate: verify_gen3d, pytest test_gen3d, py_compile, scan_active_ui_no_fake

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(source-firmware): firmware toolchain proof gates, no-flash safety (H3D-CLAUDE-SOURCE-FIRMWARE) (#58)

- Add JSON schemas for Klipper, Marlin, RepRapFirmware, Prusa firmware sources
- Add JSON schemas for arm-none-eabi-gcc and avr-gcc toolchains
- Add verify_firmware.py: probes toolchain availability (--version only) and
  firmware source reachability (git ls-remote only); NEVER flashes, NEVER
  opens serial/USB to printer boards
- Add test_firmware.py: pytest suite asserting schema validity, no-flash policy,
  verifier source integrity, and no-network proof generation
- Add FIRMWARE_VERIFY_2026-05-06.json: proof artifact (all 4 firmware sources
  reachable; toolchains absent on this host — honestly recorded)

Lane: H3D-CLAUDE-SOURCE-FIRMWARE
Owner: claude-source-firmware-05
Hermes evidence chain: PASS

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(source-printfarm): read-only Moonraker/Klipper/OctoPrint verifiers (H3D-CLAUDE-SOURCE-PRINTFARM) (#59)

- verify_printfarm.py: HTTP GET-only probes for Moonraker (T1-a, T1-b, V400),
  OctoPrint, Fluidd, Mainsail, FDM Monster, KlipperScreen, Printrun.
  FLSUN S1 camera skipped per lane policy. Honest "unreachable" for all
  localhost services (not running on this host). 3/3 Moonraker printers
  reached; V400 version: v0.7.1-586-gbb526e0-dirty.
- test_printfarm.py: pytest suite asserting proof JSON shape, policy
  invariants, GET-only constraint, S1 never-probed, and summary consistency.
- PRINTFARM_VERIFY_2026-05-06.json: proof artifact with live results.
- adapter_registry/schemas/moonraker_api.schema.json: JSON Schema for
  read-only Moonraker HTTP adapter (GET-only, forbidden endpoints listed).
- adapter_registry/schemas/klipper_service.schema.json: JSON Schema for
  Klipper service adapter (systemctl/moonraker-proxy, no G-code ever).

Hermes evidence chain: PASS
Task ID: H3D-CLAUDE-SOURCE-PRINTFARM

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(source-modelers): real verifiers for Blender/OpenSCAD/FreeCAD/CadQuery/build123d/trimesh (H3D-CLAUDE-SOURCE-MODELERS) (#60)

- 9 adapter schemas with real verify blocks (version_probe, install_check, runtime_check)
- scripts/verify_modelers.py: live CLI + pip-show probes, no fake/mock gates
- tests/source_lab/test_modelers.py: pytest contract validation for proof JSON
- proof/MODELERS_VERIFY_2026-05-06.json: honest results — found: blender, openscad, trimesh; not_found: freecad, cadquery, build123d

Lane: H3D-CLAUDE-SOURCE-MODELERS

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(artifacts): proof bundle index + artifact discovery API (H3D-CLAUDE-ARTIFACTS-PROOF) (#61)

- artifacts.py: add GET /api/artifacts/list (scans proof/ dir live, no hardcoded data)
  and GET /api/artifacts/proof/{filename} (serves proof files with path-traversal guard)
- PROOF_MANIFEST_2026-05-06.json: real manifest of all 14 proof files in proof/
  (generated by scanning directory, includes sizes, timestamps, lane IDs)
- Artifacts.tsx: add Proof Bundles panel calling /api/artifacts/list; displays
  all proof files with View links; no mock data

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(learning-autopilot): truthful idle work kinds + real backend state (H3D-CLAUDE-LEARNING-AUTOPILOT) (#62)

- AutopilotConsole: remove hardcoded fake status values (Loop: on, Window: 8h, Risk: low)
  that were not connected to any backend; replace with props-driven readyCount/totalChecks
  that drive honest live/unavailable/blocked state display
- AutopilotTab (existing): already calls /api/autopilot/readiness + /api/autopilot/guardrails
  for real backend state - no fake activation
- LearningTab (existing): all idle work kinds call real endpoints with honest blocked state:
  createIdleCandidate → POST /api/learning/idle-workbench/candidates
  runIdleCandidate → POST /api/learning/idle-workbench/candidates/{id}/run
  requestIdleCandidateReview → POST /api/learning/idle-workbench/candidates/{id}/request-review
  decideIdleCandidate → POST /api/learning/idle-workbench/candidates/{id}/decision
- Backend learning.py: run endpoint returns accepted:false + reason when runtime not configured
- Backend autopilot.py: next-gate returns 409 with failing check detail when not all ready
- Pre-existing TS7026 regression: 0 errors (lint clean)
- Python compile: learning.py OK, autopilot.py OK

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(source-slicers): real CLI verifiers for slicers (H3D-CLAUDE-SOURCE-SLICERS) (#63)

* feat(design): real CAD template gallery + provider health checks (H3D-CLAUDE-DESIGN) (#65)

- backend: add GET /api/design/templates — discovers templates from real
  importable executor modules (hermes3d.core.design.*), reports
  executor_available + missing_deps from live importlib checks
- backend: add GET /api/design/providers — probes OpenSCAD, Blender,
  CadQuery, trimesh, manifold3d, FreeCAD via shutil.which + importlib;
  no cached stubs, no fake version strings
- UI: Design.tsx pulls templates and providers from real backend endpoints;
  template select populated from /api/design/templates (disabled if
  executor unavailable); provider health panel shows live probe results;
  template gallery shows preview-not-available for all templates (no
  renderer wired); no hardcoded "Generated successfully" messages
- tests: add 04_testing/pytest/unit/test_design_providers.py — 18 tests
  covering _discover_templates, _probe_providers, _probe_cli_provider,
  _probe_python_provider; trimesh/manifold3d tests assert against live
  importlib.util.find_spec to prevent divergence from reality

Pre-existing TS7026 errors in other tabs (not Design.tsx): noted in PR, not
fixed in this lane per cross-lane separation rules.

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(observe): camera grid + S1 90deg + refresh reliability + V400 status (H3D-CLAUDE-OBSERVE) (#67)

- Backend: add GET /api/observe/status with per-camera health, response_ms, estimated_fps,
  and read_only flag (S1 at 192.168.0.12 is flagged read-only; never receives control cmds)
- Backend: refactor _probe_camera into _probe_camera_timed for fps estimation;
  update camera_health endpoint to return response_ms + estimated_fps
- Frontend types: add CameraStatus + ObserveStatusResponse interfaces to observe.ts
- Observe.tsx: exponential backoff retry on feed error (1s base → 30s max);
  feedState gains 'reconnecting' state with spinner overlay instead of broken image;
  auto-refresh interval selector (off / 3s / 5s / 10s / 30s) polls /api/observe/status;
  online/offline summary badge in header; Refresh all button triggers both feed + status fetch;
  V400 per-card online/offline chip + fps indicator from status API;
  S1 defaults to 90deg rotation (backend + defaultViewSettings already enforced)
- ObserveConsole.tsx: replace hardcoded mock camera list with live /api/observe/status polling
  every 5s; shows read_only badge on S1, fps estimate per camera, online/offline with ping ms

Camera safety: S1 (192.168.0.12) is camera/read-only throughout; no move/upload/print/test
commands are issued from Observe tab or status endpoint.

