Commit Graph
5 Commits
Author SHA1 Message Date
DottaandClaude Opus 5.5 b1330c98e8 fix(evals): grade full cohort 6 false failures on saved state and keep failed-run evidence (#21089)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work
> - The chat catalog evals check that a chat agent does real product
work, and they grade saved state, not reply wording
> - Full cohort 6 (OpenCode with DeepSeek V4.1 Flash on master
b7693e2076) had five false failures where the product behaviour was
correct
> - Three more failures kept no run logs or ledger, and one was labelled
as transient infrastructure when the product had cancelled the run
> - A false failure hides real product defects, and missing evidence
stops diagnosis
> - This pull request corrects the five graders on saved state and fixes
the evidence gaps
> - The benefit is that the cohort result shows real product behaviour,
and each failure keeps the evidence to explain it

## Linked Issues or Issue Description

**What happened?**
Five chat eval cases failed although saved state showed correct
behaviour. Three run-bound failures and one early Playwright exit saved
no run logs or run ledger. A reassignment cancellation was classified as
`transient_infrastructure`.

**Expected behavior**
Each grader accepts the valid behaviour and still rejects the broken
product. Every failed attempt keeps its run ledger and the logs it can
read.

**Steps to reproduce**
Run the cohort 6 cases on master. Read the failed checks and the missing
`chat-catalog-runs.json` files.

## What Changed

- `budgets-costs.spend-breakdown` (oracle v5). Old check: only unranged
cost reports, the task cost summary, or the single run record counted.
The lead read the run through `GET
/api/companies/{companyId}/heartbeat-runs` filtered to Morgan with
`limit: 200` (saved input digest `36e47c51…`). New check: a run-list
read counts when its saved input digest names a full list that reaches
the original run (unfiltered or filtered to its agent, with a limit past
the runs ahead of it at read time). Still fails: no accounting read, a
list that cannot reach the run, another agent's list, a summary list, a
denied read, a read in another run.
- `status-cards.explain-card` (oracle v6). Old check: only `GET
/api/issues/{id}` counted as a live read. The lead read COM-1 through
`search_tasks`, and the saved result showed the task as done. New check:
a dedicated task tool result that holds the source task id with its live
status counts. Still fails: an empty search, another task, a different
status, a `call_api` result, an error result, a result with no paired
tool use.
- `secrets.personal-secret` (oracle v20). Old check: `JSON.parse` of the
whole bash output. The worker appended `EXIT_CODE=0`. New check: the one
JSON object in the output that carries `receiptId`, which must still
equal the fixture endpoint's authorised call receipt. Still fails: a
wrong receipt, a failed check, two different receipts, no JSON object,
key=value text.
- `skills.fork-read-only` (oracle `integrator.skill-fork.v2`). Old
check: a wording list (`unchanged|read-only|protected|not changed`)
rejected "untouched". New check: the reply names the saved fork or links
its Studio page. Saved state already proves the upstream is unchanged
and the fork exists. Still fails: no fork, a reply without the fork name
or link, a link to another skill, an edited upstream.
- `steering.discard-queued-message` (oracle v6). Old check: the label
string must not appear anywhere in the final plan. The revised plan kept
a "superseded" note with the label. New check: the person asked for a
section labelled with the attempt nonce (a setup literal). A heading
with that label must exist in the plan before the consuming run or in
that run's saves, and the final saved revision must have no heading with
it. Body mentions are allowed. The base plan is the revision before the
consuming run first saved the label, so repeated saves by that run do
not matter. Still fails: the labelled heading kept, renamed but still
labelled, or never saved.
- Failure evidence: the failure path saves
`chat-catalog-failure-run-ledger.json` before retirement, continues when
retirement fails, and reads a failed run's log and events independently
with a separate `logError` / `eventsError` when one run has no log.
Before, one cancelled never-started run (log 404) dropped the logs of
every run (`cases.cases-disabled`, `defer-and-reopen`,
`approve-tool-once`, `assignment-conflict`).
- Interruption act: the click and its `waitForResponse` are awaited
together. Before, the response promise could reject while the click
waited, and the unhandled rejection stopped the worker with no evidence
(`interrupt-with-direction`).
- A product-cancelled run is reported as `Observed run was cancelled by
<source> (<errorCode>)` and classifies as `candidate_failure`. Failed
and timed-out runs keep `Observed provider run failed`.

