paperclip/packages/paperclip-runner/docs/runner-protocol-live-evals.md

16 KiB

Direct live Runner protocol evals

One Evalbook presentation

Every new report uses the canonical Evalbook grid and the existing Runner Lab chat viewer for attempt drill-downs. There is no plain-HTML attempt fallback. The grid, Latest, test-design, inventory and server-gate pages load the same built stylesheet and fonts as the chat viewer. The long-term S3 history index references that stylesheet inside an immutable campaign, so every page keeps the same dark theme. Static site styles live in devtools/issue-thread/src/evalbook-site.css, scoped to .evalbook-site; colors and typography come from the Runner Lab token layer. Missing recordings show a notice in the same viewer; missing viewer builds fail generation. Build with pnpm --filter @paperclipai/paperclip-runner build:issue-thread and provide --viewer-root or PAPERCLIP_EVAL_VIEWER_ROOT to the canonical Python renderer.

The Actions artifact contains full evidence. S3 uses the same viewer with a closed public DTO: only isolated mock-run conversation text, scrubbed private references, named tool outcomes, and checks. Tool arguments/results, reasoning, provider identities, and company snapshots stay private. The public notice explains these redactions. An unverified isolation boundary yields no public conversation, not a guessed reconstruction.

Public attempts use inert JSON and one shared viewer asset directory. The publisher verifies each shell and asset against the exact same-run viewer build, checks the public payload contract and local links, and rejects other scripts. The CSP prohibits network calls, forms and external resources. Supply PAPERCLIP_RUNNER_PROTOCOL_EVAL_VIEWER_DIR to the publisher. The workflow sends a viewer-only artifact to that job; raw attempts and provider secrets stay out. After publication succeeds, the publishing job writes Open this run's Evalbook and All eval runs links to the Actions summary. Its deployment URL also points to the exact immutable report, not to the downloadable ZIP. Failed publication does not advertise a successful deployment.

Before uploading, the report job runs the actual built application in Chromium against representative passing, failing and missing-recording pages in both reports. It checks initial rendering, tool expansion, read-only controls, navigation, reload, a narrow viewport and the public no-API-request boundary. In full-evidence reports, a tool's View in Evidence link selects the Evidence tab and highlights its record, including when reopening that same record after switching tabs. The browser check exercises this cross-link too. Run the same check locally with node packages/paperclip-runner/scripts/verify-runner-evalbook-viewer.mjs --report-root /path/to/report. Add --screenshots /path/to/proof for visual evidence. Public screenshots are retained in the aggregate Actions artifact, not added to the public report's closed file allowlist.

These screenshots are generated acceptance-test evidence from the scrubbed completed campaign, not a separate report design:

Passing chat replay with expanded tool evidence Failed chat replay with visible assertions Missing recording shown in the same viewer

Refresh a completed report without calling models

Download the aggregate Actions artifact, then:

node packages/paperclip-runner/scripts/refresh-runner-protocol-eval-report.mjs \
  --source /path/to/downloaded-aggregate \
  --evals-root /path/to/paperclip-evals \
  --viewer-root packages/paperclip-runner/dist-issue-thread \
  --output /path/to/new-refresh-directory \
  --revision chat-v1

Publish the returned reportRoot with the normal history publisher. The new ID is gha-RUN-ATTEMPT-report-chat-v1. Original reports remain immutable; history adds a labeled refresh and retains the source campaign, original measurement timestamp, renderer digest, and providerCalls: 0. Scores and evaluated source revisions do not change. This is not a new model qualification run. Future live runs create chat reports automatically. Refreshes are ordered by their render time in the history list, but keep the original measurement timestamp and never replace the latest or latest-green qualification identities. The HTML history links each measurement to its newest presentation, while retaining every original bundle and listing refreshes separately. Refreshes also recover retry-inclusive estimated and provider-list costs from the retained attempt records. They never add another model-cost measurement.

This is the provider-backed, one-turn protocol qualification layer in paperclipai/paperclip-evals/evals/paperclip-runner. It is intentionally separate from both the browser full-stack model E2E and the stress-derived workflow schedule in runner-workflow-evals.md.

The canonical unit of work is one live roster plus one authored case. A full campaign selects every enabled lane declared by campaigns/live-direct-full.json at one immutable paperclip-evals commit. That includes the complete 35-case provider rosters and the smaller ACPX Codex control roster. A disabled roster remains available for an explicit diagnostic selection, but is never inferred into a paid all run from the files present on disk. The native resume reliability gate is not a normal one-turn roster: it requires its separately governed external-resource campaign and remains opt-in.