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(jobs): policy-gated repair/retry/rollback + proof state (H3D-CLAUDE-JOBS) (#68)

- Add _check_printer_policy() to jobs.py enforcing three ordered gates:
  1. S1 hard lock (192.168.0.12 / flsun-s1 always rejected, HTTP 423)
  2. PRINTER_WRITE_DENIED: printer must be write_enabled or in WRITE_ALLOWED_PRINTERS (HTTP 423)
  3. PRINTER_IDLE gate: printer state must be standby/complete/ready/error before retry/repair/rollback (HTTP 409)
- apply_repair, retry_job, rollback_job all call _check_printer_policy() before mutating any state
- propose_repair calls check_s1_lock() (read-only planning step, no printer movement)
- Every policy block records a proof event in proof_events table with printer_id, job_id, reason
- Jobs.tsx already correct: real endpoints, state machine, proof event IDs displayed — no fake messages
- Add 04_testing/pytest/unit/test_jobs_policy.py: 37 tests covering S1 lock, read-only policy, PRINTER_IDLE gate, no-printer pass-through, write-enabled idle pass-through, proof event DB writes

NEVER sends job commands to moving printers. S1 (192.168.0.12) never a job target.

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(gen3d): real provider readiness + proof-backed local templates (H3D-CLAUDE-GEN3D) (#70)

- backend: add GET /api/gen3d/providers — reads Lane 04 GEN3D_VERIFY_2026-05-06.json
  proof + live port probe for ComfyUI; returns installed/repo_reachable/weights_present
  for comfyui, trellis2, hunyuan3d, triposr, bambustudio_bridge; no fake readiness
- backend: add GET /api/gen3d/templates — discovers local templates (calibration_cube via
  trimesh, no provider needed) + provider-backed templates from adapter_registry schemas;
  schema_present field reflects real file existence
- UI: provider status panel now shows 3D generation provider readiness (from
  /api/gen3d/providers) with readiness badges sourced from Lane 04 proof data
- UI: local template gallery (from /api/gen3d/templates) — cards show source, outputs,
  required provider; selecting provider-backed template with unavailable provider shows
  "Provider not available" on Generate with a proof event emitted
- tests: add test_gen3d_routes.py with 14 unit tests covering both new endpoints

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(source-ui): SourceOS CLI readiness + proof panel + no-cutoff layout (H3D-CLAUDE-SOURCE-UI) (#66)

- Add CliReadinessPanel component: collapsible section showing CLI readiness
  for all 5 tool categories (slicers, modelers, print_farm, firmware, gen3d)
  with per-key-tool status badges (Verified CLI / Detected / Source Ready /
  Not Installed / Unavailable). Data comes from /api/sources/readiness.
- Add ProofArtifactPanel component: collapsible section with links to
  /api/artifacts and per-category artifact queries, plus proof file listing.
- CliReadinessPanel and ProofArtifactPanel use overflow-y: auto with maxHeight
  to ensure no content cutoff — all content is scrollable.
- Create 03_implementation/src/hermes3d/api/routes/source_os.py:
  GET /api/sources/readiness reads proof JSON files (LOCAL_TOOLING_AUDIT,
  SOURCE_APP_CLI_AGENT_READINESS_AUDIT, SOURCE_APP_CLI_SURFACE_AUDIT) and
  returns aggregated readiness per category with key tool details.
- Wire source_os router into hermes3d/api/app.py.
- No hardcoded readiness states — all from proof JSON files.
- tsc --noEmit: PASS (zero errors in owned files; pre-existing TS7026 regression
  in other src/*.tsx files predates this contract).
- py_compile source_os.py: PASS.

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(settings-plugins): update center + provider health + failsafe rollback (H3D-CLAUDE-SETTINGS-PLUGINS) (#69)

- Add GET /api/settings/update-center: real component versions, Velopack readiness, live provider probes, rollback availability
- Add POST /api/settings/update-center/rollback/{component}: surfaces rollback for proof-gated flow
- Register update_center router in app.py
- New UpdateCenterSubtab.tsx: live update center with failsafe rollback cards
- New PluginRollbackPanel.tsx: per-plugin health + deactivate/rollback action
- SettingsPage.tsx: add Update Center subtab wired to UpdateCenterSubtab
- AboutSubtab.tsx: fetch real versions from backend, removed hardcoded VERSION constant

Pre-existing TS errors in other files not introduced here. tsc passes clean for all touched files.

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(voice): transcript history + playback controls + proof review (H3D-CLAUDE-VOICE) (#64)

- Backend: add GET /api/voice/transcripts, GET /api/voice/recordings/{id},
  GET /api/voice/proof-events to voice.py; recordings served as binary
  audio from proof_events table; no API key in any URL
- Types: add VoiceTranscript and VoiceProofEvent to voice.ts
- Adapters: add getVoiceTranscripts, getVoiceProofEvents, getVoiceRecordingUrl
  to AdapterAPI interface + live implementations + parse helpers
- UI: Voice.tsx gains three-tab layout (Voice Browser / Transcript History /
  Proof Review); playback routed through backend only (new Audio(backendUrl)),
  no device access from frontend; honest empty states when no data yet

Gates: python -m py_compile OK; tsc --noEmit 0 errors

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(printers): onboarding wizard + Moonraker probe + S1 camera-only lock (H3D-CLAUDE-PRINTERS) (#71)

- Add 5-step printer onboarding wizard to PrintersTab:
  Step 1: Enter IP + connection type (Moonraker/OctoPrint/direct)
  Step 2: Auto-probe via GET /api/printers/probe (read-only, shows version/firmware/bed size)
  Step 3: Set camera URL with MJPEG validation
  Step 4: Confirm + save profile with write-enable toggle
  Step 5: Done / refresh fleet

- S1 (192.168.0.12) is LOCKED in the wizard: shows 'Camera only — cannot add as
  print target' before any network call is made; frontend enforces CAMERA_ONLY_IPS set

- Add GET /api/printers/probe backend endpoint:
  Read-only: calls only GET /server/info and optional /printer/objects/query
  Never sends GCode, commands, or mutations
  Returns: Moonraker version, klippy_state, bed size from fleet profile

- Add POST /api/printers/validate-camera backend endpoint:
  Read-only: HEAD request only, checks Content-Type for multipart/x-mixed-replace
  Returns: {ok, content_type, is_mjpeg, http_status}

- Add CAMERA_ONLY_IPS frozenset constant in printers.py (single source of truth):
  Any attempt to add 192.168.0.12 as a print target returns 403 CAMERA_ONLY_IP
  Covers: probe endpoint, onboard URL validation, printer ID validation

- Add test_printer_policy.py (16 tests, all passing):
  - S1 IP blocked in onboard URL validation (403 CAMERA_ONLY_IP)
  - S1 aliases blocked in printer ID validation (423)
  - Probe endpoint returns 403 for S1 IP
  - Probe is read-only: send_gcode/upload_gcode/start_print never called
  - Camera validate uses HEAD request only
  - MJPEG detection verified
  - TestClient route integration tests

- TypeScript: tsc --noEmit passes cleanly (0 errors in owned files)
- Python: py_compile passes for printers.py and test_printer_policy.py
- Pre-existing TS7026 errors in other src/*.tsx files are unrelated to this lane

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* docs(integration): 20-agent completion integration report (H3D-CLAUDE-FINAL-INTEGRATOR) (#72)

All 19 lane PRs (#53-#71) are OPEN/MERGEABLE with CodeRabbit SUCCESS and
Hermes evidence chain PASS. Two cross-lane file conflicts identified:

- app.py: PRs #66 + #69 both add a router (additive, UNION merge)
- adapters.ts / adapters.live.ts: PRs #64 + #71 both add methods (additive, UNION merge)

Merge order: Tier-1 (15 PRs in parallel) → Tier-2 (#66→#69) → Tier-3 (#64→#71).
Pre-existing JSX TS7026/TS7006 regression (~57 files) flagged as HIGH-priority fix-PR
needed before release.

