PaperClipAI

Jannik/PaperClipAI

0 Fork 0

mirror of https://github.com/paperclipai/paperclip.git synced 2026-10-06 21:05:21 +02:00

Commit Graph

Author	SHA1	Message	Date
DottaandPaperclip	72ff3a9f27	Measure native tool context and expand bounded workflow evals (#15218 ) ## Thinking Path > - Paperclip manages persistent agents and their assigned work. > - Agents receive both fixed instructions and tool definitions. > - Moving a procedure into a tool description still adds model context. > - We need to measure the complete delivered catalog and test real outcomes. > - The tested reductions saved little space and introduced failing outcomes. > - This PR keeps measurement, bounded eval coverage and the original evidence. > - Production prompts, tools and runtime behavior stay unchanged. ## Linked Issues or Issue Description Refs: #15151, #14961, #14948, #14985. This adds the measurement and eval coverage needed to assess further native instruction changes. The attempted hiring and dependency reduction failed qualification and is excluded from the final diff. ## What Changed - Measure the actual standard-mode tool authority, including all 39 tools and their input schemas. Capture scripted native start, resume and continuation payloads and the OpenCode MCP declaration list. - Add OpenCode to the two explicit-only local hiring/reuse and delegation/feedback stories. Preserve the original task requests and independent oracles. - Apply one attempt per selected story and explicit company and lead-agent budget stops. - Select managed hiring credentials from the requested profile, including OpenRouter. - Run the existing Node test files under Node instead of collecting them as Vitest suites. - Retain sanitized comparison reports, original failed grades, source hashes and evidence gaps. ## Verification Final source: `2e7cef78eef7cdfe02265e0dcb03e855b8e50bd8`. Local verification passes: - Full repository `pnpm -r typecheck` and `pnpm build`. - Eval-support unit tests: 1,252 Vitest checks and 128 Node checks. - Eval TypeScript check and six final-source measurement tests. All 36 normalized components across nine scripted deliveries and the OpenCode MCP catalog match the baseline exactly; the complete standing projection is 49,200 bytes. Fresh review of this exact head is [5/5 with no remaining findings](https://github.com/paperclipai/paperclip/pull/15218#issuecomment-5996723170); all three review threads are resolved. [Final-head CI](https://github.com/paperclipai/paperclip/actions/runs/37354539126) passes all 47 jobs, including repository typecheck, build, tests, runner checks, browser shards and the canary check. One initial annotation-test timeout is retained in attempt 1; its seven-test suite passed locally, and the affected CI lane passed on one targeted retry. No source or paid eval rerun was needed. Both ready-transition security scans passed with zero annotations. The PR is out of draft and conflict-free. GitHub still requires code-owner approval for the `package.json` test-script change; its requested reviewers are already set. A byte-for-byte comparison against master context `a65ca0950834a85bb93bcc4b4042ecacdebfef53` confirms no production changes under `packages/` or production `server/` paths. The only server addition is a measurement test. No new paid rerun is needed to compare unchanged production bytes. This does not claim that existing product defects have been fixed. The rejected corrected experiment had baseline 5 PASS / 1 FAIL and candidate 3 PASS / 3 FAIL, including two newly failing pairs. The later readiness experiment had candidate 3 PASS / 3 FAIL and baseline 3 PASS / 2 FAIL / one setup cell without a behavioral grade. Its five comparable pairs had two new failures, two new passes and one unchanged pass. The missing baseline Codex cell never reached its provider step because Docker setup timed out. The observed missing behaviors include parent continuation, waiting for the latest child revision, revised ZIP delivery and OpenCode credential persistence. Those failures remain failures. Source review and passing CI do not regrade them. The original reduction saved 460 bytes; its first repair saved only 125 bytes, and the larger unqualified runtime repair increased the full projection. None of those production changes is shipped here. Read the [report](https://github.com/paperclipai/paperclip/blob/2e7cef78eef7cdfe02265e0dcb03e855b8e50bd8/doc/plans/2026-10-05-native-procedure-guidance.md) and its linked sanitized receipts for exact sources, original campaign links, pair-level results and evidence limitations. ## Risks - Full-catalog bytes are not model tokens, invoices, private vendor prompts, lazy-loading behavior or truncation proof. - The two stories are explicit-only and do not prove general coding quality or arbitrary resume behavior. Single trials do not establish causation or performance trends. - The retained OpenCode candidate credential guard failed. Cleanup removed the original provider database, so the precise persistence mechanism remains unknown. The guard is unchanged. - ACPX provider-execution IDs lack proven host-call mapping. No extra-work or feedback-consumption claim is inferred by matching names, order or counts. - PostgreSQL cannot start locally while the host's shared-memory slots are exhausted. Hosted CI must supply the full database checks; the full local database suite is not claimed green. ## Model Used OpenAI Codex, based on GPT-6, with repository tools and code execution. The exact deployment model ID and context-window limit are not exposed in this session. ## 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 --------- Co-authored-by: Paperclip <noreply@paperclip.ing>	2026-10-05 20:03:35 -05:00
DottaandPaperclip	b17019e14d	fix(agents): reduce default instructions and qualify stock harnesses (#14948 ) ## Thinking Path > - Paperclip is the open source app people use to manage AI agents for work. > - Its adapters supply task context and access to Paperclip skills and tools. > - The default hire manual and shared prompts also repeat general work procedures. > - Those procedures overlap with stock provider instructions and the Paperclip skill. > - Existing E2E fixtures supply a QA manual, so they do not qualify the production default. > - This pull request reduces the generic instructions and adds real default-hire coverage. > - The benefit is less competing guidance, with inspectable evidence for preserved skills and task context. ## Linked Issues or Issue Description Refs: #14920. That merged change preserves native Codex base instructions. This PR covers the default manual, shared legacy prompts, operational skill guidance, and the narrowly approved ACP skill-discovery/session-environment repair for measured delivery and credential-persistence failures. What existing behavior does this improve? New non-CEO hires without a custom bundle and legacy task/chat startup and continuation prompts. Current behavior The shipped default manual contains 602 words. Generic