mirror of
https://github.com/paperclipai/paperclip.git
synced 2026-10-08 11:13:44 +02:00
Dotta
af8439a70b
## Thinking Path > - Paperclip is the open source app people use to manage AI agents for work. > - The Runner executes agents through native and managed provider drivers. > - The direct live eval layer had drifted from the current Runner contracts. > - The old local workflow did not provide a complete parallel campaign or durable report history. > - The Runner also needed current native OpenCode and OpenRouter qualification. > - This pull request restores the direct campaign, corrects the runtime gaps that the campaign found, and adds safe hosted Evalbook history. > - The benefit is repeatable model comparison against an immutable Runner and eval source revision. ## Linked Issues or Issue Description Refs #11297 Refs #11634 **What existing behavior does this improve?** This improves the direct live `paperclip-runner` eval workflow, provider execution contract, and static Evalbook reporting path. **Current behavior** The direct evals do not have one maintained full campaign on current `master`. OpenCode has no qualified multi-model OpenRouter roster. Parallel provider bursts can compact committed events before the transport observes them. Local reports do not have a separate safe S3 history index. **Proposed behavior** Run one immutable roster-plus-case matrix. Use the shared paid AWS runner fleet. Keep raw artifacts access-controlled. Publish a sanitized canonical Evalbook report under the separate `runner-protocol-evals` S3 prefix. Keep immutable campaign directories plus root history, latest, and latest-green pointers. **Reason and benefit** Maintainers can compare native Codex, native OpenCode, ACPX, Claude Managed, and AWS AgentCore behavior over time. They can inspect failures without mixing this direct protocol layer with browser full-stack E2E. **Breaking changes** None. The new workflow and S3 prefix are additive. The existing Runner full-stack E2E workflow and report remain separate. ## What Changed - Added a trusted two-shard direct live workflow for up to 393 roster-plus-case cells. - Reused the numeric actor allowlist, protected paid environment, and RunsOn fleet controls from Runner full-stack E2E. - Added immutable Runner and eval revision resolution, exact credential boundaries, bounded retries, and cost ceilings. - Added a public report projection that removes sessions, transcripts, tool payloads, state, traces, raw failures, remote profile identities, and credential-shaped values. - Added additive S3 history under `runner-protocol-evals`, with immutable campaigns and mutable root index pointers. - Added native OpenCode model injection and current OpenRouter pricing contracts. - Fixed direct eval completion, workflow execution, semantic discovery, warm-attach state reset, executable binding, and event-burst handling. - Kept Runner browser full-stack E2E behavior and publication separate. - Documented local and hosted direct eval operation. ## Verification - `pnpm --filter @paperclipai/paperclip-runner test:runner-protocol-eval-publish` — 15 passed. - `pnpm --filter @paperclipai/paperclip-runner build:typescript` — passed. - `actionlint .github/workflows/runner-protocol-live-evals.yml .github/workflows/runner-full-stack-e2e.yml` — passed. - Local current matrix at the revision in [paperclip-evals#17](https://github.com/paperclipai/paperclip-evals/pull/17) — 323 cells across 10 enabled configurations completed. - Final local current matrix — 269 passed, 11 behavior failures, and 43 expected macOS-only ACPX platform failures. - Targeted Runner checks — 13/13 eval-session tests, 15/15 publisher/security tests, and package typecheck passed; complete PR CI is green, including all browser E2E shards. ## Risks - Paid live campaigns can consume provider budget. Actor authorization, exact per-cell ceilings, protected environments, and explicit schedule enablement bound this risk. - Public reports can leak provider data. The workflow publishes only a separately projected report and validates every file before upload. - The new workflow cannot publish until it is present on the default branch. This pull request does not change the existing `runner-full-stack-e2e` publication path. - The campaign is large. It uses two GitHub matrices and caps combined concurrency at the shared fleet limit. > For core feature work, check [`ROADMAP.md`](ROADMAP.md) first and discuss it in `#dev` before opening the PR. Feature PRs that overlap with planned core work may need to be redirected — check the roadmap first. See `CONTRIBUTING.md`. ## Model Used OpenAI Codex on GPT-5. The exact deployment ID and context-window size are not exposed. The model used reasoning, code editing, browser inspection, repository tools, and live provider execution. ## Checklist - [x] I have included a thinking path that traces from project context to this change - [x] I have specified the model used (with version and capability details) - [x] I have checked ROADMAP.md and confirmed this PR does not duplicate planned core work - [x] I have searched GitHub for duplicate or related PRs and linked them above - [x] I have either (a) linked existing issues with `Fixes: #` / `Closes #` / `Refs #` OR (b) described the issue in-PR following the relevant issue template - [x] I have not referenced internal/instance-local Paperclip issues or links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip` URLs) - [x] My branch name describes the change (e.g. `docs/...`, `fix/...`) and contains no internal Paperclip ticket id or instance-derived details - [x] I have run tests locally and they pass - [x] I have added or updated tests where applicable - [x] I have updated relevant documentation to reflect my changes - [x] I have considered and documented any risks above - [x] All Paperclip CI gates are green - [x] Greptile is 5/5 with no open P2s, recommendations, or follow-ups - [x] I will address all Greptile and reviewer comments before requesting merge
88 lines
4.8 KiB
Markdown
88 lines
4.8 KiB
Markdown
# Stress-derived Runner workflow evals
|
||
|
||
The Runner workflow eval system turns the `STRESS-001`–`STRESS-044` campaign
|
||
into complementary deterministic, live, and chaos lanes. It is additive to the
|
||
capability inventory, capability cases, and existing scoring/report readers.