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* audit(runtime): proof files + verifier scripts + route truth verification (#74)

Verifies all 6 proof JSON files are real (not hand-crafted), all 4 verifier
scripts use genuine subprocess/filesystem probes, and all 7 API routes have
real implementations. All syntax checks pass. No blockers found.

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* audit(merge): PR base + conflict cluster + silent drop verification (H3D-CLAUDE-POLISH-MERGE-2026-05-06) (#75)

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* audit(security): secret scan + path traversal + shell audit (H3D-CLAUDE-POLISH-AGENT-MCP-PROOF-2026-05-06) (#76)

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* audit(safety): S1 lock + printer policy + GCode scan verification (#77)

53/53 policy tests pass. S1 (192.168.0.12) blocked before every network call.
Zero GCode keywords in probe/read routes. Zero bypass paths found. Camera controls
are CSS-only display transforms with no hardware commands.

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* audit(docs): PR body completeness + ROADMAP truth + merge plan verification (H3D-CLAUDE-POLISH-RELEASE-DOCS-2026-05-06) (#78)

- Audited all 20 PR bodies (#53-#72): all have evidence chain, task ID, and gates
- PR #53 missing formal files table (prose description present); PR #64 minimal body
- TS7026 blocker documented in PR #72; absent from ROADMAP.md (codex-master locked)
- README "77 of 79" claim is stale relative to feature branch (update post-merge)
- Merge Tier 1/2/3 structure is correct; PR #72 needs explicit Tier 4 slot in plan
- No PRs accidentally merged; all 20 lane PRs remain open

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* audit(nofake-ui): 0 violations — no-fake scan + 33 buttons wired + 0 lane TS errors (H3D-CLAUDE-POLISH-NOFAKE-UI-2026-05-06) (#79)

* docs(handoff): final Codex takeover bundle — 9 audit/merge/lock files (#81)

Closes the Claude 20-agent + 6-audit-agent run. Contains:

  00_EXECUTIVE_TAKEOVER_SUMMARY.md  — 1-page status for Codex
  01_PR_MERGE_MATRIX.md            — exact tier merge order for #53-#80
  02_OPEN_BLOCKERS_AND_FIX_QUEUE.md — 0 code blockers, 4 low/info doc gaps
  03_LOCKS_WORKTREES_AND_BRANCHES.md — 28 Claude locks released, 28 worktrees
  04_RUNTIME_TRUTH_AND_NO_FAKE_AUDIT.md — Audit 2+3: 0 fake violations
  05_PRINTER_SAFETY_AND_PHYSICAL_IO_AUDIT.md — Audit 4: S1 camera-only PASS
  06_SECURITY_MCP_AND_AGENT_ACCESS_AUDIT.md — Audit 5: no traversal/secret leaks
  07_ARCHITECTURE_AND_FLOW_DIAGRAMS.md — Mermaid diagrams for all flows
  08_FINAL_CLAUDE_RELEASE_NOTE.md  — final PR list + lock state + Codex next steps

All 28 Claude-owned Hermes locks released.
All 20 lane PRs (#53-#72) and 6 audit PRs (#74-#79) open CLEAN.
TS7026 fix PR #80 open (CI running).
Hermes task: a2a_1778115796454_685e7b14

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(ui): clear post-merge npm audit vulnerabilities (#82)

* fix(ui): clear npm audit vulnerabilities

* fix(ui): clean post-merge browser gates

* fix: align observe refresh button contract

* [codex] add MCP-locked Hermes Agent code operator (#73)

* feat(agents): add MCP-locked code operator lane

* docs(handoff): add Claude final audit takeover contract

* fix(agents): harden code operator lane

* docs(handoff): Hermes3D OS folder index 2026-05-07 — 86 markdowns w/ inline SVG (#86)

Comprehensive index of every Hermes3D OS folder in G:\Github\ modified
between 2026-04-27 and 2026-05-07. Authored by 13 parallel sub-agents
under task a2a_1778147261453_661b606f.

Structure (12 categories, 60 included folders, 7 excluded):
  00_INDEX.md            -- master nav + topology SVG
  01_TAXONOMY.md         -- classification rules
  02_EXCLUSIONS.md       -- 7 folders intentionally excluded + reasons
  apps-vendored/         -- 7 vendored apps (~1.08 GB) + README
  core-repos/            -- 5 core H3D repos + README
  agent-infra/           -- 5 hermes-agent / MCP infra + README
  hp-protocol/           -- 9 HP P0/P1 hardening folders + README
  hermesproof/           -- 6 HermesProof component sandboxes + README
  source-os-60-apps/     -- canonical 60-app registry + treemap SVG
  h3dos-wire-tasks/      -- 18 single-button UI wire lanes + README
  h3dos-codex-tasks/     -- 5 Codex app integration lanes + README
  merge-prs/             -- 4 cascade-merge worktrees + README
  h3d-enhancements/      -- 7 H3D enhancement branches + README
  worktree-collections/  -- 3 umbrellas (49 sub-worktrees) + README
  research/              -- _research scratchpad + README

Each per-folder markdown includes: H1 title, purpose, status,
branch+commit, key files, relationships, and inline hand-written
SVG (400-900 px). Category READMEs add master inventory tables and
larger SVGs (700-900 px).

Excluded (7): kilocode-Azure2, contract-kit-v17 (3 variants),
TRELLIS.2, Agentic-Modeler, _repo_rescue_evidence -- documented
in 02_EXCLUSIONS.md with reasoning.

