test: the plugin suites on Haiku 5.5

[mauricios]

Re-runs the plugin-hooks and plugin-docs session evals on Claude Haiku 5.5. Their reports from 2026-10-04 and 2026-10-05 ran on Haiku 4.5. Full write-up: evals/dynamic/reports/2026-10-09-haiku-5-5.md.

What changes

• run-session-evals.sh: agents that ask for opus in plugin-docs now run on claude-haiku-5-5. The stand-in still pinned claude-haiku-4-5-20251001.
• A report of this run, linked from evals/dynamic/README.md.

Results

Run	Sessions	Checks
plugin-hooks	7	8 passed, 0 failed — same as Haiku 4.5
plugin-docs	3	7 passed, 0 failed — same as Haiku 4.5
agent-spec-model without decision 0019's note	1	2 passed, 0 failed

Every transcript names claude-haiku-5-5, including the agents'.

Decision 0019 added its note after spec-analyzer, on Haiku 4.5, read the project's docs/SPEC-MODEL.md instead of the plugin's. Without the note, Haiku 5.5 read the plugin's copy.

The note stays, for three reasons:

• One run doesn't show the failure can't happen.
• The line states a fact: the docs live outside the project.
• In projects the agent runs on Opus.

Caveats

• Claude Code v2.1.286 doesn't recognize claude-haiku-5-5. It prices the model at another rate, so the cost figures are unreliable, and it checks session budget caps against the same estimate. One plugin-docs session stopped at its $0.75 cap before its final reply; its checks had already passed. Update the CLI before reading the costs.
• These suites test plumbing, not judgment. triage and debug stay on sonnet and opus.

Verification

./evals/run-evals.sh passes. The session runs are the ones the report describes.

🤖 Generated with Claude Code