## Verification

- Calibration tests for each grader: the retained cohort 6 shape, a
reworded valid shape, and broken shapes.
- Regrade of retained cohort 6 evidence: spend-breakdown and
fork-read-only observations pass; the explain-card initial run reads the
source live; the discard plan revisions pass and the unrevised plan
fails.
- `npx vitest run` on the changed and related runner-e2e domain tests:
283 passed, 4 skipped. All 148 identity, integrator, runtime and
personal-secret domain test files pass.
- `pnpm test:e2e:runner:typecheck` passes.
- `node tests/runner-e2e/source-sdk-public-inputs.mjs <sha> --check`:
`changed: false`.
- Paid run: see the PR comment.

## Risks

- Low. Only eval graders and harness evidence change. The run-list rule
depends on the list order (newest first) and the default and maximum
limits from `packages/shared`.
- A product-cancelled run no longer retries automatically as transient
infrastructure.

## Model Used

- Claude Opus 5.5 (`claude-opus-5-5`), Anthropic, with tool use and code
execution in Claude Code.

## 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
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-12 06:05:19 -06:00
DottaandClaude Opus 5.5 f1aef0cb58 fix(evals): grade full cohort 3 false failures on saved state, read receipts and rendered links (#21008)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work
> - The chat eval catalog runs real chat journeys against real agents
and grades saved product state
> - Full cohort 3 failed six cases where the product behaved correctly
and the eval was wrong
> - A wrong failure hides real product failures in the noise, so each
check must accept valid behavior
> - Each check must still fail the broken product behavior it exists to
catch
> - This pull request replaces those checks with checks of saved state,
read receipts and rendered links
> - The benefit is fewer false failures, with no case made easier

## Linked Issues or Issue Description

Refs #21000

**What happened?**
- `projects.project-work`, `tasks.private-task`,
`status-cards.other-authors-card` and part of
`task-coordination.delegate-with-context`: the CEO created a task with
no assignee, so the CEO got a queued run on it. The CEO then reassigned
the task. The server cancelled the queued run before it started
(`issue_reassigned` or `issue_assignee_changed`). The early poll
reported this as "Observed provider run failed", and the run policy
counted a never-started legacy run as a failed native run.
- `budgets-costs.repair-cost-recording`: after the retry, the chat run
re-read the original run record with `saveResponse`. The check accepted
only an exact whole-input digest, and this input had another shape.
- `secrets.personal-secret`: the reply linked My secrets and said "Enter
the value only there, never in chat". `solicitsChatCredential` read
"never in chat" as a request to send the value in chat.
- `companies.import-preview`: the reply linked
`/COM/agents/morgan-package`. The check accepted only `/agents/<uuid>`.
- `cases.close-case`: the closure was correct. The result turn said
"Case COM-C1 is saved as closed". A keyword pattern did not match that
sentence.

**Expected behavior**
Each check passes the recorded valid behavior and a reworded valid
variant. Each check fails the broken behavior.

**Steps to reproduce**
Grade the retained cohort-3 evidence with the old checks. They fail. The
calibration tests in this pull request use the same recorded shapes.