task/chat prompts and ordinary resume deltas repeat work procedures already available through the harness and Paperclip skill. Proposed behavior The default manual contains only the eight-word company identity. Shared startup prompts retain identity and connection guidance. Ordinary resume deltas retain current work context without the generic execution contract. Reason and benefit Let the stock harness guide general work. Keep Paperclip-specific capabilities and independently test default hires, skills, ordered comments, and chat restart. Breaking changes New default hires receive less guidance. Existing saved manuals, explicit custom bundles, CEO templates, and specialized wake contracts retain their behavior. The obsolete includeExecutionContract option remains accepted for source compatibility. ## What Changed - Reduce the default hire manual to one sentence. - Reduce shared task/chat defaults and remove the generic ordinary-resume contract. - Keep connection guidance, auth, skills, custom prompts, and specialized wake context. - Add credential-free instruction-boundary gates and 26 explicit Product E2E cells across eight legacy/native profiles, including two focused Paperclip-storage cases. - Capture public hire receipts before providers run, then grade delivered prompts and independent task/chat outcomes. - Add an early legacy skill API recipe for saving a task document, checking the saved revision receipt and linking the document. Improve stock task/heartbeat skill-selection metadata and show a clickable Markdown UI-link example. Keep native tool completion separate. - Advertise bounded routing descriptions and exact successfully staged SKILL.md paths in legacy ACP Claude; keep full bodies on demand and preserve remote path rebasing. - Remove only the provider environment from copied persisted ACP session records, while loading current run credentials and preserving all other options/conversation state. - Regenerate both capability metadata inventories and reject stale manifests/inventories before provider admission. - Publish the original reduction and focused skill-repair comparisons, preserving all failures, automatic recovery, cost coverage and limitations. ## Verification Behavioral qualification remains pending. Original legacy ACP Claude loses the issue document only in the reduced cohort beneath an unchanged credential failure. A source-backed diagnosis finds that neither ordinary assignment reads the staged operational skill, while the runtime persists provider environment in session state. The new common repairs expose skill metadata/path and omit persisted env; strict document and credential guards stay intact. [Inspectable diagnosis and retained hashes](https://github.com/paperclipai/paperclip/blob/9f654db4541d3d002769c988f6e51fc0b08dadbd/doc/plans/2026-10-03-legacy-acp-claude-readiness.md). Current repair head `de0965984ff3edf611ae6d0e7ca5c7d5ae3947bb` incorporates master `569c7203aa24b95440682983ce7940ba1d4247bd` (merged #14961/#15007). All 222 affected adapter tests, adapter-utils/E2E typechecks, and final 96 variant/grader/retry calibrations pass. The frozen historical comparator is `c25697f4260b6f3adfea143c3ae9932e2f42986d`: 8,280 of 8,291 paths identical, exactly two production instruction paths plus nine declared unit expectations differ. The operational skill/discovery/environment repairs, selected model/profile/task/core grader/auth/permissions/retry policy are identical. Both actual launcher prepare→verify admissions pass with zero providers. [Immutable manifest and exact receipts](https://github.com/paperclipai/paperclip/blob/9f654db4541d3d002769c988f6e51fc0b08dadbd/doc/plans/2026-10-03-legacy-acp-claude-evidence/manifest.json). One original legacy ACP Claude cell per variant is authorized, with enforced single campaign attempts, 12-minute deadlines and company/agent 1,000-cent hard stops; every product recovery run/cost is counted. Actual live outcomes are pending. Current normal CI has one failed server shard and failed aggregate verify under diagnosis; other normal gates including typecheck/build/Rust/all eight browser shards pass. Fresh review completed successfully; the valid historical startup/resume masking finding was fixed with per-invocation task/chat checks and strict complete-snapshot capture, calibrated and resolved. Prior heads, failures and campaigns below remain historical evidence, not checks on this repair head. - Prior head `36aa4d81c49a1a8f6f04b1a068fae19aa901955f` is replayed on merged hiring master `862a5758ba0e88a33232c1f1fa645e85c38a3113`. All 52 current-head checks pass with two intentional Storybook skips, including repository typecheck/test/build and the browser shard. Fresh Greptile is 5/5 with zero unresolved review threads. Exact-head stock prerequisites pass 599 assertions (598 TypeScript + 1 Rust), all six gates and retained receipt verification, zero providers/source errors. Fingerprint `a7f5a22d860a88fe20cce213c6d5e0004788f32c8363930729aea4fd740ad16d`. Combined catalog/hiring calibrations pass 67 assertions, E2E typecheck and 26-cell stock discovery pass. Canonical contract/inventory checks and the later issue-derived reference calibration are retained; that reference-only follow-up is not live-qualified by earlier frozen runs. - Prior full repository typecheck/build passed. The complete local Vitest run executed 14,956 tests: 14,870 passed, 83 skipped, three timing failures. All three affected files passed unchanged narrow reruns; original failures remain retained. Current-head CI now passes the full general checks; the original local failures remain retained. - The original 24-pair default-manual/shared-prompt comparison has two new overall classic Claude/OpenCode document-delivery failures plus an additional legacy ACP Claude document loss beneath an unchanged credential-guard failure (not closed by later runs), two newly passing OpenCode ordered cases, seven unchanged failures and 13 unchanged passes. Equal 15/24 totals do not establish behavioral equivalence. [Complete original report](https://github.com/paperclipai/paperclip/blob/875f4c397d9e8c3f12f39dedd59abaf1eaf5236e/doc/plans/2026-10-02-stock-harness-live-comparison.md). - The skill-only repair holds the eight-word manual/shared prompts and merged #14920 fixed. All four matched profile configurations and 203 fixture/behavior files match. Candidate `abd0b628ca642c09a54a4edc56a5227402f6686e` varies only the two skill sources against baseline `bc83fe030234439ac51279502a28803958963e2e`. [Candidate workflow](https://github.com/paperclipai/paperclip/actions/runs/37060885547) and [baseline workflow](https://github.com/paperclipai/paperclip/actions/runs/37060888047) each pass 571 exact-source prerequisites before providers; all eight cells clean up successfully. Failed campaigns publish successfully and remain failed. - Repair pairs: Claude original Fail → Pass; Claude explicit Pass → Pass; both OpenCode cases Fail → Fail. Explicit OpenCode's handoff worsens beneath the unchanged failing