fix(profiles): bind routing receipts and harden review paths

This commit is contained in:
haelyra
2026-09-27 18:01:40 -04:00
parent a9e3ecb77f
commit befde0d2cd
11 changed files with 99 additions and 18 deletions
@@ -136,7 +136,7 @@ node docker/context-profiles/ai-eval.js --plan \
> registration.json
# 5. Run (requires your own Claude subscription login or API key).
node docker/context-profiles/ai-eval.js --allow-real-provider \
node docker/context-profiles/ai-eval.js --allow-real-provider --allow-credentialed-tools \
--registration registration.json \
--corpus docker/context-profiles/complex-corpus.json \
--provider claude --model <model> --executable /absolute/path/to/claude \
@@ -144,6 +144,11 @@ node docker/context-profiles/ai-eval.js --allow-real-provider \
--artifact-dir /absolute/path/for/transcripts > report.json
```
Claude task tools inherit the provider credential through the CLI process and can read it. Use
`--allow-credentialed-tools` only with a trusted local corpus and credential. Without that
explicit flag, real Claude task evaluation stops before a provider call; selection-only calls
remain tool-free. This development evaluator does not provide a credential isolation boundary.
The registration digest binds the exact corpus, evaluator source, model, and
executable; the run refuses to start if any of them drift, and aborts if the
tree changes mid-run. `--artifact-dir` retains per-trial session transcripts