Managed-provider evidence identifies the immutable deployed provider artifact: Claude Managed uses its Agent version, while AgentCore uses its qualification revision. Transport API or beta versions remain separate protocol metadata and must not replace that runtime identity in a report.

Hosted campaign

Use the Runner Direct Live Protocol Evals workflow. Dispatch the workflow from the default branch and provide:

  • target_branch: the Paperclip branch to build and test;
  • evals_sha: an exact 40-character commit from paperclipai/paperclip-evals;
  • rosters: all for every enabled lane in the canonical live-direct-full.json campaign, or a comma-separated diagnostic subset. Disabled lanes remain available only through an explicit diagnostic selection; all never spends against a lane that the eval program has marked disabled;
  • max_infrastructure_retries: zero through three, applied only when an attempt explicitly reports a retryable infrastructure failure.

The authorization job resolves the Paperclip branch to a commit and verifies the supplied eval commit before any checkout. A short-lived bot token generated from COMMITPERCLIP_KEY authorizes each checkout of the private eval repository; the token is masked and is never forwarded to a provider process. The workflow uses the same numeric actor allowlist, protected runner-e2e-paid environment, RunsOn fleet selector, and RUNNER_E2E_MAX_PARALLEL ceiling as the full-stack E2E workflow. Two balanced GitHub matrices keep each matrix below GitHub's 256-job limit while keeping their combined concurrency at or below that shared ceiling. Runner TypeScript, the native daemon, provider dependencies, and the attempt viewer are built once and reused by every cell. Because the complete suite requires two matrix shards, this workflow accepts a shared concurrency ceiling from 2 through 100.

The reused provider runtime is created with pnpm deploy --prod from the frozen workspace lock. That preserves the repository's qualified dependency versions and patched ACP server bytes. Do not replace this step with a fresh npm install of the packed Runner tarball: npm cannot apply the workspace's patchedDependencies, so the resulting ACPX executables no longer match their qualified digests.

The direct eval CLI also materializes a minimal immutable native runtime context in each isolated attempt workspace. This keeps the direct layer on the same paperclip.native-execution-input.v3 contract as production, including the AgentCore HarnessSkill upload path, without borrowing any browser E2E setup.

The direct CLI resolves native Codex from the pinned Codex ACP dependency, as the browser E2E launcher does; no globally installed codex is required. Its isolated fixture sessions explicitly use unattended ACPX approve-all permissions, matching the full-stack ACPX test profiles. OpenCode retains its existing semantic-only permission boundary. The seeded mock control plane still enforces tool grants, company boundaries, and governance checks. These test-only choices do not change production permission defaults and are not a qualification of interactive approval UI.

AgentCore eval admission waits up to 125 seconds, capped by the configured whole-turn timeout. Its network worker already allows 120 seconds for cold invocation delivery; the generic 30-second facade wait must not cut that delivery short during parallel fleet startup. Other providers retain their existing admission timeout.

The runtime instructions distinguish the fixture's finish_task/block_task state changes from native paperclip_finish/paperclip_block run-result reporting. This layer has no production server to apply a reported run result to a task. Its completion cases still require the actual semantic operation and durable mock-state change; reporting success alone does not pass them. The current turn request defines the requested work. Shared seed descriptions, old eval notes, or past accepted interactions are background state, not a replacement objective or proof that a newly requested action is already done. Finishing the provider turn does not authorize an unrequested mock task-state change or completion comment. These harness instructions keep single-operation cases bounded while leaving their operation and state-effect assertions intact.

ACPX accounting uses the qualified server's billable token semantics: Claude and Codex already include reasoning in output, and Codex has no cache-write billing category. Other missing categories remain unknown. ACPX stores Codex's terminal prompt-response usage without streaming it, so the sidecar forwards the single newly persisted receipt bound to that exact turn before its terminal event. It never substitutes a previous turn's receipt or context occupancy for billable usage. These receipt corrections also apply to non-eval ACPX sessions. The durable redactor preserves these known numeric token-counter fields; string/object values under the same names remain redacted as sensitive data.

all means the maintained enabled campaign, not every matching file in the roster directory. A missing campaign file fails closed. Within an enabled lane, a missing remote profile or unavailable provider is retained as an infrastructure result; it is not silently omitted. The current ACPX Pi roster is declared disabled because its Runner security profile is not qualified, so it runs only when selected explicitly for diagnosis. Use a roster subset only for diagnosis, not to claim the full maintained campaign is green.

A provider turn that reaches a durable failed, interrupted, or otherwise non-completed terminal still produces an attempt artifact and is scored as a behavior result. It is an infrastructure failure only when the harness cannot produce usable evidence, such as an unavailable provider, invalid profile, or transport failure. Automatic retries therefore never rerun a measured behavior failure merely to improve its score.