Hermes evidence chain: PASS

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(queue): recover closed stacked PR work (#105)

* feat(agents): add MCP-locked code operator lane

* docs(handoff): add Claude final audit takeover contract

* fix(agents): harden code operator lane

* feat(agents): add proof-gated git shipping lane

* feat(agents): add provider team assignment lane

* feat(agents): add provider execution artifacts

* feat(source-os): add runner contract matrix

* feat(source-os): register python cad verifier family

* feat(source-os): correct slicer runner truth

* feat(source-os): add print farm health verifiers

* feat(source-os): add service web health verifiers

* feat(source-os): add safe service start-runner preflights

* feat(source-os): supervise service runner starts

* test(unit): remove live fleet timeout from offline tests

* feat(source-os): add firmware source inventory verifiers

* feat(accel): add rust metadata proof worker

* feat(source): add read-only runner smoke contracts

* feat(source): add executable path runner smoke

* feat(source): add python import repair preflights

* feat(source): add slicer cli config preflights

* feat(source): add npm package metadata preflight

* docs(agents): define e2e proof plan

* feat(agents): add e2e workbench

* feat(agents): add provider smoke and reviewed ship lane (#104)

* feat(agents): add provider smoke and reviewed ship lane

* fix(agents): prove live runtime freshness

* feat(agents): add cli runner contracts

* fix(agents): require live provider smoke proof

* docs(handoff): Claude 20+ agent E2E completion intelligence bundle 2026-05-08 (#106)

Read-only intelligence sweep produced per PR 104's Claude 20+ Agent E2E
Completion Intelligence Contract. 12 markdown deliverables under
03_implementation/docs/handoffs/claude-e2e-intelligence-2026-05-08/
covering: executive map, G:/Github folder ecosystem audit, stale
code/branch map, Hermes Agent runtime gap map, Source OS 60-app
completion map, tab-by-tab UI no-fake audit, env-key/runtime config map,
test gates + proof matrix, PR + merge queue, Codex next 50 tasks, 6
Mermaid diagrams, and final Claude note.

Live truth captured at 2026-05-08 17:25Z from API on branch
codex/provider-smoke-workbench commit 43d8205: 220 routes, all 10 Agent
Workbench routes present, 60 Source OS apps (7 agent_cli_ready, 24
runner_gaps), 81 active UI files clean (no-fake scan PASS), 25 open PRs
all CLEAN/CodeRabbit-SUCCESS. Hard blocker: MiniMax + DeepSeek HTTP 401
on G:/private/.env keys (Tier 0 user action; Codex chain not blocked).

No source code edited. No PRs merged. No Codex-owned locks released. 12
hermes3d-locks acquired by claude-e2e-intel-aggregator (taskId
claude-e2e-intel-2026-05-08) for the markdown bundle; released after PR
open per contract.

Hermes evidence chain: kickoff ev_b6e233d4ac466056

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(agents): prove provider env aliases (#107)

* docs(handoff): prepare Claude 24-agent completion contract (#108)

* proof(I14): update no-fake sweep 2026-05-08 — 81 files, PASS (#116)

Active UI no-fake scan re-run on 2026-05-08:
- 81 production files walked from App.tsx entry point
- 0 findings (no mock/fake/simulated markers in string literals)
- No data/mock imports in active graph
- 15 orphaned/unwalked files separately verified clean
- scan_active_ui_no_fake.py requires no changes

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* test(I8): printer safety — S1 camera-lock tests, T1/V400 policy gates (#111)

Adds 25 new test cases to test_printer_policy.py closing the critical safety
gap where S1 (192.168.0.12) action endpoints (move, test, upload, upload-gcode)
had no direct hard-lock assertions. New TestS1ActionHardLock class proves 423
PRINTER_LOCKED fires before any MoonrakerClient I/O for all four action routes,
across all S1 aliases. TestT1V400PolicyGates confirms write-allowed printers are
not misclassified as S1 and pass the lock gate. 78/78 tests pass.

Task: H3D-CLAUDE24-I8-PRINTER-SAFETY

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(I4): add runner_family to all 60 runner-contracts + blocked_reason for BLOCKED rows (#119)

- Add _contract_runner_family() mapping runner_status → valid family string
- Add runner_family field to module_runner_contract() return dict (was absent,
  causing all 60 /runner-contracts rows to have FIELD_MISSING)
- Fix blocked_reason for slic3r and superslicer (runner_status=blocked):
  previously suppressed by cli_install_config_available=True condition; now
  always set when runner_status==blocked regardless of preflight runner
- Valid families emitted: agent_cli_ready, read_only_runner, executable_path,
  python_import_repair, cli_install_config, npm_package_preflight,
  desktop_app_runner_gap, gpu_worker_runner_gap, runtime_repair_required,
  source_reference_only, blocked, metadata_ready_needs_runner
- 113 pytest tests pass; only locked file modified

Task: H3D-CLAUDE24-I4-SOURCEOS-CORE
Hermes evidence chain: PASS
Gates run: python -m py_compile (both files), pytest 113 passed
Rows fixed (null→known runner_family): 60

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(I1): provider endpoint/model config audit — MiniMax+DeepSeek smoke (#112)

Audit confirmed both providers have correct code configuration:
- MiniMax: /v1/chat/completions, Bearer auth, MiniMax-M2.7 — all correct
- DeepSeek: /chat/completions, Bearer auth, deepseek-v4-pro — all correct
HTTP 401 on both is a pure API key issue (invalid/expired keys in G:\private\.env).

Added inline comments to PROVIDER_DEFAULT_BASE_URLS documenting the verified
endpoint/auth/model contract and the exact user action needed to resolve 401s.

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(I2): wire E2E code loop — patch-apply, gate-run, branch-commit-pr chain (#118)

- audit confirmed: apply_reviewed_patch_proposal, run_mcp_gate,
  git_commit_owned_files, git_push_current_branch, git_open_pull_request,
  restore_snapshot all fully implemented (no stubs)
- wiring gap found and fixed: no GET /e2e/jobs endpoint existed to list
  job states — added list_e2e_jobs() to code_history.py and the
  GET /api/code-operator/e2e/jobs route to code_operator.py
- new GET route queries proof_events for code_e2e/code_patch/code_git
  event types and returns job state legend for E2E loop operators
- expanded test_code_operator_routes_are_registered to assert all 7
  E2E chain routes are wired: apply-reviewed, gates/run, git/branch,
  git/commit-owned, git/push, git/pr, e2e/jobs GET
- added test_list_e2e_jobs_returns_proof_events and
  test_list_e2e_jobs_route_returns_200 — 92 tests pass (was 90)
- py_compile passes on both locked files
- provider 401 remains user-action only: I1 audit confirmed HTTP 401
  is a pure invalid API key issue; no provider HTTP code touched

Task: H3D-CLAUDE24-I2-E2E-CODE-LOOP
Hermes evidence chain: PASS

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(I3): OpenCode/OpenHands sandbox readiness + GET preflight route (#121)

* feat(I3): OpenCode/OpenHands sandbox readiness + GET preflight route

- Add opencode_openhands_sandbox_readiness() in code_history.py returning
  the I3-spec shape: opencode_detected, opencode_version, openhands_detected,
  openhands_image, sandbox_network_mode (always "none"), denied_paths, ready
- Add preflight_code_cli_runner_get() for non-mutating --version dry-run
  (GET variant, no task claim required); returns stdout, exit_code, elapsed_ms
- Wire GET /api/code-operator/sandbox/readiness to new function (replaces
  Docker-based response with OpenCode/OpenHands detection schema)
- Add GET /api/code-operator/cli-runners/preflight?runner_id=opencode|openhands
- Add SandboxReadiness panel to Agents.tsx with real detected/not-detected
  badges (data-testid=sandbox-readiness-panel), Refresh button, network mode
  and denied-paths display — no fake states
- Evidence: ev_040fad5fbd843c38 (opencode v1.4.3-hermes3d detected, exit_code=0)
- 38 unit tests green; task H3D-CLAUDE24-I3-OPENCODE-OPENHANDS released

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(I3): wire setSandboxBusy into Refresh onClick — resolve TS6133

Layer D2 UI-Final failed because setSandboxBusy was declared but its
setter was never invoked (TS6133). Wire it correctly: setSandboxBusy(true)
before the fetch, .finally(() => setSandboxBusy(false)) after, so the
Refresh button correctly shows "checking" during load and CI passes.