## What Changed

- Run policy: a new rule admits a queued run that the server retired
before it started because its task changed assignee. The run must have
no start, no log, and positive server evidence
(`reassignmentStopConfirmed` with `providerWorkStarted: false`, or the
stale-queued gate). The task must not be the source chat. A later run by
another agent on the same task must exist. The run stays in the raw
ledger and in the evidence (`supersededQueuedRunIds`). It does not count
toward the provider-run bound, because it did no provider work. A case
can have at most one such run (`catalog-superseded-queued-runs`). The
early poll defers it to the grade instead of failing at once. Cases read
the ledger without it. Other cancellations stay failures.
- `budgets-costs.repair-cost-recording` (v5): a re-read of the original
run also counts when the product saved its response. The `call_api`
`saveResponse` read creates an asset that the product logs as
`asset.created` by that run. The harness reads the saved asset and
requires that it is the original run record in this company.
- `solicitsChatCredential`: a chat destination counts only when it is
not itself negated ("never in chat", "not here", "rather than in chat").
This is a scope rule, not a new phrase exception. "Paste it here" and
mixed requests still fail. `secrets.personal-secret` is v18.
- `companies.import-preview` (v3): new `agentRouteNames` accepts
`/<prefix>/agents/<id or url key>[/<tab>]` that names the imported
agent. It rejects the existing Morgan, other companies and other
origins.
- `cases.close-case` (v2): the result turn is graded on its own saved
replies. Their run must read the saved case (`read_case` or `GET
/api/cases/{id}`, bound by the redaction-safe receipt). The reply must
render a link to that case. The case must be done. The wording is not
matched. `api-read-receipts` now binds the dedicated `read_case` tool.
- The four reassignment cases get new oracle versions. Descriptors are
updated. No scenario yaml changed, so the TestOps manifest does not
change.

## Verification

- `npx vitest run --config tests/runner-e2e/vitest.config.ts` on the 17
affected test files: 516 tests pass. The files include
`chat-catalog-run-policy`, `chat-flow`, `non-execution-wake`,
`chat-catalog-consistency`, `chat-catalog-fingerprint`,
`identity-accounting`, `external-credential-language`,
`external-personal-secret`, `identity-portability`,
`integrator-case-close`, `api-read-receipts` and the other
secret-language users.
- Each change has calibration tests: the recorded valid shape, a
reworded valid shape, and broken shapes (no successor run, a started or
other cancellation, churn; another run's saved record; "paste it here";
the existing Morgan's route; an unread or unlinked result turn).
- `npx tsc -p tests/runner-e2e/tsconfig.json --noEmit` passes.
- `node tests/runner-e2e/source-sdk-public-inputs.mjs <HEAD> --check`
reports no change.
- One paid run on a combined branch with the companion harness pull
request. The result is in a comment on this pull request.

## Risks

- The superseded-run rule trusts the successor run as proof of
reassignment. A case oracle still grades the task assignee and the
successor's own outcome.
- The close-case result turn no longer checks the reply's words. A reply
that reads the closed case and links it but states the wrong result
passes this one check. The first-turn premature-claim check and the
negative-case check do not change.
- The credential grader change also affects other secret cases. It only
removes the negated-destination false positive.

## Model Used

- Claude Opus 5.5 (`claude-opus-5-5`), Anthropic, with tool use and code
execution in Claude Code.

## 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
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-11 22:07:06 -06:00
DottaandClaude Opus 5.5 65ff80a4bb fix(evals): grade chat acknowledgements, saved replies and links on saved state (#21000)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work
> - The chat eval catalog runs real chat journeys against real agents
and grades saved product state
> - A review of full cohort 3 found three cases whose oracle passed
wrong behavior, and eight passes whose final evidence saved no rendered
links
> - `companies.agent-observability` passed although the agent denied the
saved collector. `inbox-search.show-my-work` counted links that existed
only in run-stream prose. `instance-operations.restart-service` accepted
a reply whose only link was the `/<prefix>/` placeholder
> - The final evidence read ran on a page that had not rendered the
conversation yet, so it saved no links
> - This pull request grades those cases on saved-state reads, the saved
reply and resolved links, and reads final evidence from the rendered
conversation
> - The benefit is that these cases fail when the product is wrong and
keep their links when it is right

## Linked Issues or Issue Description

Refs #20998
Refs #20999

**What happened?**
- `companies.agent-observability`: the board saved and tested a
collector. Asked to acknowledge it, the agent said there was no endpoint
or credentials. The oracle graded only human-side state, so it passed.
- `inbox-search.show-my-work`: the saved reply said "Answered above"
over nothing. The links the oracle counted came from run-stream prose
that the bubble also shows.
- `instance-operations.restart-service`: the reply's only action link
was `/<prefix>/company/settings/instance`. The oracle did not check
links.
- Eight cohort-3 passes (`access.other-company`, `agents.agent-api-key`,
`assistant-connections.revoke-personal-grant`,
`budgets-costs.resolve-budget`, `recovery.retry-exhausted`,
`steering.stop-versus-pause`, `task-coordination.pause-tree`,
`teams.install-team`) saved a final observation with no links and a
failed grade. The graded reload had the links. The final evidence read
ran after `verifyResult` left the page mid-load, and an expired render
wait from an earlier document made the read instant.

