Repository navigation
test(eval): collect reproducible trial inputs - #436
Merged
Merged
Conversation
This was referenced Nov 9, 2026
fnzko
reviewed
Nov 9, 2026
fnzko
left a comment
Contributor
There was a problem hiding this comment.
Thanks for this. I've gone through it mainly by testing, not read the code in full detail (deferred that to Claude's). I left inline comments on #438, #439, #441 and #443
Besides those, a few things the RFD already asks for that seem worth doing in this stack, because this runner looks like the base for plugin authors' own evals (Benches like SWE-benches) we discussed a few days ago:
- I love the approach, but the whole thing is done with a simulated client. To leave this in a better position for follow-up, I'd add a real agent run (if the real one has a different transcript, isn't reloading skills/hooks, or behaves differently, nothing ends up making sense).
- Numbering trials by round shall be here, since in the RFD one round is
trial-1-without-symposium+trial-1-with-symposium, but a capability batch now givestrial-1-...andtrial-2-.... Easy to change now, after--trialsit means aschema_versionbump.
gmemuriuki
force-pushed
the
feat/agent-eval-evidence
branch
from
November 9, 2026 16:57
df6cd19 to
4f20634
Compare
Contributor
Author
I don't follow. WDYM? |
Contributor
Author
Would you run it for me, so we can inspect the retained artifacts to confirm the above. If it exposes a real-client difference then, I'll address the concrete failure. |
fnzko
approved these changes
Nov 10, 2026
Read the model, token usage and final message from the transcript, for result.json and response.txt. Usage is unsupported, reported or unknown, so an empty field never hides a reader failure. Sum modelUsage over models, since it covers subagents. Count an error subtype as an error even when is_error is false.
Find the SKILL.md behind each expected skill as Symposium finds a registry plugin's: a skill it would load, named by its frontmatter, one origin per name. Reject a needed skill no plugin provides, any skill two origins provide, and links a trial cannot copy. Read each plugin manifest once for hooks and skills, and conclude nothing when a plugin directory is missing. Co-authored-by: Codex <codex@openai.com>
Copy the task, its fixtures, and its plugins into the batch once, and create every trial from that snapshot, which create_trial now requires and checks is its own batch's. The content hash is the snapshot's digest, and the staged cargo-agents copy is hashed too. A tree's digest covers sorted relative paths, entry kinds, and file bytes, each length-prefixed; timestamps, permissions, and absolute paths are left out, and links refused.
gmemuriuki
force-pushed
the
feat/agent-eval-evidence
branch
from
November 16, 2026 11:31
4f20634 to
4dc5878
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do?
The evaluator needs to record exactly what a trial ran before it can produce a useful result.
This reads Claude's reported model, version, usage, and final response, resolves expected skills back to their sources, and records stable digests for the task, plugins, and staged tools. The readers are kept independent from Symposium's production readers so that the evaluator
does not verify the product using the same implementation.
The tests cover recorded Claude transcripts, malformed and ambiguous skill sources, and the digest encoding.
Stack: #436 → #437 → #438 → #439 → #441 → #442 → #443
Disclosure questions
AI disclosure.
Questions for reviewers.
Are these the right inputs to record for a reproducible trial?