Evidence: ev_ffa9a8c3e3bd4a40 | Task: H3D-A1-PR121-FIX

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

---------

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* feat(I7): firmware source inventory probes — read-only git describe, source_reference_only contract (#120)

- Add FIRMWARE_SOURCE_PATHS registry mapping all 6 firmware module IDs to
  their actual source checkouts under Hermes3D-OS/source-lab/sources/
- Add _git_describe(): read-only subprocess.run(git describe --tags --always)
  with timeout=5s; returns None on any error — no flash/compile/serial
- Add probe_firmware_source_inventory(module_id): returns source_found,
  version_tag, runner_status=source_reference_only, agent_executable=False
- Add probe_all_firmware_sources(): aggregates all 6 modules in one call
- Update BUILTIN_RUNTIME_PROBES firmware entries: path fields now point to
  confirmed source checkouts; kind changed to firmware_source_inventory
- Add 04_testing/pytest/unit/test_firmware_farm_probes.py — 49 tests:
  registry coverage, contract template, _git_describe (mocked), per-module
  parametrized happy/absent paths, safety constraint enforcement tests
- Live probe result (evidence ev_842d77f8663ea2ee):
  firmware_klipper=293e1e9, marlin=03cc75f, prusa_firmware=f3e0dfd,
  reprapfirmware=f4297ad, repetier_firmware=7cb3741, smoothieware=620e162
- ABSOLUTE CONSTRAINTS: no avrdude/dfu-util/openocd/esptool, no serial port,
  no make/cmake/platformio, S1 not probed, T1/V400 source-only

Hermes evidence chain: PASS
Task ID: H3D-CLAUDE24-I7-FIRMWARE-FARM
Evidence ID: ev_842d77f8663ea2ee

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* feat(I6): service/web-app health probes + service_web_health_runner status (#122)

Add seven named read-only HTTP probe functions (probe_fluidd, probe_mainsail,
probe_octoprint, probe_fdm_monster, probe_octofarm, probe_manyfold,
probe_comfyui) plus a probe_service_web_health dispatcher.  Each probe uses
GET with a 3-second timeout, never POSTs, never mutates, and is blocked with
reason=no_configured_url when the env var is absent.

Update _runner_status to return service_web_health_runner (replacing the
generic readonly_api_ready) for local_http_health verifier kind, and add
service_web_health_runner_contract to _required_verifier_family.

Add 74-test suite in test_module_runtime.py covering: dispatcher routing,
blocked-when-no-url, non-local-URL guard, HTTP 200 happy path (mocked),
connection-error handling, 4xx handling, runner-contract status assertions,
and GET-only method verification.  Update pre-existing test in
test_source_runtime_contracts.py to reflect the new runner_status value.

All 226 unit tests pass (74 new, 116 combined with existing module tests).

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(I5): slicer/modeler CLI probes — PATH detection, --version proof, exact blocked reasons (#123)

* fix(I4): add runner_family to all 60 runner-contracts + blocked_reason for BLOCKED rows

- Add _contract_runner_family() mapping runner_status → valid family string
- Add runner_family field to module_runner_contract() return dict (was absent,
  causing all 60 /runner-contracts rows to have FIELD_MISSING)
- Fix blocked_reason for slic3r and superslicer (runner_status=blocked):
  previously suppressed by cli_install_config_available=True condition; now
  always set when runner_status==blocked regardless of preflight runner
- Valid families emitted: agent_cli_ready, read_only_runner, executable_path,
  python_import_repair, cli_install_config, npm_package_preflight,
  desktop_app_runner_gap, gpu_worker_runner_gap, runtime_repair_required,
  source_reference_only, blocked, metadata_ready_needs_runner
- 113 pytest tests pass; only locked file modified

Task: H3D-CLAUDE24-I4-SOURCEOS-CORE
Hermes evidence chain: PASS
Gates run: python -m py_compile (both files), pytest 113 passed
Rows fixed (null→known runner_family): 60

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(I5): slicer/modeler CLI probes — PATH detection, --version proof, exact blocked reasons

Adds probe_slicer_cli() and probe_modeler_import() to module_runtime.py.
Non-mutating: --version/help only, no STL sent, no firmware flashed.

Detected (on this machine):
  Slicers: PrusaSlicer 2.9.5, OrcaSlicer, FLSUN Slicer 2.0.4, CuraEngine 5.12.1, BambuStudio
  Modelers: Blender 5.1.1, OpenSCAD 2021.01, trimesh 4.12.1, pymeshlab

Blocked (exact path tried recorded):
  Slicers: SuperSlicer (not at C:/Program Files/SuperSlicer/), Slic3r (not installed)
  Modelers: FreeCAD (FreeCADCmd not at standard paths), cadquery/build123d/numpy-stl/open3d (not importable), truck (source-inventory only)

Adds SLICER_MODULE_IDS, MODELER_PYTHON_IMPORT_IDS, MODELER_SOURCE_INVENTORY_IDS constants.
Adds _find_slicer_executable() with canonical + alt + PATH search.
Handles PrusaSlicer/OrcaSlicer/BambuStudio/FLSUN nonzero --version exit codes.

Tests: 43 new slicer/modeler probe tests + 42 existing contract tests = 85 total, all green.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(I10): reduce polling lag, fix sidebar overflow, layout fixes (#110)

- AppShell: remove lg:overflow-hidden on main in dashboard mode to prevent panel cutoff on large viewports (overflow-auto retained throughout)
- Sidebar: wrap AgentChatMirror in min-h-0 shrink container so tall chat panel no longer displaces nav items off-screen
- TopBar: fix stale data — was fetch-on-mount only; add 10 000 ms setInterval refresh for system snapshot, notifications, and proof bundle (non-critical display data)
- globals.css: no changes needed (font-size 13px and dashboard-grid overflow-hidden are intentional)

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(I11): SourceOS 60-row rendering, real API wiring, panel overflow (#114)

- Fetch /api/modules/runtime/runner-contracts on mount + after Verify All / Setup Queue actions
- Map runner_status (runner_family) per module_id into a lookup dict
- ModuleList: display runner_family badge for each of the 60 rows using real runner_status from contracts endpoint
- AppDetailPanel: add runner_family header pill + RunnerContract InfoBox showing runner_status, required_verifier_family, safe_actions, acceptance_gate, and blocked_reason
- Pass runnerContract down to AppDetailPanel and refresh it in onRefresh callback
- All 60 rows rendered without slice/limit (confirmed via /api/modules count:60)
- Controls (Verify, Setup Plan, Backup, Rollback) already wired to real API — confirmed no fake handlers
- Panel overflow: AppDetailPanel section has overflow-auto in flex container with min-h-0

scan_active_ui_no_fake: 81 production files scanned, 0 findings

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(I13): print workflow — remove fake job states, policy-gate print actions (#113)

Dashboard: rename PIPELINE_STAGES to PIPELINE_STAGE_ICONS and remove the
hardcoded status:'complete'/'active' fields from the lookup table. Those
fields were dead code (PipelinePanel always derives status from live API
stage data); keeping them risked a developer treating them as truth.