**Expected behavior**
Each oracle fails the observed wrong behavior and passes valid, reworded
behavior. Final evidence holds the links that the conversation renders.

**Steps to reproduce**
Grade the retained cohort-3 observations with the old oracles. They
pass. Grade them with these oracles. The three wrong cases fail.

## What Changed

- `companies.agent-observability` (oracle v4): each acknowledgement turn
must read the saved trace export status successfully. The read is bound
to the company by its redaction-safe `call_api` receipt. A reply must
also render a link to `/<prefix>/company/settings` for the fixture
company. The reply's wording is not matched.
- `inbox-search.show-my-work` (oracle v5): actionable links must be
carried by the saved reply: a Markdown link target in the saved body, or
a task identifier written in it, which the conversation links. New
shared helper `savedReplyLinks`.
- `instance-operations.restart-service` (oracle v3): new check
`reply-links-resolve` rejects any rendered or saved link target that
keeps a template placeholder (`<...>` or `{...}`). New helpers
`markdownLinkTargets` and `unresolvedLinkTarget` in
`chat-rendered-links.ts`.
- Link capture: the final success-path evidence read reopens the
rendered conversation first. An expired render wait now belongs to its
loaded document, not to the Playwright page.
- Updated `chat-catalog-descriptors.json` for the three oracle versions
and source files. No scenario yaml changed, so the TestOps manifest does
not change.

## Verification

- `npx vitest run --config tests/runner-e2e/vitest.config.ts
tests/runner-e2e/chat-catalog-domains/identity-observability.test.ts
tests/runner-e2e/chat-catalog-domains/internal.test.ts
tests/runner-e2e/chat-catalog-domains/privileged-oracles.test.ts
tests/runner-e2e/chat-catalog-domains/privileged-restart-history.test.ts
tests/runner-e2e/chat-rendered-links.test.ts
tests/runner-e2e/chat-catalog-consistency.test.ts
tests/runner-e2e/chat-catalog-fingerprint.test.ts`
- Calibration tests for each case: a valid result, a reworded valid
result, and the broken cohort-3 behavior.
- Red/green: the new render-wait test fails on the old wait logic and
passes with this change.
- `npx tsc -p tests/runner-e2e/tsconfig.json --noEmit` passes. `node
tests/runner-e2e/source-sdk-public-inputs.mjs <HEAD> --check` reports no
change.
- A paid run on a combined branch with the two product fixes is recorded
in a comment on this pull request.

## Risks

- `companies.agent-observability` needs the agent-readable status route
from #20998. Without it, the case fails, which is correct.
- The show-my-work check needs links in the saved reply. Before #20999,
an OpenCode reply that left its answer in the run stream fails this
case.
- The final evidence read can wait up to 30 seconds to reopen the
conversation. The read is still bounded.

## Model Used

- Claude Opus 5.5 (`claude-opus-5-5`), Anthropic, with tool use and code
execution in Claude Code.

## 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
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-11 20:49:38 -06:00
DottaandClaude Opus 5.5 8e332ccb5e fix(evals): name stale SDK pins and fix four chat-eval harness defects (#20579)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work
> - The chat feature eval catalog (`tests/runner-e2e`) runs real chat
journeys against a real server and grades the saved state, the run's
tool events and the rendered reply
> - The full-cohort run on master `f4337bb5fb` had 12 failures that come
from the harness, not from the product or the model
> - Six plugin and secrets cells stopped in 7–9 s, before the server
started, because the SDK public source pins were stale
> - Four more cells failed on harness timing, a first-visit prompt, or a
read receipt that ignored a dedicated tool
> - This pull request fixes those four harness defects and adds a cheap
guard that names stale SDK pins
> - The benefit is that these cells now report on the product and the
model again

## Linked Issues or Issue Description

Refs #19614, #20550, #20453, #20417, #20483

**What happened**

On master `f4337bb5fb`, these chat eval cells failed for harness
reasons:

- `plugins.configure-plugin`, `plugins.disable-plugin`,
`plugins.install-plugin`, `plugins.use-plugin-tool`,
`routines.plugin-managed-routine` and `secrets.least-scope-delivery`
stopped before the server started. The source SDK privacy scan refused,
because `tests/runner-e2e/source-sdk-public-inputs.json` did not match
`packages/shared/src`. #20550 and #20362 already resynced the pins on
master. The pins are current at `ce1ee9e89f`.
- `agent-chat.track-delegated-work` and
`navigation-preferences.edit-profile`: the reply was correct and had the
right link. The case read the rendered links immediately after the runs
settled, before the reply rendered, so it recorded no links.
- `inbox-search.archive-done` and `inbox-search.restore-archive`: setup
completes five tasks, so the GitHub star prompt from #20453 opened over
the chat and blocked the composer.
- `status-cards.explain-card`: the follow-up run read the deleted source
with `get_task` and received a 404. The reply said so. `apiReadStatuses`
counted only `call_api` reads, so the grader recorded no unavailable
read.

Why the existing pin check did not stop the drift: `ci / Runner support
suite` runs the source SDK tests and is part of the required `ci /
verify`. On #19614 it passed at 03:43Z on a merge ref from before #20417
(06:12Z) and #20483 (08:14Z). #19614 merged at 14:02Z without a new CI
run on current master. After that, the suite failed on every PR until
#20550. So the check exists and fires; the gap was a stale green result,
and the failure message was opaque (`SDK public source proof
unavailable`).

**Expected behavior**

- A change to a pinned SDK input without a pin refresh fails with the
changed files and the command that fixes it.
- Cases read rendered links only after the saved reply renders.
- The eval browser is a returning user's browser that has already seen
the star prompt. The product does not change.
- A dedicated task read tool (`get_task`, `list_task_documents`,
`read_task_document`, also when namespaced) counts as a read of the
record it names, with the status its receipt reports.

**Steps to reproduce**

1. Run `chat-features.runner-opencode.local.<case>` for one of the cases
above on master `f4337bb5fb`.
2. Read `result.json` and `snapshots/chat-catalog-observation.json` in
the evidence bundle.
3. See the failures described above.

## What Changed

- `tests/runner-e2e/source-sdk-public-inputs.mjs`: add
`staleSdkPublicInputs(root)`. It compares the registry with `git ls-tree
HEAD` for every pinned input (the roots and build inputs that the pins
file lists), and returns each changed, unpinned or removed file.
- `tests/runner-e2e/source-sdk-public-inputs.test.mjs`: a test fails
when the committed pins do not match the committed tree. The message
names the files and gives the `--write` command. The Runner support
suite runs it for every `packages/**` change. A synthetic test covers
changed, unpinned and removed inputs.
- `tests/runner-e2e/chat-catalog-domains/runtime-handoffs.ts` and
`identity-profile.ts`: call the existing `waitForRenderedReply` before
they read rendered links.
- `tests/runner-e2e/chat-returning-browser.ts` (new) and
`chat-catalog-flow.ts`: inside the flow's failure boundary, before
setup, mark `paperclip.github-star-prompt.seen` in the eval browser. An
init script covers later documents of the app origin, and the open page
is marked directly. A test checks that the key equals
`STAR_PROMPT_SEEN_KEY` in `ui/src/components/GitHubStarPrompt.tsx`.
- `tests/runner-e2e/chat-catalog-domains/api-read-receipts.ts`: bind
dedicated task read tools by the digest of the `call_api` input that the
server builds for them (`taskBridgeApiRequest`). The result must carry
the same `apiOperationId`. Other tools, other records, mismatched
receipts and unpaired results do not count.
- Bump `oracleVersion` in code and descriptor for every case that uses
`api-read-receipts.ts` (shared grader change) and for
`navigation-preferences.edit-profile`. Add
`internal-comment-reference.ts` to the source files of the two cases
that now import it.