UI-link grade: baseline gives a clickable API URL, candidate gives a code-formatted path without an anchor. The request's usable-link wording is narrower in the UI-only oracle. [Complete repair report and safe projection](https://github.com/paperclipai/paperclip/blob/875f4c397d9e8c3f12f39dedd59abaf1eaf5236e/doc/plans/2026-10-02-legacy-document-skill-repair.md). - The subsequent narrow stock metadata/link correction has two matched Pass → Pass cases, zero new machine failures/passes and no pending pairs. Both original-case handoff links remain deficient: candidate uses a wrong PAP prefix, baseline supplies a bare prefix-less slug path; the preserved original oracle only requires a durable document. Both explicit clickable UI-link cases pass revision/content/link grading. All four exact-source 587-check gates, single assignment runs and cleanup pass. This does not establish fix causality because baseline also succeeds. [Candidate workflow](https://github.com/paperclipai/paperclip/actions/runs/37069547401) freezes `fe9dc1e3c518825242ed889ab9c8352986f8c2ed`; [matched baseline](https://github.com/paperclipai/paperclip/actions/runs/37069552374) freezes `0d7ecfa96d72fba79b7f0a25052b42c0686c0488`. This is a skill-only comparison with reduced manuals/shared prompts held constant, not a repeat of the historical-manual comparison. Only original and clarified explicit classic OpenCode cases are selected, two per variant/four expected turns. 8,242 other tracked files and both profile hashes match; protected workflows admit each exact source before credentials. [Complete qualification report](https://github.com/paperclipai/paperclip/blob/74d0d3d945f4c52d0814b5a845ab5bd09f33cd6b/doc/plans/2026-10-02-opencode-skill-routing-link-qualification.md). Candidate original loads Paperclip/reference before saving publicly; baseline original loads it after writing locally, then saves publicly within the same assignment. Reported cost totals are $0.0107824490 candidate / $0.0107909015 baseline, with unmetered runtime. The later reference-only issue-derived link correction is provider-free calibrated and not live-qualified by these frozen runs; no further paid runs. - Retained tool calls show the repaired original OpenCode assignment loads only its assigned output skill before writing locally. Operational Paperclip is first loaded during automatic disposition recovery; its early recipe is visible then, but it never saves the missing document. Explicit candidate loads Paperclip and reads the new reference before saving successfully. All nine actual runs are counted. Reported LLM totals are $0.3802537209 baseline and $0.4918990161 candidate; local runtime is unmetered. - Initial setup, packaging, cancelled/missing-cell recovery, callback test and relative-output attempts remain retained. No completed provider failure was rerun. Frozen measurement branches are unchanged by later canonical metadata maintenance. - Run `pnpm test:e2e:runner:stock-harness`, `pnpm test:e2e:runner:unit`, and `pnpm test:e2e:runner:typecheck`. Select `stock-harness` explicitly for paid execution; it is excluded from `--all`. Prior-head integration: `36aa4d81c49a1a8f6f04b1a068fae19aa901955f` replays this PR on merged hiring #14985 (`862a5758ba0e88a33232c1f1fa645e85c38a3113`), preserving the four explicit custom-CEO-bundle checks, minimal generic manual boundary, and both suites. The combined fixture catalog and hiring calibrations pass 67 assertions; exact-head stock prerequisites pass 599 assertions (598 TypeScript + 1 Rust), all six gates and retained-receipt verification, zero providers/source errors, fingerprint `a7f5a22d860a88fe20cce213c6d5e0004788f32c8363930729aea4fd740ad16d`. E2E typecheck and 26-cell stock discovery pass. Fresh current-head CI passes all 52 checks with two intentional skips, and fresh Greptile is 5/5 with zero unresolved review threads. The prior source-plan browser failure is retained: a deterministic process fixture replayed its last `fixture:plan` command on `chat_task_completed`, writing revision 2 with identical body after the approval handoff. This was not paid provider execution. Rebased current-head CI passes the same assertion without an old-head retry or a change to that browser fixture. The merged hiring change was measured separately on immutable matched unions, with this reduced/shared/operational context and native completion guidance held constant. [Complete original two-profile report](https://github.com/paperclipai/paperclip/blob/f0512647656be78e48abd8c22a3078db8bf6bcd2/doc/plans/2026-10-02-hiring-template-live-comparison.md): [candidate](https://github.com/paperclipai/paperclip/actions/runs/37075466208) / [historical baseline](https://github.com/paperclipai/paperclip/actions/runs/37075469463), 705 provider-free prerequisites each. Both pairs are unchanged Fail → Fail on the exact-five count, with six core delivery checks passing all four cells; 28 actual successful runs include eight automatic completion wakes, zero retries, four successful cleanups. Source-read coverage is uncomparable, actual model charges unknown. Separately versioned provider-free accounting remains analytical work; original verdicts are preserved. This does not rerun or qualify the completed default-manual or native campaigns. ## Risks - Legacy ACP Claude's additional delivery loss is not closed by any later matched run and blocks the no-extra-failing-behavior merge criterion. Legacy document delivery may have relied on the prior manual/shared prompts. The early skill repair improves Claude in one trial; the later OpenCode pairs pass in both variants and cannot establish causality or robust recovery. Both original-case links remain deficient beneath the storage-only grade. The later issue-derived reference correction has only provider-free validation. Native finish/block descriptions must not be supplied to legacy agents. - The comparison holds merged native Codex fix #14920 constant; it cannot measure that fix's before/after task performance. - These bounded skill/context/chat workflows do not measure general coding quality. Unrepresented providers remain unqualified. - Saved manuals and old Codex sessions are not automatically migrated. Codex through ACP still has a separate base-instruction follow-up. ## Model Used OpenAI Codex, GPT-6 family as identified by this session. The exact deployment ID and context-window size are not exposed. The assistant used reasoning, repository tools, code execution, and delegated PR/eval work. ## 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 (relevant suites and all three unchanged narrow reruns pass; complete-run timing failures retained in Verification) - [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 - [ ] All Paperclip CI gates are green on the new repair head (prior-head checks retained above) - [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups on the new repair head - [x] I will address all Greptile and reviewer comments before requesting merge --------- Co-authored-by: Paperclip <noreply@paperclip.ing>	2026-10-03 12:32:42 -05:00
