PaperClipAI/tests/runner-e2e/CHAT-FEATURES.md

# Chat feature evals

`chat-features` promotes executable entries from the [scenario catalog](chat-scenarios/README.md)
into real Product E2E executions. The two original task journeys remain intact.
Additional domain cases use `chat-catalog-contract.ts` and the explicit registry;
listing a cell or calibrating its oracle is not live acceptance. The complete
244-scenario denominator and unchanged requests are frozen in
`chat-campaign-manifest.json`. Its raw-file/defaults inventory hash is
`9dc9cf6a5426cf1b56bc30a3571851f233f9dd5e93a1292eaee414d33b52df77`.

Domain cases own production API setup, explicit user followups, observations,
case-specific grading and optional cleanup. The shared driver submits the exact
catalog query, saves prerequisite state before action, records real API comments,
run attribution and rendered links, and rechecks after browser reload. Every case
declares run/time limits and an oracle version. Missing checks and contradictory
grades fail. Known `coverageGaps` make a diagnostic case fail acceptance until its
conditions are implemented; they are not blocked cases reported as passing.

The registry contains only executable setups/actions/oracles. Scenarios absent
from it remain in the denominator as unimplemented; optional external services
must be real or explicitly labeled simulations. The frozen campaign ledger owns
implementation readiness and all attempts separately from catalog discovery.
No model judge supplies a generic pass.

Discovery reads pure `chat-catalog-descriptors.json` metadata and does not import
any domain. The runtime loads the selected area's module and rejects descriptor
bounds/oracle/policy drift. `chat-catalog-fingerprint.ts` compiles imports to find
real local dependencies, including transitive helpers and the selected domain,
and binds the normalized execution, exact scenario/defaults, package versions and
lockfile. Whole source revision and global suite hash remain separate provenance.
The six reviewed dynamic resolutions are pinned to their callsite source hashes;
source drift falls back to conservative source coverage. Installed or external
journey overrides refuse the reviewed local-chat lane. No global loader hook is
installed. This permits unchanged cases to retain a measurement when an unrelated,
unloaded domain changes; a changed executable dependency invalidates that case.

Standard funding/readiness and successful native runs are defaults. A scenario
that preserves a special funding or paused state declares a setup policy and
must independently verify its named prerequisites before the query. Declared
terminal counts distinguish cancellation and denial from successful execution.
Creation-only journeys may explicitly allow bounded separate worker execution
while requiring a settled source and their own exact task/assignee oracle.
Authenticated identity cases select the existing human-session fixture path;
private secondary-user credentials join the attempt's redaction sink. Scenarios
whose stated context has no working chat require a genuine boundary journey;
these cannot receive a native chat pass from seeing an onboarding or login gate.


| Scenario | Actual user request | Independent outcome |
| --- | --- | --- |
| `tasks.create-backlog` | Add a task for Morgan to draft the launch announcement in Launch, but don't start it yet. | Exactly one task, Launch project, Morgan assignee, backlog at creation, no worker execution, usable chat link after reload. |
| `tasks.delegate-announcement` | Have Morgan write a two-sentence launch announcement using the Launch project brief, save it on the task, and send me the result here when it's finished. | Exactly one task, actual Morgan execution, current worker-authored document grounded in the project brief, task Done, automatic completion response in the original chat, usable result after reload. |

The queries are read directly from `chat-scenarios/tasks.yaml`. Context and
expected outcomes stay evaluator-side. The fixture creates a disposable company,
CEO, Morgan and Launch project through production APIs; both agents use the
shipped role instructions. The brief contains a product, date and random invitation
code absent from the user query. The browser sends one message through Agent Chat.
The harness never creates the result task, writes the announcement, completes the
task, or sends a hidden follow-up.

## Run the first slice

Follow the installation steps in [README.md](README.md#local-commands). The current
profile is local native Codex using the catalog's recorded default model and an
explicit `CODEX_AUTH_JSON` login. Load that login privately into the launcher
environment; do not print it or save it in tracked configuration. The existing
subscription fixture stages only `auth.json` in the disposable company login slot
and removes it with the isolated instance.

```sh
pnpm test:e2e:runner:typecheck
pnpm test:e2e:runner:unit
pnpm test:e2e:runner -- --list --suite chat-features
pnpm test:e2e:runner -- --id chat-features.runner-codex-subscription.local.create-backlog --max-automatic-retries 0
pnpm test:e2e:runner -- --id chat-features.runner-codex-subscription.local.delegate-announcement --max-automatic-retries 0
```

Run creation first and inspect its packaged report before delegation. For the historical first-slice repeatability probe, each cell used three fresh fixtures on the same definition revision. The full-catalog campaign instead allows one initial attempt and one documented fix-and-rerun per failing scenario, with automatic retries disabled. Keep every attempt, including failures; a retry does not erase one. A
fixture or product correction starts a new three-sample measurement. Both cells
are explicit-only and excluded from `--all`; this does not implement or qualify
the rest of the YAML inventory.

## Bounds and grading

Creation expects one accounted native work run; delegation expects three (CEO
handoff, Morgan work, CEO completion response). Runs have a 120-second timeout
and attempts a seven-minute deadline. The visible task and completion reply must
settle within five minutes after the human request, including a five-second stable
observation. This is an independent journey acceptance bound: three runs may not
each consume their maximum timeout and still pass. The outer attempt also covers
fixture setup, reload and link checks, and evidence collection. A failed provider
run rejects immediately; the historical missing turn-start event was not a
five-minute poll timeout. Automatic retries are disabled. Company
and agent budgets are 500 cents each, CEO daily work runs are capped at three,
Morgan at one, and periodic heartbeats are disabled. Initial production hello
probes fall outside heartbeat usage and budget accounting; detailed setup tokens
and cost are unavailable. These limits are not an absolute inference or cash cap.
Subscription cost coverage must remain explicit in the report.