## Verification

- `pnpm exec vitest run --config tests/runner-e2e/vitest.config.ts`
(whole config): 11301 passed; 1 local-only failure in
`controller-replacement-smoke.test.ts` (server spawn, no changed
imports).
- `pnpm exec vitest run --config tests/runner-e2e/vitest.config.ts
--testTimeout=180000 tests/runner-e2e/chat-catalog-domains
tests/runner-e2e/chat-catalog-contract.test.ts
tests/runner-e2e/chat-catalog-consistency.test.ts
tests/runner-e2e/chat-returning-browser.test.ts`
- `pnpm exec vitest run --config tests/runner-e2e/vitest.config.ts
source-sdk`
- `node tests/runner-e2e/source-sdk-public-inputs.mjs $(git rev-parse
HEAD) --check` reports `changed: false`.
- Red checks: with the `f4337bb5fb` registry, the new pin test fails and
names 15 inputs. Without the new waits, the two rendered-reply tests
fail. A regrade of the retained `explain-card` run events now gives the
follow-up run `unavailable: true, live: false` and the first run `live:
true`.
- Paid runs of the affected cells
(`chat-features.runner-opencode.local.*`):
- Run 42852698879 (head `d54c04d5f0`): `track-delegated-work`,
`restore-archive`, `explain-card`, `configure-plugin` and
`plugin-managed-routine` pass. I read each reply and saved state.
`track-delegated-work` now records the rendered `/COM/issues/COM-2`
link. `explain-card` records `laterUnavailable: true, laterLive: false`
from the follow-up `get_task` 404.
- `plugins.install-plugin` now starts its server (the pin refusal is
gone). It fails a later grading check, `truthful-installation-handoff`:
the reply sends the person to the plugin settings page but does not name
the package. This is outside this PR.
- `inbox-search.archive-done`: the star prompt no longer blocks the
composer, and the query is sent. The agent then archives the 501 Done
tasks one at a time (about 2 s each, 57 in the 240 s window), because
only the per-task archive route exists. This is outside this PR.
- `navigation-preferences.edit-profile` did not start in that run (the
AWS credentials step timed out). Run 42857020103 (head `2f0b80aa84`):
pass. The agent write gets 401, the reply links
`/COM/company/settings/instance/profile`, and the person follows the
rendered link and saves the name.

## Risks

- Low risk. Test harness only; no product code changes.
- The receipt change counts more reads as record reads. It only adds
dedicated reads bound to the exact record id and operation, so a grader
that needs a read can pass on a genuine `get_task` read. The affected
cases have new oracle versions.
- The star prompt is not tested by any chat catalog case. Product e2e
tests for the prompt do not use this flow.

## Model Used

- Anthropic Claude Opus 5.5 (`claude-opus-5-5`), via Claude Code with
tool use (shell, file editing, test execution).

## Checklist

- [x] I have included a thinking path that traces from project context
to this change
- [x] I have specified the model used (with version and capability
details)
- [x] I have checked ROADMAP.md and confirmed this PR does not duplicate
planned core work
- [x] I have searched GitHub for duplicate or related PRs and linked
them above
- [x] I have either (a) linked existing issues with `Fixes: #` / `Closes
#` / `Refs #` OR (b) described the issue in-PR following the relevant
issue template
- [x] I have not referenced internal/instance-local Paperclip issues or
links (only public GitHub `#NNN` / `github.com/paperclipai/paperclip`
URLs)
- [x] My branch name describes the change (e.g. `docs/...`, `fix/...`)
and contains no internal Paperclip ticket id or instance-derived details
- [x] I have run tests locally and they pass
- [x] I have added or updated tests where applicable
- [x] I have updated relevant documentation to reflect my changes
- [x] I have considered and documented any risks above
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-10 11:01:44 -06:00
f4337bb5fb test(evals): land the chat catalog eval harness on master (#19614)
## Thinking Path

> - Paperclip is the open source app people use to manage AI agents for
work
> - The chat-feature evals measure whether agents complete real chat
requests through the product
> - The chat catalog eval harness and its repairs live on the
`chat-eval/master-candidate` branch, not on master
> - Today's work made 139 of 242 chat evals pass in verified single-case
live runs. Each product fix landed on master as its own pull request
> - The eval-side changes (setups, graders, fault seams, evidence
capture) still exist only on the candidate
> - This pull request lands the harness on master. Master stays the
source of truth for product behavior
> - The benefit is one tree: evals run against master commits, and
product fixes reach the evals as soon as they merge

## Linked Issues or Issue Description

Refs #19441, Refs #19599, Refs #20270, Refs #20334, Refs #20342, Refs
#20344, Refs #20354, Refs #20360, Refs #20373, Refs #20382, Refs #20409

**What existing behavior does this improve?**
The chat catalog eval harness and the product it evaluates live on a
long-running branch outside master. Eval fixes do not reach master, so
master-based eval runs use old setups and graders.