Dotta	5716fe907e	test(runner): add full-stack acceptance and eval gates (#12700 ) ## Thinking Path > - Paperclip is the open source app people use to manage AI agents for work. > - The runner subsystem executes agent work across local and managed provider backends. > - The lower pull requests restore the task runtime, provider backends, and managed-provider control plane. > - The restored system needs repeatable full-stack checks before it can ship safely. > - Paid live checks also need clear access, cost, and secret controls. > - This pull request adds acceptance, live evaluation, chaos, and release gates for the restored runner stack. > - The benefit is measurable runner parity with safer release decisions. ## Linked Issues or Issue Description Subsystem affected Cross-cutting. This change covers runner tests, release workflows, server contracts, and evaluation tools. Problem or motivation The runner stack did not have one complete acceptance surface for native Codex, ACPX, Claude Managed, and AWS AgentCore. Release checks could miss provider drift, task-view regressions, cost-policy errors, and destructive cleanup errors. Proposed solution Add a 57-cell full-stack catalog, a Daytona image, and opt-in paid workflows. Add live evaluation, chaos, cost-limit, redaction, and release contract checks. Add AWS AgentCore infrastructure and guarded provisioning tools. Keep the native runner experimental flag off by default. Alternatives considered We considered manual smoke tests only. They do not give repeatable evidence and they do not protect release branches. We also considered one large pull request. The stacked pull requests keep each review below the Greptile file limit. Roadmap alignment This work supports the shipped Cloud / Sandbox agents milestone and the shipped Agent evals & feedback milestone in `ROADMAP.md`. Related stack: - #12699 adds managed provider backends and lifecycle support. - #12691 adds qualified OpenCode and ACPX provider backends. - #12685 restores task runtime rendering and steering. ## What Changed - Add the runner full-stack harness with 57 catalog cells and 60 unit tests. - Add a Daytona runner image with digest-pinned base images and base-aware image-content checks. - Add guarded live evaluation and chaos workflows with a fixed 40-execution matrix; live and full-stack paid schedules now run only on Sundays or by manual dispatch. - Add in-flight reported-usage cost stops, post-turn cost caps, exact-threshold failure classification, secret redaction, retry classification, and actor authorization. - Reattach stream and hard-budget listeners before restart-recovery continuations so restored paid sessions cannot bypass in-flight interruption. - Preserve OpenCode usage and cost across tool-loop messages and turns while exposing an explicit current-run delta to durable accounting. - Keep PNG/WebM evidence in access-controlled artifacts only, reject SVG, and publish only pruned inert structured per-attempt evidence. - Add AWS AgentCore infrastructure, provisioning checks, and smoke tools; reject unsafe model identifiers, require exact stack ownership markers, and make failed-stack replacement explicit. - Add evaluation-session contracts and capability reports. - Add release workflow checks for immutable action pins, frozen dependency installs, exact weekly cron shape, paid-run guards, provider-secret isolation, and chaos test paths. - Reauthorize the original and triggering numeric actor IDs as the first step of every provider-secret job, including partial reruns, before checkout or provider access. - Give each full-stack matrix cell only its matching provider credential, expose Daytona only to Daytona cells, and disable shared dependency caches anywhere paid credentials or OIDC write access are present. - Protect the legacy manual E2E workflow with the same default-branch, allowlist, environment, and per-job authorization boundary. - Rotate live-eval candidates by week and retain 120 days of compatible history so the seven-week trend window remains viable. - Restore the root runner-acceptance commands and reconcile reported snapshots, raw receipts, and terminal usage without double counting or losing late usage. - Mark ACPX token deltas exact only when every budget field is present, keep cumulative cost/request authority separate, reject non-USD cost labeling, and include thought tokens in output-token budgets. - Keep `enableNativeRunner` off by default. The acceptance harness enables it only in its isolated test instance. ## Verification Passed locally: - `pnpm --filter @paperclipai/paperclip-runner typecheck` - `pnpm test:runner-acceptance:typecheck` - `pnpm test:runner-acceptance` (19 tests) - focused OpenCode proxy, driver, runnerd transport, live-session, and turn-stream tests (106 tests) - `pnpm --filter @paperclipai/paperclip-runner exec vitest run src/live/clean-room-server.test.ts` (22 tests) - `pnpm test:e2e:runner:typecheck` - `pnpm test:e2e:runner:unit` (62 tests) - `node --test scripts/__tests__/release-verify-workflow.test.mjs` - `pnpm --filter @paperclipai/paperclip-runner test:runner-workflow-evals` (22 tests) - `pnpm -r typecheck` - `pnpm build` - `node --test packages/paperclip-runner/scripts/aws-agentcore-provisioning.test.mjs` (6 tests) - `git diff --check` - `cargo test --manifest-path packages/paperclip-runner/runner/Cargo.toml -p paperclip-runner-core --lib --locked` (161 tests) - focused ACPX provider-event tests (10 tests) - The rebased PR changes 92 files. `pnpm-lock.yaml` is unchanged. I did not run paid live provider jobs or provision AWS resources. Those checks need credentials and can create cost. ## Risks The paid workflows can create provider cost. They require an allowlisted original and triggering actor, the protected `runner-e2e-paid` environment, explicit opt-in variables, and cost limits. The four provider credentials exist only in that master-only environment, which requires allowlisted reviewer approval and disables administrator bypass; repository and organization Actions scopes contain no copies. Provider usage arrives after a billable request, so the live guard cannot prevent one request from crossing a threshold. It interrupts immediately on the first reported threshold hit and permits no continuation. Visual evidence can contain secrets rendered as pixels. PNG/WebM remain only in access-controlled workflow artifacts; SVG and per-attempt XML are excluded, and S3/Pages receive a pruned structured dashboard. The AWS scripts can create cloud resources. They use explicit commands, least-privilege roles, KMS encryption, saved nonsecret metadata, and explicit teardown. This pull request does not enable the experimental native runner for existing instances. > 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 with GPT-5. The model used extended reasoning, tool use, code execution, and parallel subagents. ## 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	2026-09-02 08:55:08 -05:00