|
||
|
||
The workspace-private `@paperclipai/paperclip-eval-kernel` package owns only
|
||
structural scenario-by-candidate orchestration. Runner-specific cases,
|
||
observations, scoring, and traceability remain package-local. The HTML matrix is
|
||
rendered by the canonical `paperclip-evals` report program so live results use
|
||
the same grid and drill-down pages as the direct Runner eval suite.
|
||
|
||
## Lanes
|
||
|
||
- `pnpm --filter @paperclipai/paperclip-runner test:runner-workflow-evals`
|
||
runs the credential-free PR gate over sanitized Codex, OpenCode, and ACPX
|
||
normalization fixtures.
|
||
- `pnpm --filter @paperclipai/paperclip-runner report:runner-workflow-evals`
|
||
validates the deterministic fail-closed fixture matrix and writes JSON,
|
||
Markdown, JUnit, and GitHub-safe artifacts under
|
||
`.paperclip-local/evals/workflows/`. It makes no network requests.
|
||
- `pnpm --filter @paperclipai/paperclip-runner report:runner-live-evals` runs
|
||
the balanced forty-execution schedule against real provider sessions. Live
|
||
candidate failures are trend-only; missing credentials, qualification
|
||
failures, and provider outages remain unscored.
|
||
- `pnpm --filter @paperclipai/paperclip-runner report:runner-chaos-evals`
|
||
writes the eight-scenario fault schedule consumed by weekly and pre-release
|
||
restart, replay, trace, finalization, interaction, and wake-race suites.
|
||
|
||
The checked-in live manifest contains only adapter/model settings,
|
||
qualification variable names, and budgets. Credentials remain in the
|
||
environment. `PAPERCLIP_EVAL_MAX_CAMPAIGN_COST_USD` must be a positive finite
|
||
number and defaults to 12 USD for scheduled runs.
|
||
|
||
Live executions export one immutable Evalbook attempt per workflow/candidate to
|
||
`.paperclip-local/evals/workflows/evalbook-runs/`, then invoke
|
||
`evals/paperclip-runner/tools/eval_program.py report` from a `paperclip-evals`
|
||
checkout. That program writes the canonical matrix to
|
||
`.paperclip-local/evals/workflows/index.html`, plus `latest.html`, test pages,
|
||
and attempt pages. Set `PAPERCLIP_EVALBOOK_PROGRAM` to the program's absolute
|
||
path or `PAPERCLIP_EVALS_ROOT` to its repository root. Conventional sibling
|
||
worktree locations are discovered automatically. GitHub Actions checks out a
|
||
pinned `paperclip-evals` revision, so every hosted run uses the same reviewed
|
||
report implementation rather than a copied or package-local renderer.
|
||
Filtered reports with only a few candidates expand their result columns to the
|
||
available viewport, keeping the PASS, FAIL, and INFRA labels visible without a
|
||
horizontal scroll. Full matrices retain the canonical scrollable grid and
|
||
sticky test-name column.
|
||
|
||
The exported artifact contains only the safe workflow observation and
|
||
scorecard. Prompts, credentials, raw provider frames, tool arguments, and
|
||
reasoning remain excluded. `evalbook-manifest.json` records the generator path
|
||
and SHA-256 digest used for the render.
|
||
|
||
Local and manual GitHub runs can bound paid execution with comma-separated
|
||
`--candidate` and `--case` selectors plus `--limit`. For example,
|
||
`report:runner-live-evals -- --candidate codex-luna --limit 2` executes only
|
||
the first two Codex entries in that week's validated schedule. Subsets receive
|
||
a distinct bundle identity and do not contaminate full-campaign trend history.
|
||
|
||
The hosted live workflow is default-branch-only and requires an allowlisted
|
||
numeric actor plus the protected `runner-e2e-paid` environment. Scheduled runs
|
||
also remain disabled until `RUNNER_LIVE_EVALS_NIGHTLY_ENABLED` is explicitly
|
||
set to `true`. The paid job uses the full-stack workflow's literal runner
|
||
selection: `RUNNER_E2E_AWS_ENABLED=true` routes it to the RunsOn Fleet, and any
|
||
other value uses `ubuntu-latest`.
|
||
|
||
## Trace and reasoning safety
|
||
|
||
Live executions capture provider frames in a run-local mode-`0600` sidecar.
|
||
The evaluator verifies byte lengths, SHA-256 digests, order, dispositions, and
|
||
lineage, retains only redacted observations plus a digest, and destroys the
|
||
temporary trace after execution. Prompts, credentials, tool arguments, and
|
||
reasoning text never enter reports or uploaded artifacts. Evals measure visible
|
||
progress and activity; they do not inspect or grade hidden chain of thought.
|
||
|
||
## Compatibility and trends
|
||
|
||
Live bundle identity includes the Runner version/build, prompt policy, schedule
|
||
seed, adapters, resolved models, and reasoning settings. Seven-day comparisons
|
||
use only matching bundle IDs, and alerts stay disabled until seven compatible
|
||
reports exist. Safe reports are retained for 30 days; raw traces are not
|
||
uploaded.
|
||
|
||
The checked traceability manifest is
|
||
`spec/evals/stress-workflow-traceability.json`; CI fails for missing findings,
|
||
unknown workflow IDs, or missing regression-test anchors.
|