eval(harbor): stop the adapter injecting a deadline into clawcodex's prompt

[ericleepi314]

The Harbor adapter appended Harbor's wall-clock deadline to clawcodex's system prompt via settings.append_system_prompt:

"This run has a hard external execution deadline at <ts>. By <ts>, preserve the best valid deliverable, stop broad exploration, and switch to the narrowest checks needed for the explicit requirements … You can use date -u to compare the current time."

The built-in claude-code adapter appends nothing (claude_code_subscription is a thin subclass touching only credentials), so this was harness-shaped guidance one agent got and the other didn't — and it told the model to narrow its own verification, which is precisely the behavior a trajectory comparison is trying to measure. eval/harbor/README.md already notes official submissions may not modify task timeouts, so it wouldn't exist on a scored run.

It never took effect — but not for the reason you'd guess

I initially wrote that the deadline resolved to None because hub trials don't materialize a task.toml. That was wrong. Re-running the deleted resolver against the recorded lock files and cached task definitions returns a real deadline for every task:

task	resolved
cancel-async-tasks, regex-log, fix-code-vulnerability, sparql-university	900s
large-scale-text-editing	1200s
circuit-fibsqrt, distribution-search	3600s

The adapter did write it. The value was then silently dropped: settings.append_system_prompt is a dead config key — it parses into the settings object (src/settings/types.py:261) and nothing reads it. The headless chain ends at src/entrypoints/headless.py:504-513, which passes only the output-style prompt into build_effective_system_prompt. Verified with a sentinel under a fake $HOME: it loads into settings and never reaches the assembled prompt. (settings.effort is live via src/query/query.py:575-586, which is what made the sibling key look plausible.)

So prior clawcodex-vs-Claude-Code measurements weren't confounded by this — by accident, not design.

Follow-up worth filing: settings.append_system_prompt and settings.custom_system_prompt are silently ignored on both headless and TUI.

Adds eval/harbor/compare_trajectories.py

The comparisons driving recent prompt work were ad-hoc, and several were wrong the same way: clawcodex figures over a 7-task job compared against Claude Code figures over an 89-task job. Restricted to the shared tasks, several conclusions invert:

metric	as reported	actual (shared 7)
exact-repeat rate	5.3% vs 6.8% — CC repeats more	5.3% vs 0.0% — CC repeats nothing
heredoc-authored	12.1% vs 7.4% — clawcodex more	12.1% vs 23.4% — CC ~2× more
Bash statements/call	8.7 vs 9.3 ≈ same	4.4 vs 5.5 — CC packs more
env-probe share	13.3% vs 11.5% ≈ same	13.6% vs 8.5%
mean steps	12.9 vs 7.4	10.67 vs 7.43

The script only computes over the task intersection; reports harness-killed trials separately and excludes them from step means (iter2's large-scale-text-editing died on a TimeoutError — averaging a truncated run against completed ones inflated the gap from +3.2 to +5.5); states that "steps" means steps carrying a tool call (raw len(steps) is 14.7 vs 9.4); and always prints rewards, because all 14 trials on this set scored 1.0 — a step delta where both sides are already perfect is a cost difference, not a quality one.

Tests: 79 passing in the affected area; both adapters import under Harbor's interpreter. No orphan references to time_budget anywhere (src, tests, CI, Makefile, pyproject, or either eval README).

🤖 Generated with Claude Code