Author

SHA1

Message

Date

DottaandPaperclip

72ff3a9f27

Measure native tool context and expand bounded workflow evals (#15218 )

## Thinking Path

> - Paperclip manages persistent agents and their assigned work.
> - Agents receive both fixed instructions and tool definitions.
> - Moving a procedure into a tool description still adds model context.
> - We need to measure the complete delivered catalog and test real
outcomes.
> - The tested reductions saved little space and introduced failing
outcomes.
> - This PR keeps measurement, bounded eval coverage and the original
evidence.
> - Production prompts, tools and runtime behavior stay unchanged.

## Linked Issues or Issue Description

Refs: #15151, #14961, #14948, #14985.

This adds the measurement and eval coverage needed to assess further
native
instruction changes. The attempted hiring and dependency reduction
failed
qualification and is excluded from the final diff.

## What Changed

- Measure the actual standard-mode tool authority, including all 39
tools and their input schemas. Capture scripted native start, resume and
continuation payloads and the OpenCode MCP declaration list.
- Add OpenCode to the two explicit-only local hiring/reuse and
delegation/feedback stories. Preserve the original task requests and
independent oracles.
- Apply one attempt per selected story and explicit company and
lead-agent budget stops.
- Select managed hiring credentials from the requested profile,
including OpenRouter.
- Run the existing Node test files under Node instead of collecting them
as Vitest suites.
- Retain sanitized comparison reports, original failed grades, source
hashes and evidence gaps.

## Verification

Final source: `2e7cef78eef7cdfe02265e0dcb03e855b8e50bd8`. Local
verification passes:

- Full repository `pnpm -r typecheck` and `pnpm build`.
- Eval-support unit tests: 1,252 Vitest checks and 128 Node checks.
- Eval TypeScript check and six final-source measurement tests. All 36
normalized
components across nine scripted deliveries and the OpenCode MCP catalog
match
the baseline exactly; the complete standing projection is 49,200 bytes.

Fresh review of this exact head is [5/5 with no remaining
findings](https://github.com/paperclipai/paperclip/pull/15218#issuecomment-5996723170);
all three review threads are resolved.
[Final-head
CI](https://github.com/paperclipai/paperclip/actions/runs/37354539126)
passes all 47 jobs, including repository typecheck, build, tests, runner
checks,
browser shards and the canary check. One initial annotation-test timeout
is
retained in attempt 1; its seven-test suite passed locally, and the
affected
CI lane passed on one targeted retry. No source or paid eval rerun was
needed.

Both ready-transition security scans passed with zero annotations. The
PR is
out of draft and conflict-free. GitHub still requires code-owner
approval for
the `package.json` test-script change; its requested reviewers are
already set.

A byte-for-byte comparison against master context
`a65ca0950834a85bb93bcc4b4042ecacdebfef53` confirms no production
changes under
`packages/` or production `server/` paths. The only server addition is a
measurement test. No new paid rerun is needed to compare unchanged
production
bytes. This does not claim that existing product defects have been
fixed.

The rejected corrected experiment had baseline **5 PASS / 1 FAIL** and
candidate
**3 PASS / 3 FAIL**, including **two newly failing pairs**. The later
readiness
experiment had candidate **3 PASS / 3 FAIL** and baseline **3 PASS / 2
FAIL / one
setup cell without a behavioral grade**. Its five comparable pairs had
two new
failures, two new passes and one unchanged pass. The missing baseline
Codex
cell never reached its provider step because Docker setup timed out.

The observed missing behaviors include parent continuation, waiting for
the
latest child revision, revised ZIP delivery and OpenCode credential
persistence.
Those failures remain failures. Source review and passing CI do not
regrade them.
The original reduction saved 460 bytes; its first repair saved only 125
bytes,
and the larger unqualified runtime repair increased the full projection.
None
of those production changes is shipped here.

Read the
[report](https://github.com/paperclipai/paperclip/blob/2e7cef78eef7cdfe02265e0dcb03e855b8e50bd8/doc/plans/2026-10-05-native-procedure-guidance.md)
and its linked sanitized receipts for exact
sources, original campaign links, pair-level results and evidence
limitations.

## Risks

- Full-catalog bytes are not model tokens, invoices, private vendor
prompts, lazy-loading behavior or truncation proof.
- The two stories are explicit-only and do not prove general coding
quality or arbitrary resume behavior. Single trials do not establish
causation or performance trends.
- The retained OpenCode candidate credential guard failed. Cleanup
removed the original provider database, so the precise persistence
mechanism remains unknown. The guard is unchanged.
- ACPX provider-execution IDs lack proven host-call mapping. No
extra-work or feedback-consumption claim is inferred by matching names,
order or counts.
- PostgreSQL cannot start locally while the host's shared-memory slots
are exhausted. Hosted CI must supply the full database checks; the full
local database suite is not claimed green.

## Model Used

OpenAI Codex, based on GPT-6, with repository tools and code execution.
The exact deployment model ID and context-window limit are not exposed
in this session.

## 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

---------

Co-authored-by: Paperclip <noreply@paperclip.ing>

2026-10-05 20:03:35 -05:00

DottaandPaperclip

b17019e14d

fix(agents): reduce default instructions and qualify stock harnesses (#14948 )

## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work.
> - Its adapters supply task context and access to Paperclip skills and
tools.
> - The default hire manual and shared prompts also repeat general work
procedures.
> - Those procedures overlap with stock provider instructions and the
Paperclip skill.
> - Existing E2E fixtures supply a QA manual, so they do not qualify the
production default.
> - This pull request reduces the generic instructions and adds real
default-hire coverage.
> - The benefit is less competing guidance, with inspectable evidence
for preserved skills and task context.

## Linked Issues or Issue Description

Refs: #14920. That merged change preserves native Codex base
instructions. This PR covers the default manual, shared legacy prompts,
operational skill guidance, and the narrowly approved ACP
skill-discovery/session-environment repair for measured delivery and
credential-persistence failures.

**What existing behavior does this improve?**

New non-CEO hires without a custom bundle and legacy task/chat startup
and continuation prompts.

**Current behavior**

The shipped default manual contains 602 words. Generic task/chat prompts
and ordinary resume deltas repeat work procedures already available
through the harness and Paperclip skill.

**Proposed behavior**

The default manual contains only the eight-word company identity. Shared
startup prompts retain identity and connection guidance. Ordinary resume
deltas retain current work context without the generic execution
contract.

**Reason and benefit**

Let the stock harness guide general work. Keep Paperclip-specific
capabilities and independently test default hires, skills, ordered
comments, and chat restart.

**Breaking changes**

New default hires receive less guidance. Existing saved manuals,
explicit custom bundles, CEO templates, and specialized wake contracts
retain their behavior. The obsolete includeExecutionContract option
remains accepted for source compatibility.

