Skip to content

feat: integrate upstream with proxy hardening and observability - #165

Closed
AndrewMoryakov wants to merge 34 commits into
teamchong:mainfrom
AndrewMoryakov:integrate/upstream-20260801
Closed

feat: integrate upstream with proxy hardening and observability#165
AndrewMoryakov wants to merge 34 commits into
teamchong:mainfrom
AndrewMoryakov:integrate/upstream-20260801

Conversation

@AndrewMoryakov

Copy link
Copy Markdown

Summary

  • merge the current upstream main baseline and retain the fork's local dashboard, health, session, and Codex observability work
  • harden proxy ingress and credential handling: bounded buffered bodies, safe Codex OpenAI-key fallback, and fail-closed non-loopback credential guard
  • make dashboard/session accounting safer and consistent across live and persisted views, including provider-specific GPT/Gemini accounting and safe event-log pruning
  • add local-only Docker dashboard/health exposure and health-check diagnostics

Validation

  • pnpm run typecheck
  • pnpm test — 1004 passed
  • pnpm run build
  • git diff --check

Notes

The Docker Compose runtime could not be exercised in this environment because Docker is unavailable. The configuration is covered by static/type/test validation.

…260713

# Conflicts:
#	src/dashboard/fragments.ts
#	tests/dashboard-api.test.ts
Bundles pre-existing local WIP (baseline/proxy/tracker/types/sessions/codex-usage
+ their tests) with the approved model-scope persistence feature (model-scope-store
+ node/fragments/dashboard chips). Checkpoint only — reorganize/split later as needed.
…e getter

Final-review Minor: DashboardState already had getCompressionEnabled(); Task 3
added a redundant isCompressionEnabled() twin. buildHealthState now consumes the
existing getter.
Enabling a model that reads pxpipe-imaged content measurably worse (Opus 4.7/4.8,
GPT-5.5, GPT-5.6 Sol, Grok 4.5 — per FINDINGS) now pops an hx-confirm with the
concrete numbers; a lit weak reader shows a ⚠ marker. Fable/Sonnet/Terra/Lun and
broad gpt-5.6 are unaffected. Disabling never prompts.
…/unmeasured)

Replace the ad-hoc weak-reader list with an objective MODEL_READINESS table keyed
to committed eval receipts. Severity follows evidence strength:
- validated (Fable) -> no warning
- below-bar (Opus 4.7/4.8, GPT-5.5, Sol, Grok, broad gpt-5.6->enables Sol) ->
  blocking confirm on enable + concrete numbers + persistent tooltip
- unmeasured (Sonnet, Terra, Lun, and any unlisted id by default) -> non-blocking
  ⚠ marker + 'unmeasured' tooltip
Disabling never prompts. Opus/Sol now sit together honestly (both 0/15 verbatim);
Terra/Lun no longer get a false clean pass.
@AndrewMoryakov

Copy link
Copy Markdown
Author

Superseded by #167, which contains the reviewed backend-only proxy hardening changes without UI changes.

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