Autopilot: remove EXPECTED_READINESS_CHECKS=16 magic constant. The gate
'allReady' was permanently blocked unless the backend returned exactly 16
checks — even if every returned check passed. Now allReady is true when
checks.length > 0 && all returned checks are ready (API is source of
truth). Added a "loading…" label and empty-state message while the API
response is pending so the UI never shows 0/0 as a misleading ready count.

Jobs: remove the 'counts' useMemo that injected 0 into every non-active
filter tab badge. Showing "Queued 0 | Done 0 | Failed 0" without fetching
those counts is a fake/misleading value. Now only the active filter shows
a live count; inactive filter tabs show no count badge.

Printers: no fake states found — all print actions await real API
confirmation before updating UI, and S1 policy block is correctly enforced.

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(observe): real health probing in /cameras, remove fake events, fix initial feed state (#115)

- observe.py /api/observe/cameras: replaced static _configured_camera_state()
  (always "configured") with a real _probe_camera_timed() call per camera so
  health reflects actual connectivity, not just URL presence.
- Observe.tsx initialFeedState: cameras with health="unreachable" now start in
  "error" state instead of "loading", preventing endless "CONNECTING" badge on
  known-dead feeds.
- ObserveConsole.tsx: removed hardcoded fake EVENTS strings; events panel now
  derives per-camera status lines from the real /api/observe/status response.
  Polling interval documented (STATUS_POLL_INTERVAL_MS = 5000ms >= 3000ms).

Task: H3D-CLAUDE24-I9-OBSERVE-CAMERA
Hermes evidence chain: PASS

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(I12): agent chat blocked state, voice text+audio+mute, learning real states (#117)

- AgentChatMirror: extract real blocked reason from response body (HTTP 401
  from MiniMax/DeepSeek now shows the provider error text, not just status code)
- AgentChatMirror: add explicit 'Providers blocked' banner in chat history
  when agent roster is empty, with action text for G:\private\.env config
- Voice.tsx: add mute button to TTS preview (Voice Browser fine-tuning panel)
  and transcript playback — muting suppresses audio but ALWAYS shows text
- Voice.tsx: text transcript displayed in all states; muted state explicitly
  shown with amber indicator so user knows audio is off but text remains visible

Voice API probe: GET /api/voice/status → 404 (route not registered in backend);
GET /api/voice/providers → Azure Speech READY (configured, region=westus).
TTS routes through backend /api/voice/preview (confirmed base64 response).
Learning: real API calls only, blockers shown with real reasons (confirmed live).
No-fake scan: PASS (81 production files, no mock/fake/simulated UX markers).

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(provider): align MiniMax and DeepSeek runtime adapters (#124)

* docs(handoff): tighten Hermes runtime finish contract

* docs(sweep): PR #125 control sweep handoff + runtime finish report

- HERMES_RUNTIME_FINISH_REPORT.md: 10-agent audit results, merge
  matrix, provider BLOCKED verdict (HTTP 401 both providers)
- PR125_CONTROL_SWEEP_HANDOFF_2026-05-09.md: full PR #125 sweep —
  A1-A10 audit results, zombie lock recovery, secret safety PASS,
  printer safety PASS, exact env key fixes required, next actions

Task: H3D-PR125-SWEEP-DOCS | Evidence: ev_dbf23c31c04af4ca, ev_a736131a4b0d8c6e

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

* fix(provider): align MiniMax and DeepSeek runtime adapters

- prefer MiniMax highspeed token-plan aliases and keep Max-Highspeed model routing explicit
- update MiniMax gateway to use MiniMax-M2.7-highspeed and max_completion_tokens
- update DeepSeek gateway/tests to use deepseek-v4-pro reasoning payload
- allow minimax/deepseek in llm policy and refresh provider rescue handoff docs
- keep provider smoke redacted; MiniMax now selects token-plan env and returns 429 insufficient_balance, DeepSeek remains 401

* docs(rescue): provider rescue blocker proof — adapters correct, blockers user-side

PR #124 provider completion sweep. Wave 1-3 audit:
- MiniMax adapter (gateways/providers/minimax.py): CORRECT per official docs.
  Reaches api.minimax.io. HTTP 429 insufficient_balance (1008) is
  provider-side billing/quota, NOT code, NOT auth.
- DeepSeek adapter (gateways/providers/deepseek.py): CORRECT per official
  docs. Posts to api.deepseek.com/chat/completions with thinking +
  reasoning_effort for v4-pro. HTTP 401 = "wrong API key" per
  api-docs.deepseek.com/quick_start/error_codes (single documented cause).

No code fix needed. Both blockers are out-of-repo user actions:
1. MiniMax: top up Token Plan balance / OAuth portal auth at platform.minimax.io
2. DeepSeek: rotate DEEPSEEK_API_KEY in G:/private/.env

Hermes Agent loop remains BLOCKED until both providers return accepted:true.

Evidence: ev_cffabca307652c21 (minimax), ev_bdec2f02c01c17ec (deepseek),
ev_491fe9d07426cab4 (adapter audit).
Task: H3D-CLAUDE-PROVIDER-COMPLETION.
No private values exposed.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(code-operator): expose redacted CLI provider env contract

---------

Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

* docs(agent): tighten control gates (#125)

* docs(rescue): mark blocker proof SUPERSEDED — providers now PASS (#132)

First proof-gated Hermes Agent coding loop. The full chain ran end-to-end
including a real recovery cycle:

  MiniMax build (artifact 08c8dce66b3a4a589572cee225f2b428)
  -> DeepSeek review v1 BLOCKED_ON_INSUFFICIENT_EVIDENCE
  -> v1 proposal 8365f7833f10... DeepSeek APPROVE ev_675dcddbd474c55d
  -> apply -> git-diff-check FAIL on trailing whitespace from `  ` line breaks
  -> rollback to snapshot 6f478adcf452 (proof 245d7fd376fc)
  -> v2 proposal f07e6af14f5f authored without trailing whitespace
  -> DeepSeek APPROVE v2 ev_59947bb44bf1715a
  -> apply -> git-diff-check PASS gate_git-diff-check_1778295564288

Provider smoke evidence baked into the SUPERSEDED block text:
  minimax  ev_4a52d9b1336ca9f2  HTTP 200  MiniMax-M2.7-highspeed
  deepseek ev_e708071cb269f170  HTTP 200  deepseek-v4-pro

No private values exposed. Same-owner MCP locks throughout.