## What Changed

- Reduce the default hire manual to one sentence.
- Reduce shared task/chat defaults and remove the generic
ordinary-resume contract.
- Keep connection guidance, auth, skills, custom prompts, and
specialized wake context.
- Add credential-free instruction-boundary gates and 26 explicit Product
E2E cells across eight legacy/native profiles, including two focused
Paperclip-storage cases.
- Capture public hire receipts before providers run, then grade
delivered prompts and independent task/chat outcomes.
- Add an early legacy skill API recipe for saving a task document,
checking the saved revision receipt and linking the document. Improve
stock task/heartbeat skill-selection metadata and show a clickable
Markdown UI-link example. Keep native tool completion separate.
- Advertise bounded routing descriptions and exact successfully staged
SKILL.md paths in legacy ACP Claude; keep full bodies on demand and
preserve remote path rebasing.
- Remove only the provider environment from copied persisted ACP session
records, while loading current run credentials and preserving all other
options/conversation state.
- Regenerate both capability metadata inventories and reject stale
manifests/inventories before provider admission.
- Publish the original reduction and focused skill-repair comparisons,
preserving all failures, automatic recovery, cost coverage and
limitations.

## Verification

**Behavioral qualification remains pending.** Original legacy ACP Claude
loses the issue document only in the reduced cohort beneath an unchanged
credential failure. A source-backed diagnosis finds that neither
ordinary assignment reads the staged operational skill, while the
runtime persists provider environment in session state. The new common
repairs expose skill metadata/path and omit persisted env; strict
document and credential guards stay intact. [Inspectable diagnosis and
retained
hashes](https://github.com/paperclipai/paperclip/blob/9f654db4541d3d002769c988f6e51fc0b08dadbd/doc/plans/2026-10-03-legacy-acp-claude-readiness.md).

Current repair head `de0965984ff3edf611ae6d0e7ca5c7d5ae3947bb`
incorporates master `569c7203aa24b95440682983ce7940ba1d4247bd` (merged
#14961/#15007). All 222 affected adapter tests, adapter-utils/E2E
typechecks, and final 96 variant/grader/retry calibrations pass. The
frozen historical comparator is
`c25697f4260b6f3adfea143c3ae9932e2f42986d`: 8,280 of 8,291 paths
identical, exactly two production instruction paths plus nine declared
unit expectations differ. The operational skill/discovery/environment
repairs, selected model/profile/task/core grader/auth/permissions/retry
policy are identical. Both actual launcher prepare→verify admissions
pass with zero providers. [Immutable manifest and exact
receipts](https://github.com/paperclipai/paperclip/blob/9f654db4541d3d002769c988f6e51fc0b08dadbd/doc/plans/2026-10-03-legacy-acp-claude-evidence/manifest.json).

One original legacy ACP Claude cell per variant is authorized, with
enforced single campaign attempts, 12-minute deadlines and company/agent
1,000-cent hard stops; every product recovery run/cost is counted.
Actual live outcomes are pending. Current normal CI has one failed
server shard and failed aggregate verify under diagnosis; other normal
gates including typecheck/build/Rust/all eight browser shards pass.
Fresh review completed successfully; the valid historical startup/resume
masking finding was fixed with per-invocation task/chat checks and
strict complete-snapshot capture, calibrated and resolved. Prior heads,
failures and campaigns below remain historical evidence, not checks on
this repair head.

- Prior head `36aa4d81c49a1a8f6f04b1a068fae19aa901955f` is replayed on
merged hiring master `862a5758ba0e88a33232c1f1fa645e85c38a3113`. All 52
current-head checks pass with two intentional Storybook skips, including
repository typecheck/test/build and the browser shard. Fresh Greptile is
5/5 with zero unresolved review threads. Exact-head stock prerequisites
pass 599 assertions (598 TypeScript + 1 Rust), all six gates and
retained receipt verification, zero providers/source errors. Fingerprint
`a7f5a22d860a88fe20cce213c6d5e0004788f32c8363930729aea4fd740ad16d`.
Combined catalog/hiring calibrations pass 67 assertions, E2E typecheck
and 26-cell stock discovery pass. Canonical contract/inventory checks
and the later issue-derived reference calibration are retained; that
reference-only follow-up is not live-qualified by earlier frozen runs.
- Prior full repository typecheck/build passed. The complete local
Vitest run executed 14,956 tests: 14,870 passed, 83 skipped, three
timing failures. All three affected files passed unchanged narrow
reruns; original failures remain retained. Current-head CI now passes
the full general checks; the original local failures remain retained.
- The original 24-pair default-manual/shared-prompt comparison has two
new overall classic Claude/OpenCode document-delivery failures plus an
additional legacy ACP Claude document loss beneath an unchanged
credential-guard failure (not closed by later runs), two newly passing
OpenCode ordered cases, seven unchanged failures and 13 unchanged
passes. Equal 15/24 totals do not establish behavioral equivalence.
[Complete original
report](https://github.com/paperclipai/paperclip/blob/875f4c397d9e8c3f12f39dedd59abaf1eaf5236e/doc/plans/2026-10-02-stock-harness-live-comparison.md).
- The skill-only repair holds the eight-word manual/shared prompts and
merged #14920 fixed. All four matched profile configurations and 203
fixture/behavior files match. Candidate
`abd0b628ca642c09a54a4edc56a5227402f6686e` varies only the two skill
sources against baseline `bc83fe030234439ac51279502a28803958963e2e`.
[Candidate
workflow](https://github.com/paperclipai/paperclip/actions/runs/37060885547)
and [baseline
workflow](https://github.com/paperclipai/paperclip/actions/runs/37060888047)
each pass 571 exact-source prerequisites before providers; all eight
cells clean up successfully. Failed campaigns publish successfully and
remain failed.
- Repair pairs: Claude original Fail → Pass; Claude explicit Pass →
Pass; both OpenCode cases Fail → Fail. Explicit OpenCode's handoff
worsens beneath the unchanged failing UI-link grade: baseline gives a
clickable API URL, candidate gives a code-formatted path without an
anchor. The request's usable-link wording is narrower in the UI-only
oracle. [Complete repair report and safe
projection](https://github.com/paperclipai/paperclip/blob/875f4c397d9e8c3f12f39dedd59abaf1eaf5236e/doc/plans/2026-10-02-legacy-document-skill-repair.md).
- The subsequent narrow stock metadata/link correction has two matched
Pass → Pass cases, zero new machine failures/passes and no pending
pairs. Both original-case handoff links remain deficient: candidate uses
a wrong PAP prefix, baseline supplies a bare prefix-less slug path; the
preserved original oracle only requires a durable document. Both
explicit clickable UI-link cases pass revision/content/link grading. All
four exact-source 587-check gates, single assignment runs and cleanup
pass. This does not establish fix causality because baseline also
succeeds. [Candidate
workflow](https://github.com/paperclipai/paperclip/actions/runs/37069547401)
freezes `fe9dc1e3c518825242ed889ab9c8352986f8c2ed`; [matched
baseline](https://github.com/paperclipai/paperclip/actions/runs/37069552374)
freezes `0d7ecfa96d72fba79b7f0a25052b42c0686c0488`. This is a skill-only
comparison with reduced manuals/shared prompts held constant, not a
repeat of the historical-manual comparison. Only original and clarified
explicit classic OpenCode cases are selected, two per variant/four
expected turns. 8,242 other tracked files and both profile hashes match;
protected workflows admit each exact source before credentials.
[Complete qualification
report](https://github.com/paperclipai/paperclip/blob/74d0d3d945f4c52d0814b5a845ab5bd09f33cd6b/doc/plans/2026-10-02-opencode-skill-routing-link-qualification.md).
Candidate original loads Paperclip/reference before saving publicly;
baseline original loads it after writing locally, then saves publicly
within the same assignment. Reported cost totals are $0.0107824490
candidate / $0.0107909015 baseline, with unmetered runtime. The later
reference-only issue-derived link correction is provider-free calibrated
and **not live-qualified** by these frozen runs; no further paid runs.
- Retained tool calls show the repaired original OpenCode assignment
loads only its assigned output skill before writing locally. Operational
Paperclip is first loaded during automatic disposition recovery; its
early recipe is visible then, but it never saves the missing document.
Explicit candidate loads Paperclip and reads the new reference before
saving successfully. All nine actual runs are counted. Reported LLM
totals are $0.3802537209 baseline and $0.4918990161 candidate; local
runtime is unmetered.
- Initial setup, packaging, cancelled/missing-cell recovery, callback
test and relative-output attempts remain retained. No completed provider
failure was rerun. Frozen measurement branches are unchanged by later
canonical metadata maintenance.
- Run `pnpm test:e2e:runner:stock-harness`, `pnpm test:e2e:runner:unit`,
and `pnpm test:e2e:runner:typecheck`. Select `stock-harness` explicitly
for paid execution; it is excluded from `--all`.