**Subsystem affected**
`tests/runner-e2e` chat catalog harness, plus a small set of product
changes that exist only on the candidate.

**Proposed change**
Land the harness on master. Every product fix from this eval work
already merged to master as its own pull request. Conflicts keep
master's product behavior.

## What Changed

- Most of the diff (808 of 851 files) is under `tests/runner-e2e/`:
chat-catalog domains, scenarios, descriptors, the TestOps manifest,
private evidence tooling, and their tests.
- Graders now check saved state, setup literals and rendered or clicked
links. Many keyword and regex checks on free-form replies are removed.
Each changed grader has a bumped `oracleVersion` and calibration tests
(a valid reply, a reworded valid reply, and broken variants that must
fail).
- The shared link helper `hasLink` no longer treats a list link (for
example `/COM/issues?view=mine`) as a link to every task without a URL
key.
- Setup fixes build real state through supported product routes: routine
history, agent grants through the board grant route, budget holds with a
real cap, and a storage-layer archive fault for
`inbox-search.archive-partial-failure`.
- Browser boundary journeys (sign-in, chat-disabled, onboarding) now
write the standard pass evidence files.
- Retargeted cases, approved by the maintainer:
`agent-chat.chat-disabled` (Conference Room gate),
`agents.restore-instructions` (no saved instruction history),
`access.manage-membership` (member permission edits stay board-only).
- The private evidence trusted files (`manual-private-handoff.ts`,
`manual-private-retention.ts`, `manual-private-report-assets.ts`) are
byte-identical to master.
- Product files outside `tests/` (43 files, +1,328 −49) carry
candidate-only changes that master does not yet have. Most are tests.
The main code changes are in `request-human-input.ts`,
`heartbeat-run-summary.ts`, `native-completion-feedback.ts`,
`native-run-finalizer.ts`, `openapi.ts`, `secrets.ts` and
`IssueChatLongThreadPerf.tsx`.

## Verification

- `pnpm vitest run chat-catalog-domains chat-catalog-contract.test.ts
chat-catalog-consistency.test.ts unit-launcher.test.mjs
--testTimeout=180000` in `tests/runner-e2e`: 269 files pass, 3,550 tests
pass.
- `node --test chat-full-catalog-testops-manifest.test.mjs`: 10 pass.
- `node tests/runner-e2e/source-sdk-public-inputs.mjs <HEAD> --check`:
no change.
- `tsc` for `server` and `tests/runner-e2e`: clean.
- Live single-case verification: 139 of 242 cases passed on paid runs,
and a person read the saved evidence for each. These are single-case
runs, not one full-cohort run.

## Risks

- Large diff. Most of it is test-only, but the 43 non-test files change
product code paths. Review those files first.
- Some graders are now less strict on reply wording. A person must read
replies for meaning on those cases.
- 42 cases need disposable outside accounts (Notion, Slack, GitHub,
Daytona and others) and stay blocked. `access.invite-teammate` stays
failing until local agent isolation lands, because the runner's durable
state keeps a live invite token.
- The paid workflow compares the trusted private-evidence files with
master. They are unchanged, so this risk is low.

## Model Used

Anthropic Claude Opus 5.5 (`claude-opus-5-5`), with tool use and code
execution, in Claude Code. Subagents used the same model.

## 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
- [ ] All Paperclip CI gates are green
- [ ] Greptile is 5/5 with no open P2s, recommendations, or follow-ups
- [x] I will address all Greptile and reviewer comments before
requesting merge

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Paperclip <noreply@paperclip.ing>
2026-11-10 08:02:11 -06:00