Required protected configuration

The paid jobs read only the credential selected for each roster:

  • OPENAI_API_KEY for native Codex and ACPX Codex;
  • ANTHROPIC_API_KEY for ACPX Claude and Claude Managed;
  • OPENROUTER_API_KEY for native OpenCode and ACPX Pi;
  • short-lived GitHub OIDC workload identity for AWS AgentCore.

Claude Managed also requires the four nonsecret PAPERCLIP_CLAUDE_MANAGED_* profile variables. AgentCore requires the nonsecret PAPERCLIP_AWS_AGENTCORE_* profile variables, including PAPERCLIP_AWS_AGENTCORE_EXECUTION_ROLE_ARN and the immutable PAPERCLIP_AWS_AGENTCORE_QUALIFICATION_REVISION; the eval fails closed when that deployed revision differs from the pinned roster config. The currently qualified context-aware harness revision is aws-agentcore-harness-context-v2. The workflow writes the GitHub OIDC token to a mode-0600 file and never forwards long-lived AWS access keys. Provision the AgentCore stack with --github-oidc-provider-arn so that scoped role admits only the paperclipai/paperclip repository's protected runner-e2e-paid environment as its web-identity subject. Scheduled runs additionally require RUNNER_PROTOCOL_EVAL_NIGHTLY_ENABLED=true and the pinned RUNNER_PROTOCOL_EVALS_SHA repository variable.

Reports and history

Each cell uploads its immutable run directory to an access-controlled Actions artifact. The trusted report job merges all expected cells, represents missing cell artifacts as infrastructure failures, and invokes the report program from the pinned eval commit. The full artifact contains the canonical Evalbook grid, read-only Runner issue-thread attempt pages, and raw immutable run records.

Public publishing uses a separate projection and a separate trusted OIDC job. The projection retains model/config identity, status, usage totals, and check outcomes and scrubbed mock conversation, but removes provider session identifiers, semantic-tool payloads, state revisions, traces, remote profile identities, and raw failure text. The same Evalbook report command renders that projection, so the public grid, test pages and chat viewer have the standard Evalbook layout. The publisher rejects untrusted scripts, remote resources, symlinks, unknown paths, broken links, raw session fields, and credential-shaped values.

S3 publication is additive:

runner-protocol-evals/
  index.html
  history.json
  latest.json
  latest-green.json
  campaigns/
    gha-<run-id>-<run-attempt>/
      index.html
      latest.html
      inventory.html
      tests/*.html
      attempts/*.html
      campaign.json
      bundle-manifest.json

Campaign files use immutable cache headers and a digest manifest. Reusing a campaign ID with different bytes fails closed. Only the root history and pointer files are mutable, and the publisher never deletes objects. The root history retains all run records; it no longer drops entries after 200 campaigns.

The history index includes:

  • Pass-rate and cost timelines, grouped by identical cell/model/driver membership and eval-suite SHA. Different suites and subsets cannot silently share a baseline.
  • Regression and recovery lists against the previous matching run, linking to the affected tests. Infrastructure failures stay distinct from behavior failures.
  • Estimated cost and provider-reported list cost, shown separately, never added. New campaigns include all retained attempts (including retries). Backfilled old campaigns with only final-cell usage are labeled historical final attempts only. Missing usage is unknown, not zero; partial totals use and display coverage.
  • Exact Paperclip and eval-suite commit links (full SHA on hover), the source ref, and the GitHub Actions run. These identify the code evaluated, not merely the commit used to render an old report.
  • A separate report-refresh list, excluded from trend points and model-spend totals.

history.json stores a versioned, derived analytics projection separately from immutable campaign records. The publisher backfills missing analytics from each original campaign.json; a refresh may enrich costs using retained raw attempts only when its source metadata and scores still match the original record.

The publishing job uses dedicated RUNNER_PROTOCOL_EVAL_HISTORY_* variables when present and falls back to the existing Runner E2E history role, region, bucket, and public base URL. Its default top-level prefix is runner-protocol-evals, distinct from runner-e2e. The AWS role must allow additive writes and reads for that prefix.

Local publisher checks

These tests make no provider or AWS calls:

pnpm --filter @paperclipai/paperclip-runner test:runner-protocol-eval-publish

To inspect the catalog without executing it, point the command at a local evals checkout:

pnpm --filter @paperclipai/paperclip-runner \
  report:runner-protocol-eval:catalog -- \
  --evals-root /path/to/paperclip-evals \
  --campaign-id gha-1-1 \
  --output /tmp/runner-protocol-eval-catalog.json