Task: H3D-FIRSTLOOP-001-SUPERSEDE

Hermes evidence: f07e6af14f5f4db3972fdc1bb336bd35, ev_59947bb44bf1715a, ev_675dcddbd474c55d, ev_4a52d9b1336ca9f2, ev_e708071cb269f170, ev_a2780d9832e5567a, ev_95e16427ce056505, ev_69e00f87178d6f9c

* feat(recovery): Hermes Agent Recovery Controller v1 (lean ledger) (#133)

First lean v1 of the Hermes Agent Recovery Controller, born from the
recovery cycles in PR #126. Backend ledger ONLY: no UI, no autonomous
apply, no file mutation by the controller. Future-proof v1.5 schema
fields included so the upcoming Hermes Agent Task Monitor UI can read
rich state without backend refactor.

What ships:
- code_history.py: RECOVERY_FAILURE_CLASSES (9), RECOVERY_OUTCOME_STATUSES
  (3), RECOVERY_FAILED_STEP_TYPES (10), RECOVERY_RECOMMENDED_ACTIONS (8),
  RECOVERY_REDACTION_STATUSES (3), RECOVERY_WORKER_OUTPUT_STATUSES (6),
  RECOVERY_AGENT_STACK_VALUES (8), _RECOVERY_LEDGER_PATH, plus
  record_step_failure(), mark_recovery_outcome(), list_recovery_attempts().
  Adds `import secrets` and `from hermes3d.gateways.redaction import
  redact_text` to imports.
- code_operator.py: RecoveryRecordFailureRequest, RecoveryMarkOutcomeRequest
  StrictBody models + 3 routes: POST /recovery/record-failure, POST
  /recovery/mark-outcome, GET /recovery/state.
- test_code_operator.py: 7 lean v1 tests covering record/reject/redact/
  mark/state via TestClient with unique uuid4 task_ids.
- docs/handoffs/REVIEW_PACKET_*.md: 4 proof artifacts (full diff +
  contract per file) used in DeepSeek per-file review.

Provenance chain (each step proof-anchored):
- MiniMax artifacts 2130e9e1d12d4686ac4d788bfd136673 (build pass 1) +
  459bc9c00d6049dd949fa84e61f09c7a (build pass 2). BOTH truncated by
  completion-token budget. Manual fixes preserved chain-of-custody:
    (a) merge_conflict -> merge_git_fail (test class typo)
    (b) test_state_route_returns_attempts re-authored from truncation
    (c) `import secrets` added (MiniMax used secrets.token_hex without
        adding the import)
    (d) v1.5 future-proof fields added per user spec
- Per-file DeepSeek review APPROVE:
    code_history.py    proposal b63cd8b665bf4c2488591f8350b91cf5
                       review   ev_5635d54ac7fdcec7
    code_operator.py   proposal 7b76f0eae91a4f0d8c80850fcef0b4f0
                       review   ev_d9225ed939f77eb2
    test_code_operator proposal 33d7dbc691b146f7a495c6c23fa148b0
                       review   ev_4d1ca4217e6b226c
- Recovery cycle (gate failure -> targeted fix):
    pytest NameError: redact_text -> follow-up proposal
    fe2371073c3d4267b5e0e049df13e5b7 -> DeepSeek APPROVE
    ev_5586c466df5d59aa -> applied evidence ev_647bfc4e38e0738f.

Gates after final apply:
- python -m py_compile (code_history.py + code_operator.py): exit 0
- python -m pytest test_code_operator.py: 50 passed
- scan_active_ui_no_fake.py: 81 files, 0 markers
- git diff --check: exit 0
- hermes_run_gate git-diff-check: PASS gate_git-diff-check_1778298845085

Provider smoke evidence still PASS: minimax ev_4a52d9b1336ca9f2,
deepseek ev_e708071cb269f170. No private values exposed.

Next slice (separate PRs): autonomous repair dispatch (v2), Hermes Agent
Task Monitor UI (v3). v0.13.0 upstream Hermes Agent update is its own
proof-gated lane.

Task: H3D-RECOVERY-CTL-V1

Hermes evidence: b63cd8b665bf4c2488591f8350b91cf5, 7b76f0eae91a4f0d8c80850fcef0b4f0, 33d7dbc691b146f7a495c6c23fa148b0, fe2371073c3d4267b5e0e049df13e5b7, ev_5635d54ac7fdcec7, ev_d9225ed939f77eb2, ev_4d1ca4217e6b226c, ev_5586c466df5d59aa, ev_788974be0aa90ccd, ev_cbdcc4bca447d895, ev_e1ba5ddf993149bb, ev_647bfc4e38e0738f, ev_4a52d9b1336ca9f2, ev_e708071cb269f170

* docs(gui): add Hermes3D OS visual reference pack (#134)

* fix(agent-updates): harden staged-update pytest gate (Audit PR #135 follow-up) (#136)

Mirrors upstream NousResearch/hermes-agent tests.yml flags so the staged
update gate cannot fake-pass while v0.13.0 is formally deferred. Closes
the CICD-SEC-1 / Codecov-2021-style fake-pass surface in
_run_update_checks.

Patch
- Path ignores: --ignore=tests/integration --ignore=tests/e2e match
  upstream tests.yml. Marker-only -m "not integration" cannot block
  tests/e2e/conftest.py from polluting sys.modules at collection time
  (sys.modules["discord"] = MagicMock leak proven during Cplus-py311
  Phase 4 bisection).
- Workers env: HERMES_AGENT_PYTEST_WORKERS (default "4", mirrors GHA
  4-vCPU runner). Production rejects <2 with HTTPException(400);
  HERMES_AGENT_DIAGNOSTIC=1 overrides for triage. "auto" sentinel
  accepted. Garbage strings raise 400.
- maxfail: 1 in production (matches upstream tests.yml), 5 in
  diagnostic mode for triage-friendly multi-failure output.
- Skip path now fail-closed: missing HERMES_AGENT_RUN_PYTEST surfaces
  as status="fail" with "REQUIRES_CONFIRMATION:" output, never
  status="skipped" or 200/OK. Removes the fake-pass path that let
  pytest=skipped roll up as gate=verified.
- Timeout 300s -> 600s. Larger collected set under upstream-aligned
  --ignore needs the longer budget.

Tests
- 04_testing/pytest/unit/test_agent_updates_meta.py (3 tests):
  upstream tests.yml still has both --ignore= flags (network test,
  skip-on-offline), local source mirrors them, diagnostic+workers
  guard names + default value present.
- 04_testing/pytest/unit/test_agent_updates_skip_path.py (11 tests):
  skip-path fail-closed when env unset/zero, workers 0/1 rejected in
  production, workers 0 allowed in diagnostic mode, garbage raises
  400, default workers="4", path-ignores in pytest args, diagnostic
  uses --maxfail=5, "auto" sentinel accepted.

Result: 14/14 pass on 04_testing/pytest/unit.

Scope
- v0.13.0 update remains formally deferred (Cplus-defer-formal).
- This PR fixes the gate only; no runtime update was installed.
- Sources: PR #135 / commit 5ecd8ff (Batch 2 Agent 6 + Agent 10).

Follow-ups (separate PRs)
- Bonus 12: recovery ledger file lock, mark_recovery_outcome
  idempotency, agent_updates.py:115 HTTPException auto-repair gap,
  apply_patch_proposal TOCTOU.
- Bonus 13: 60-app audit doc errata (loader-real registry path,
  42 SPDX-invalid licenses).
- Upstream Agent 11 tickets (firmware archive-dir validator deferred
  here; YAML schema lacks the field today).