Because this subscription currently reports unpriced usage, setup explicitly sets
`unpricedUsagePolicy: allow` on the disposable company and both agents through the
public budget API and verifies all three policies. The 500-cent hard stops remain
enabled for reported costs; they cannot cap unknown dollar usage. Run caps and
deadlines remain independent bounds. Without this explicit funding prerequisite,
the default policy pauses the company after the first unpriced run and cancels
the delegated worker. Such a pause is not a successful chat journey.

`chat-feature.json` retains the scenario ID, capability IDs, grader version,
saved records, audit attribution, rendered links and individual checks.
`chat-feature-runs.json` uses the existing native run evidence collector, including
on failures. Setup evidence records fixture IDs, synthetic brief, readiness and
saved budgets. Ordinary Product E2E evidence packaging, redaction, cleanup, failure
classification and dashboards own the result. The definition digest includes the
query, defaults, flow and grader sources.

The grader rejects duplicate tasks, wrong owners/projects/companies, early execution,
missing audit attribution, missing or stale deliverables, extra human nudges and
uncorrelated completion replies. Document provenance binds the current revision
to Morgan's actual run. The completion check reuses `completionDelivery`; a
handoff promise or a marker alone cannot establish delivery. Chromium verifies
the task link and reload persistence.

The checks establish the specified saved state, two-sentence structure, required
brief facts and result access. They do not score writing quality or every semantic
claim in the response. Inspect the retained conversation and announcement before
accepting the first samples; the existing completion semantic-review rubric still
applies. Oracle calibration tests exercise both valid evidence and plausible wrong
outcomes without calling a model.

Preserve the historical measurements below when promoting another scenario. Keep each
starting condition explicit: connected versus missing connection, allowed versus
approval-required, and so on. Reuse the fixture/report infrastructure and add a
case-specific outcome oracle; do not make all YAML paragraphs executable through
an unconstrained model judge.

## Initial measurement — 2026-10-26

Three fresh samples per cell on definition
`7fb20426ce3f1dd4efeddfb79f231336ea65a16b9c2a73bf6131e997acd74e9d`
produced **5/6 passes**: backlog creation 3/3 and delegation 2/3. These used local
native Codex subscription execution with `gpt-5.6-sol`. All six cleaned up
successfully. Both workflows have successful browser-to-provider evidence, but
the initial repeatability target is not met.

In campaign `local-2026-10-26T15-48-12-199Z`, the first CEO run in delegation
failed before creating a task: `runnerd did not report the provider turn identity`.
Retained events show submission and provider startup, followed by the transport's
turn-start deadline. The precise cause of the missing event remains unresolved;
the existing machine classification is retained without claiming that the provider
or infrastructure has been proven responsible. Investigate this before expanding
the suite. No replacement sample erased this failure.

The local `results/chat-feature-acceptance/acceptance.md` records every sample and
earlier failed revision; its sibling dashboard uses the standard report generator.
Results are ignored generated artifacts, so retain or publish them through the
existing Product E2E pipeline when handing this measurement to another machine.

### Follow-up measurement — 2026-10-26

The unchanged delegation cell subsequently passed three fresh executions:
`local-2026-10-26T16-50-50-751Z`, `local-2026-10-26T16-54-58-203Z`, and
`local-2026-10-26T16-58-39-074Z`. The latter two campaigns also ran backlog
creation concurrently and passed both cells. All five selected executions used
one attempt with zero automatic retries, retained the same definition above,
and passed cleanup. Five existing turn-identity, early-terminal and runner-exit
regressions passed as well.

These are green rechecks, not a demonstrated repair of the earlier startup
interruption: no production code, timing limits or oracle assertions changed,
and the missing turn-start event did not reproduce. Keep the original failure
in the history. Across the original final-definition batch and this recheck,
backlog is 5/5 and delegation is 5/6. The local
`results/chat-feature-recheck/verification.md` records the follow-up evidence;
its standard dashboard is explicitly scoped to the latest campaign.

## Full catalog campaign

The frozen denominator is 244 requests across 35 areas and 191 capabilities. The [durable progress record](../../doc/plans/2026-10-26-chat-catalog-eval-progress.md) and `chat-campaign-manifest.json` distinguish complete, partial, unimplemented, blocked and actual measured results. Discovery and oracle calibration are not live acceptance. Each canonical scenario has one initial attempt and at most one coherent correction/rerun, with automatic retries zero; failures and unrun scenarios remain in the denominator.

Domain descriptors declare real setup/actions/independent grading and exact bounds. `setupHumanMessageCount` defaults to zero; genuine earlier human turns and native setup runs remain in the declared total, with persisted comment/run IDs captured before the canonical query. No historical model reply is fabricated.

Only `steering.preserve-draft` and `tasks.attachment-brief` may currently declare one `queryAttachments` receipt. The canonical typed query remains unchanged; the production composer appends its exact attachment reference. Before and after send, the harness independently reads the public same-source receipt and actual bytes, verifies SHA-256 and human attribution, and binds it to the real canonical comment. Extra markdown, text changes, foreign receipts and hidden followups fail. Attachment fingerprints pin both production composer source hashes as data, without importing the UI graph.