[hahahahah1287]

Problem
OpencodeSubstrate.render_events and CodexSubstrate.render_events return "(no parseable events in trajectory)" or "(failed to read trajectory: ...)" when a trajectory stream is empty, contains no valid JSON events, or fails to read, whereas selection._render_pool and schedule._resume only checked startswith("(no trajectory").

As a result:

On state_deliverable=True benchmarks, _render_pool admitted "(no parseable events in trajectory)" and "(failed to read trajectory: ...)" as valid session-transcript candidates (and trivial_selection at k=1 marked them ok=True).

schedule._resume treated tasks with degraded/unparseable trajectory evidence as intact (skipped) instead of re-running them.

context._write_rollout treated sentinel strings as successful trajectory renders, preventing context.build's "no trajectory could be rendered for ANY rollout" guard from firing when all rollouts lacked parseable trajectories.

Fix

Add context.is_missing_trajectory() covering all missing/unparseable trajectory sentinels.

Use context.is_missing_trajectory() in context._write_rollout, selection._render_pool, and schedule._resume.

Add a regression test in tests/test_workflow.py covering context building, selection pool filtering, and resume validation.

[Praveen2795]

Hey, I ran into this same gap while looking at part 1b of #143, then found your PR, so I'm not opening a duplicate. I pulled the branch (df2ccd0) and tried it on macOS.

The full suite passes for me on Python 3.12.13 and 3.13.16. I also wanted to make sure the new test pins down all three fixes, so I reverted them one at a time. It failed every time:

• schedule.py: ['skipped', 'skipped'] != ['ok', 'skipped']
• selection.py: True is not false
• context.py (only the new failure line in _write_rollout): RuntimeError not raised

One thing to fix: ruff check flags the new import in schedule.py (I001, since pyproject.toml enables the I rules). ruff check --fix just merges it into the existing line:

-from . import context
+from . import context, recursion, runmark, selection
 from . import library as library_store
-from . import recursion, runmark, selection

(ruff format --check also flags core/tree.py, but that file is already unformatted on main, so it's not from this PR.)

And a small optional idea: the test types the sentinel strings out by hand. If someone later rewords a renderer's message, the test keeps passing even though the guard no longer matches. You could get the strings from the renderers instead:

from harness2.substrate import codex, opencode

garbled = self.root / "garbled.jsonl"
garbled.write_text("not json\n", encoding="utf-8")
sentinels = (
    opencode.render_events(garbled, obs_trunc=10_000, max_chars=10_000),
    codex.render_events(self.root / "missing.jsonl", obs_trunc=10_000, max_chars=10_000),
)

Otherwise this looks good to me.

[hahahahah1287]

Thanks @Praveen2795 for testing the branch thoroughly and catching the import sort issue!
I've pushed an update that:

• Merges from . import context, recursion, runmark, selection in harness2/core/schedule.py so ruff check (I001) passes cleanly.

• Derives the missing-trajectory sentinel strings in tests/test_workflow.py directly from opencode.render_events(...) and codex.render_events(...) as suggested, so the regression test stays coupled to the actual substrate renderer outputs.

[Praveen2795]

Thanks for the quick update! Re-tested 62a59ef on macOS:

• Full suite passes on Python 3.12.13 and 3.13.16.
• ruff check . is clean now. ruff format --check only flags core/tree.py and substrate/codex.py, and both are already like that on main.
• I redid the revert check with the new renderer-based sentinels. The test still fails when any of the three fixes is removed.

LGTM 👍