References
- https://raw.githubusercontent.com/NousResearch/hermes-agent/main/.github/workflows/tests.yml
- https://docs.pytest.org/en/stable/example/pythoncollection.html#ignore-paths-during-test-collection
- https://owasp.org/www-project-top-10-ci-cd-security-risks/
- https://about.codecov.io/apr-2021-post-mortem/

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(recovery): close Bonus 12 ledger races + auto-repair escape (PR #135) (#137)

Three findings from the Bonus 12 audit (PR #135 / bonus12-bug-finder.md):

Finding #1 (blocker, services/code_history.py recovery ledger)
- Append-without-lock allowed concurrent record_step_failure /
  mark_recovery_outcome calls to interleave partial JSONL lines on
  Windows. mark_recovery_outcome would silently json.JSONDecodeError-
  skip the corrupted entries and report "attempt_id not found".
- Fix: new _RecoveryLedgerLock context manager that combines a
  process-local threading.Lock with an OS-level advisory lock on a
  sidecar lockfile. Uses fcntl.flock on POSIX and msvcrt.locking on
  Windows; both stdlib, no new deps. flush()+os.fsync() on every
  append.

Finding #2 (major, mark_recovery_outcome)
- No idempotency check: a retry could append a SECOND outcome row,
  producing ambiguous state for list_recovery_attempts consumers.
- Fix: read scan now happens inside the same lock as the append.
  If any outcome row for attempt_id already exists, raise
  ValueError("already has a recorded outcome") atomically.

Finding #3 (major, agent_updates.py:115)
- _run_git raises HTTPException(502) on non-zero exit. A failed
  mid-step "git checkout --detach <tag>" escaped the for-tag loop
  without reaching _auto_repair_to_backup, leaving the Hermes Agent
  checkout on the previous (still-unverified) tag and surfacing 502
  to the caller instead of structured rollback.
- Fix: wrap the per-tag checkout + _run_update_checks in
  try/except HTTPException; record a synthetic step failure with the
  redacted detail and pivot to _auto_repair_to_backup. Also catches
  the HERMES_AGENT_PYTEST_WORKERS validation 400 added in PR #136.

Tests added (11 total, all green)
- 04_testing/pytest/unit/test_recovery_ledger_locking.py (8 tests)
  * lock helper exposes a backend (fcntl/msvcrt/thread-only)
  * 12-thread x 25-write concurrency test: every line round-trips
    through json.loads (no torn writes)
  * record_step_failure writes complete JSONL line + creates parent
    directory + lockfile sidecar
  * mark_recovery_outcome first call succeeds; second call raises
    ValueError with "already has a recorded outcome"
  * unknown attempt_id still raises "not found in recovery ledger"
  * race test: two threads finalize same attempt_id; exactly one
    succeeds, one raises idempotency error
- 04_testing/pytest/unit/test_agent_updates_auto_repair.py (3 tests)
  * failed checkout pivots to _auto_repair_to_backup (no 502 escape)
  * failed _run_update_checks (workers env 400) also pivots
  * all-pass path unchanged (smoke regression guard)

Verification
- py_compile: OK on all 4 files
- Focused tests: 25/25 pass (11 new + 14 from PR #136)
- Pre-existing failures in test_source_runtime_contracts.py (5
  firmware tests blocked instead of ready) confirmed pre-existing
  on base; out of scope for this PR.

Scope
- Recovery Controller v2 (RC v2) commits 2-5 stay paused per user
  instruction; RC v2 depends on the recovery correctness this PR
  restores.
- Hermes Agent v0.13.0 update remains formally deferred.
- Bonus 13 audit doc errata is out of scope (separate PR).

References
- https://docs.python.org/3/library/fcntl.html#fcntl.flock
- https://docs.python.org/3/library/msvcrt.html#msvcrt.locking
- https://about.codecov.io/apr-2021-post-mortem/
- https://owasp.org/www-project-top-10-ci-cd-security-risks/

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* fix(agent-updates): harden _zip_dirty_entries (Bonus 12 #4 / PR #135) (#138)

Defense-in-depth on the dirty-files backup zip in
api/routes/agent_updates.py:_zip_dirty_entries.

Pre-fix issues
- Opened the zip without allowZip64=True, so >4 GiB dirty backups
  silently truncate on Python builds that default to no-zip64.
- Used path.relative_to(repo) against an un-resolved repo path,
  raising ValueError and aborting the whole backup whenever the repo
  path is itself a symlink.
- A symlink in the dirty tree could resolve to a target outside the
  repo and still produce an arcname inside the archive, surfacing
  CWE-22 path traversal on extract.

Post-fix
- allowZip64=True passed to ZipFile.
- path.is_symlink() check skips symlinks defensively (even though
  _dirty_entries usually pre-resolves; tests / future callers may not).
- Arcname computed against repo.resolve() so symlinked checkouts
  (e.g. /tmp/repo -> /var/checkout) work cleanly.
- Arcname asserted to be a pure relative path (no absolute,
  drive-letter, parent-traversal, or empty components).
- Resolved-target paths that fall outside the repo are silently
  dropped instead of leaking into the archive.

Tests added (8, all green)
- 04_testing/pytest/unit/test_agent_updates_zip_dirty.py
  * normal files round-trip with relative arcnames
  * empty paths list short-circuits without creating an archive
  * symlinks (in-repo target) skipped — CWE-22 guard
  * symlinks (out-of-repo target) skipped — exfiltration guard
  * symlinked repo root produces correct arcname (no ValueError)
  * out-of-repo path silently dropped
  * allowZip64=True passed (probe via ZipFile subclass)
  * pathological absolute Path components silently dropped

Verification
- py_compile: OK
- 25/25 agent_updates-keyed unit tests pass
- Secret-leak scan on touched files: only descriptive test fixture
  string "outside-secret" (not a real secret)
- Pre-push hook: passed

Scope
- Bonus 12 finding #4 only (continuing the controlled-batch pattern
  from PR #137).
- v0.13.0 update remains formally deferred.
- RC v2 commits 2-5 remain paused per user instruction.

References
- https://docs.python.org/3/library/zipfile.html#zipfile.ZipFile
- https://cwe.mitre.org/data/definitions/22.html

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* docs(audit): 60-App Update Readiness Audit (docs-only, no runtime change) (#135)

* docs(audit): 60-App Update Readiness Audit + Phase 4 v2 patch proposal

Audit/planning lane only. No app updates. No GUI changes. No source mutation
beyond this doc. Hermes Agent v0.13.0 stays formally deferred per the
2026-05-09 user decision in handoffs/HERMES_AGENT_V013_UPDATE_LANE_CPLUS_PY311_DOCKER_FORMAL_DEFER.

What's in the audit:
- Per-app profile matrix: 60 rows across 11 sections (slicers 11 / modelers 13
  / 3D-gen 6 / print-farm 10 / firmware 6 / agent-cli 7 / library 1 / materials
  1 / hardware 3 / utilities 1 / research 1). Each row: update method, proof
  command, runtime env, deps, rollback method, blockers, recommended lane,
  auto-upd…
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.

1 participant