mirror of
https://github.com/paperclipai/paperclip.git
synced 2026-11-16 11:59:26 +01:00
## Thinking Path
> - Paperclip is the open source app people use to manage AI agents for
work
> - The configure-company-secret eval checks that a person saves a
deployment key securely and that Morgan receives only a reference to it
> - After the person approves the binding, a separate agent run must
confirm the outcome in chat
> - The check required that confirmation to contain one of a few words
("bound", "available", "configured", "ready", "added")
> - A correct confirmation said the binding was "applied" and "exposes"
the variable, so a correct outcome failed
> - This pull request grades the confirmation from saved state: who
posted it, after what, and which binding exists
> - The benefit is that the eval checks the real outcome, and the
reply's wording is left to human review
## Linked Issues or Issue Description
Refs #21350
**What happened?**
`chat-features.runner-opencode.local.secrets.configure-company-secret`
failed only `human-attributed-and-confirmed-after-authorization`. The
run saved the key, the binding approval was accepted and executed, and
the binding to Morgan exists as `env.DEPLOYMENT_API_KEY`. I evaluated
each part of the check against the retained observation. The secret was
created by the person. The approval path was valid. The reply came from
the bound continuation run and carried the bound interaction. The
proposal-phase check passed. `/Morgan/`, `/deploy/` and the negation
guard passed. Only `/bound|available|configured|ready|added/` failed:
the reply said "the binding you approved is applied" and "exposes it as
the environment variable `DEPLOYMENT_API_KEY`".
**Expected behavior**
The confirmation passes when the bound post-authorization run posted it
after the intended binding existed. Wording is not graded by keywords.
**Steps to reproduce**
Grade the retained confirmation with the old check. It fails although
every structural condition holds.
## What Changed
- In `external-company-secret.ts`,
`human-attributed-and-confirmed-after-authorization` no longer checks
the reply text. It requires:
- the secret was created by the current person;
- the existing authorization proof (an executed approval or native
binding authorization, or the human-attributed binding);
- a later reply whose saved comment is in this company and source
conversation, authored by the lead agent and created by that reply's
run. When a native binding is required, the run must be the bound
post-authorization run (the approval's continuation run, or the bound
continuation). When a question or confirmation card started the flow,
the run must carry that interaction;
- one binding to Morgan for the saved secret, naming the intended
variable (`env.DEPLOYMENT_API_KEY` by default, or the person's requested
name on that existing path);
- the reply comment was created at or after that binding's `createdAt`.
- The observation now always collects the source conversation's saved
comments (before, it did so only on the follow-up and card paths), so
every path can bind the reply to its run and time.
- The truthfulness pattern for the proposal phase and the other wording
patterns are dropped from this check. The credential-in-chat guard
elsewhere in the case is unchanged.
- No scenario YAML changed, so the TestOps manifest needs no entry.
## Verification
- New test built from the retained confirmation text. It passes with the
new grade and fails with the old one, on this exact check. It also fails
when the reply is saved before the binding.
- New rejections in the grade tests: no saved reply comment; a reply
saved before the binding; a reply from the initial run; a reply authored
by another agent; a reply in another conversation; a binding naming
another variable; a binding without a creation time.
- Fixtures now carry the saved reply comment and binding time. Company
and secret tests, plus catalog consistency and contract tests: 197
passed. `pnpm test:e2e:runner:typecheck` passes.
- A narrow paid run of the case on this branch is linked in a comment.
## Risks
- Low risk. The change is in eval code only.
- The confirmation's wording is no longer checked by this grade, by
design. A reply that misstates the outcome would need human review.
- One extra comment read during observation on the direct path.
## Model Used
- Claude Opus 5.5 (Anthropic), model ID `claude-opus-5-5`, used through
Claude Code with tool use and code 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
- [ ] 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>