Prior-head integration: `36aa4d81c49a1a8f6f04b1a068fae19aa901955f`
replays this PR on merged hiring #14985
(`862a5758ba0e88a33232c1f1fa645e85c38a3113`), preserving the four
explicit custom-CEO-bundle checks, minimal generic manual boundary, and
both suites. The combined fixture catalog and hiring calibrations pass
67 assertions; exact-head stock prerequisites pass 599 assertions (598
TypeScript + 1 Rust), all six gates and retained-receipt verification,
zero providers/source errors, fingerprint
`a7f5a22d860a88fe20cce213c6d5e0004788f32c8363930729aea4fd740ad16d`. E2E
typecheck and 26-cell stock discovery pass. Fresh current-head CI passes
all 52 checks with two intentional skips, and fresh Greptile is 5/5 with
zero unresolved review threads.

The prior source-plan browser failure is retained: a deterministic
process fixture replayed its last `fixture:plan` command on
`chat_task_completed`, writing revision 2 with identical body after the
approval handoff. This was not paid provider execution. Rebased
current-head CI passes the same assertion without an old-head retry or a
change to that browser fixture.

The merged hiring change was measured separately on immutable matched
unions, with this reduced/shared/operational context and native
completion guidance held constant. [Complete original two-profile
report](https://github.com/paperclipai/paperclip/blob/f0512647656be78e48abd8c22a3078db8bf6bcd2/doc/plans/2026-10-02-hiring-template-live-comparison.md):
[candidate](https://github.com/paperclipai/paperclip/actions/runs/37075466208)
/ [historical
baseline](https://github.com/paperclipai/paperclip/actions/runs/37075469463),
705 provider-free prerequisites each. Both pairs are unchanged Fail →
Fail on the exact-five count, with six core delivery checks passing all
four cells; 28 actual successful runs include eight automatic completion
wakes, zero retries, four successful cleanups. Source-read coverage is
uncomparable, actual model charges unknown. Separately versioned
provider-free accounting remains analytical work; original verdicts are
preserved. This does not rerun or qualify the completed default-manual
or native campaigns.

## Risks

- Legacy ACP Claude's additional delivery loss is not closed by any
later matched run and blocks the no-extra-failing-behavior merge
criterion. Legacy document delivery may have relied on the prior
manual/shared prompts. The early skill repair improves Claude in one
trial; the later OpenCode pairs pass in both variants and cannot
establish causality or robust recovery. Both original-case links remain
deficient beneath the storage-only grade. The later issue-derived
reference correction has only provider-free validation. Native
finish/block descriptions must not be supplied to legacy agents.
- The comparison holds merged native Codex fix #14920 constant; it
cannot measure that fix's before/after task performance.
- These bounded skill/context/chat workflows do not measure general
coding quality. Unrepresented providers remain unqualified.
- Saved manuals and old Codex sessions are not automatically migrated.
Codex through ACP still has a separate base-instruction follow-up.

## Model Used

OpenAI Codex, GPT-6 family as identified by this session. The exact
deployment ID and context-window size are not exposed. The assistant
used reasoning, repository tools, code execution, and delegated PR/eval
work.

