fix(v1): enforce trainable standing against the trainable model context

[hallerite]

What

Env.run_episode now resolves trainable standing right after setup() and before any agent runs. One model context is trainable — the run's ctx — and every trainable agent must point to it (the same model, no pinned client). The run's client decides what divergence means:

• Training run (ctx.client is a TrainClient — its tokens exist to be trained on): a divergent trainable agent is a config error; the episode fails loudly.
• Any other run (eval): the agent is demoted to untrainable, with a warning fired once per agent — an eval must not fail over a training-only concern, but standing filters metrics and records, so it must stay honest.

Either way the message names the one-line fix: agents.<name>.trainable = False in the env's setup().

Why

AgentInfo.trainable means "these tokens train the trainable model". An agent whose model context points anywhere else makes that a false statement, but nothing corrected it: its traces shipped stamped trainable: true, and consumers degraded silently — in prime-rl the trace enters GRPO's group baseline while shipping zero renderer tokens (baseline skew with no error), eval metric filtering counts judge traces as policy traces, and saved records marked trainable: true from another model would poison later SFT-from-traces.

Design notes:

• The client type is the run's purpose, not a new knob: trainable tokens only exist when sampled through a TrainClient (that's what carries renderer tokens), so "refuse when training, demote when measuring" needs no wire or config surface.
• After setup() — the only site that declares trainable (AgentConfig deliberately has no such field), so standing is final; before run() — no tokens burn on a doomed training episode, and eval traces are stamped with corrected standing from the first token.
• Compared on the resolved config: an agent explicitly pinning the run's own model string stays trainable; client is checked as "pinned at all", since _episode_agents hands the run's client to every unpinned agent.

Companion prime-rl PR: PrimeIntellect-ai/prime-rl#3108 (gates group-relative advantages to a single trainable agent — the consumer half of the invariant).

Validation

Throwaway probe over the duet-v1 fixture (no model calls; standing reported by a sentinel raised at the top of run()), both run types:

• eval ctx: unpinned agent stays trainable; model/client pins demote before run() (one warning across 3 episodes); an env that already declares the agent untrainable is untouched, no warning
• train ctx: the same pins fail the episode with ValueError: trainable agent 'a' doesn't use the trainable model ('run-model') …; unpinned and declared-untrainable envs reach run()

(ruff check / format on verifiers/v1/env.py: the import-sort and reformat complaints are pre-existing on main with the locked ruff 0.15.21; this diff's lines are clean and no unrelated churn is included.)

🤖 Generated with Claude Code

[macroscopeapp]

Approvability

Verdict: Needs human review

Unable to check for correctness in e605132. This PR changes runtime training behavior by auto-demoting mismatched trainable agents. An unresolved high-severity comment questions whether this approach is correct, suggesting it could cause silent training failures instead of explicit errors.

^{You can customize Macroscope's approvability policy. Learn more.}

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit e605132ac5583fd3037f9148df90bac389cab97f. Configure here.}

[cursor]

Demotes instead of refusing

High Severity

The PR is meant to refuse a mismatched trainable agent with a ValueError on the episode, but _resolve_trainable_standing auto-demotes it and only logs a warning. Training runs then stay alive with zero trainable traces—the silent failure the design notes explicitly reject.

Additional Locations (1)

• verifiers/v1/env.py#L434-L435

 

^{Reviewed by Cursor Bugbot for commit e605132ac5583fd3037f9148df90bac389cab97f. Configure here.}