## 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 (relevant suites and all
three unchanged narrow reruns pass; complete-run timing failures
retained in Verification)
- [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
- [ ] All Paperclip CI gates are green on the new repair head
(prior-head checks retained above)
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
on the new repair head
- [x] I will address all Greptile and reviewer comments before
requesting merge

---------

Co-authored-by: Paperclip <noreply@paperclip.ing>

2026-10-03 12:32:42 -05:00

Dotta

5716fe907e

test(runner): add full-stack acceptance and eval gates (#12700 )

## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work.
> - The runner subsystem executes agent work across local and managed
provider backends.
> - The lower pull requests restore the task runtime, provider backends,
and managed-provider control plane.
> - The restored system needs repeatable full-stack checks before it can
ship safely.
> - Paid live checks also need clear access, cost, and secret controls.
> - This pull request adds acceptance, live evaluation, chaos, and
release gates for the restored runner stack.
> - The benefit is measurable runner parity with safer release
decisions.

## Linked Issues or Issue Description

**Subsystem affected**

Cross-cutting. This change covers runner tests, release workflows,
server contracts, and evaluation tools.

**Problem or motivation**

The runner stack did not have one complete acceptance surface for native
Codex, ACPX, Claude Managed, and AWS AgentCore. Release checks could
miss provider drift, task-view regressions, cost-policy errors, and
destructive cleanup errors.

**Proposed solution**

Add a 57-cell full-stack catalog, a Daytona image, and opt-in paid
workflows. Add live evaluation, chaos, cost-limit, redaction, and
release contract checks. Add AWS AgentCore infrastructure and guarded
provisioning tools. Keep the native runner experimental flag off by
default.

**Alternatives considered**

We considered manual smoke tests only. They do not give repeatable
evidence and they do not protect release branches. We also considered
one large pull request. The stacked pull requests keep each review below
the Greptile file limit.

**Roadmap alignment**

This work supports the shipped Cloud / Sandbox agents milestone and the
shipped Agent evals & feedback milestone in `ROADMAP.md`.

Related stack:

- #12699 adds managed provider backends and lifecycle support.
- #12691 adds qualified OpenCode and ACPX provider backends.
- #12685 restores task runtime rendering and steering.

## What Changed

- Add the runner full-stack harness with 57 catalog cells and 60 unit
tests.
- Add a Daytona runner image with digest-pinned base images and
base-aware image-content checks.
- Add guarded live evaluation and chaos workflows with a fixed
40-execution matrix; live and full-stack paid schedules now run only on
Sundays or by manual dispatch.
- Add in-flight reported-usage cost stops, post-turn cost caps,
exact-threshold failure classification, secret redaction, retry
classification, and actor authorization.
- Reattach stream and hard-budget listeners before restart-recovery
continuations so restored paid sessions cannot bypass in-flight
interruption.
- Preserve OpenCode usage and cost across tool-loop messages and turns
while exposing an explicit current-run delta to durable accounting.
- Keep PNG/WebM evidence in access-controlled artifacts only, reject
SVG, and publish only pruned inert structured per-attempt evidence.
- Add AWS AgentCore infrastructure, provisioning checks, and smoke
tools; reject unsafe model identifiers, require exact stack ownership
markers, and make failed-stack replacement explicit.
- Add evaluation-session contracts and capability reports.
- Add release workflow checks for immutable action pins, frozen
dependency installs, exact weekly cron shape, paid-run guards,
provider-secret isolation, and chaos test paths.
- Reauthorize the original and triggering numeric actor IDs as the first
step of every provider-secret job, including partial reruns, before
checkout or provider access.
- Give each full-stack matrix cell only its matching provider
credential, expose Daytona only to Daytona cells, and disable shared
dependency caches anywhere paid credentials or OIDC write access are
present.
- Protect the legacy manual E2E workflow with the same default-branch,
allowlist, environment, and per-job authorization boundary.
- Rotate live-eval candidates by week and retain 120 days of compatible
history so the seven-week trend window remains viable.
- Restore the root runner-acceptance commands and reconcile reported
snapshots,
raw receipts, and terminal usage without double counting or losing late
usage.
- Mark ACPX token deltas exact only when every budget field is present,
keep
cumulative cost/request authority separate, reject non-USD cost
labeling,
  and include thought tokens in output-token budgets.
- Keep `enableNativeRunner` off by default. The acceptance harness
enables it only in its isolated test instance.

## Verification

Passed locally:

- `pnpm --filter @paperclipai/paperclip-runner typecheck`
- `pnpm test:runner-acceptance:typecheck`
- `pnpm test:runner-acceptance` (19 tests)
- focused OpenCode proxy, driver, runnerd transport, live-session, and
turn-stream tests (106 tests)
- `pnpm --filter @paperclipai/paperclip-runner exec vitest run
src/live/clean-room-server.test.ts` (22 tests)
- `pnpm test:e2e:runner:typecheck`
- `pnpm test:e2e:runner:unit` (62 tests)
- `node --test scripts/__tests__/release-verify-workflow.test.mjs`
- `pnpm --filter @paperclipai/paperclip-runner
test:runner-workflow-evals` (22 tests)
- `pnpm -r typecheck`
- `pnpm build`
- `node --test
packages/paperclip-runner/scripts/aws-agentcore-provisioning.test.mjs`
(6 tests)
- `git diff --check`
- `cargo test --manifest-path
packages/paperclip-runner/runner/Cargo.toml -p paperclip-runner-core
--lib --locked` (161 tests)
- focused ACPX provider-event tests (10 tests)
- The rebased PR changes 92 files. `pnpm-lock.yaml` is unchanged.

I did not run paid live provider jobs or provision AWS resources. Those
checks need credentials and can create cost.

## Risks

The paid workflows can create provider cost. They require an allowlisted
original and triggering actor, the protected `runner-e2e-paid`
environment, explicit opt-in variables, and cost limits. The four
provider credentials exist only in that master-only environment, which
requires allowlisted reviewer approval and disables administrator
bypass; repository and organization Actions scopes contain no copies.

Provider usage arrives after a billable request, so the live guard
cannot prevent one request from crossing a threshold. It interrupts
immediately on the first reported threshold hit and permits no
continuation.

Visual evidence can contain secrets rendered as pixels. PNG/WebM remain
only in access-controlled workflow artifacts; SVG and per-attempt XML
are excluded, and S3/Pages receive a pruned structured dashboard.

The AWS scripts can create cloud resources. They use explicit commands,
least-privilege roles, KMS encryption, saved nonsecret metadata, and
explicit teardown.

This pull request does not enable the experimental native runner for
existing instances.

> 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 with GPT-5. The model used extended reasoning, tool use,
code execution, and parallel subagents.

## 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

2026-09-02 08:55:08 -05:00