[napetrov]

Summary

 def _task_prompt(task: TaskSpec) -> str:
-    return (
-        "Solve this coding task. Work only inside the task environment. "
-        "When finished, leave the files ready for the verifier.\n\n"
-        f"# Task: {task.name}\n\n{task.instruction}"
-    )
+    return task.instruction

{prompt}, {prompt_quoted} and {prompt_escaped} in a command arm's template now expand to the bare instruction.md. The wrapper is text a user's agent never sees, so it can move a command arm's score for no reason the task measures.

Evidence

• Before: a command arm's argv carries "Solve this coding task…" and "# Task: " around the instruction.
• After: test_command_prompt_is_the_raw_instruction asserts argv == ["echo", "Fix the off-by-one in sum()."]. tests/test_task_harnesses.py and tests/test_payload_audit.py: 69 passed. ci-local static and typecheck: 0 failed.

Merge Danger

Door: two-way. The prompt text is a template input; reverting restores the wrapper.

Command-arm runs measured before this change are not comparable in their prompt text (noted in the changelog fragment).

Blast Radius: command and Harbor-style arms only (_task_prompt has one caller). docker-claude arms are not affected.

[Copilot]

🟢 Approval recommended

No active review issues block the change to pass task instructions directly to command and Harbor-style harnesses.

No open issues 🎉

What changed in this PR

0 new findings; nothing to fix.

This PR removes dokimi’s prompt wrapper from command and Harbor-style harnesses so agents receive only the task instruction. Docker-Claude arms are unchanged.

Changes:

• Makes _task_prompt return task.instruction directly.
• Adds a regression test verifying the exact instruction reaches command arguments.
• Documents the prompt change and its effect on comparisons with earlier runs.

File	Description
dokimi/​tests/​test_task_harnesses.py	Tests that command arguments contain the unwrapped instruction.
dokimi/​dokimi/​harnesses/​command.py	Removes the introductory wrapper and task heading.
dokimi/​changelog.d/​20261104_120000_napetrov_raw-task-prompt.md	Records the behavior change and historical prompt-comparability caveat.

_{Review effort: Balanced}

[exai-bot]

✅ OpenCodeReview: Review complete: 0 finding(s) across 3 selected item(s).

CI at `f36cef8`: 3 still running, 11 passing, 2 skipped

Check	Status
CLI interface diff	still running
CodeQL (actions)	still running
CodeQL (python)	still running

Review details

• Commit reviewed: f36cef8 against merge base 272d0b2
• Coverage: 3 files in 1 groups
• Engine: OpenCodeReview v1.13.5-exai.3.d7112371, us.openai.gpt-6.1-sol
• Usage: 41k tokens (26k from cache), 5s
• Tool calls: 2 (2 file_read)
• small change set: 3 files

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 94.76%. Comparing base (8d54c93) to head (f36cef8).
⚠️ Report is 2 commits behind head on main.
❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@           Coverage Diff            @@
##             main     #401    +/-   ##
========================================
  Coverage   94.75%   94.76%            
========================================
  Files         352      352            
  Lines       46507    46407   -100     
  Branches     7970     7949    -21     
========================================
- Hits        44070    43976    -94     
+ Misses       1544     1541     -3     
+ Partials      893      890     -3     

Files with missing lines	Coverage Δ
dokimi/dokimi/harnesses/command.py	97.10% <100.00%> (ø)

... and 2 files with indirect coverage changes

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[napetrov]

No change. This review has 0 findings, so there is no claim about the code to answer. The round only merges main (8d54c93) into the branch.

No change. This report has 0 findings across 3 files, so there is no claim to answer.

No change. This is the Codecov report: modified lines are fully covered. The "behind main" note is informational, not a code claim.

[antonwolfy]

LGTM. _task_prompt returns the bare instruction, it has one caller (build_command), and the new test fails against the old wrapper. Two nits inline.

What I ran, on a Linux review host at c4e06c9:

• make ci-local with Docker: 35 passed, 0 failed, 7 skipped (jobs that do not run on a PR, or need tools absent here). The changelog step passes.
• test_command_prompt.py, test_task_harnesses.py and test_payload_audit.py: 69 passed.
• A mutation run of the new test against 9 variants of _task_prompt and the prompt_* context.

Checked, no issues

• Wrapper gone: a grep for "Solve this coding task", "Work only inside" and "# Task:" finds only the two changelog entries and the new docstring.
• Other arms: docker-claude builds its prompt in docker_solver.py from the notice plus task.instruction, so it is unaffected.
• Commits: both non-merge commits carry Signed-off-by.

[napetrov]

Both nits are fixed in dba386c and cdcb9c8 above; the rest is not a claim.

[vlad-perevezentsev]

LGTM

I checked f36cef8. A command arm now gets instruction.md byte for byte, the same text a docker-claude arm gets without its skill notice, and the tests pin it.

Merge: safe now: it merges cleanly into current main, 0 commits behind, and no other open PR touches its files.

What I ran, at f36cef8:

• make ci-local --offline: 19 passed, 0 failed, 13 skipped (jobs that need the network or Docker, or do not run on a PR).
• tests/test_command_prompt.py on the merge-base 272d0b2: both tests fail against the old wrapper.
• Mutations of _task_prompt and the prompt context: .strip(), splitlines()[0], prompt_escaped = prompt and prompt = "" each fail the multi-line test.
• The Docker smoke (docker-oracle and docker-probe on the pinned catalog task): 2/2 pass, every isolation probe closed.

Checked, no issues

• Callers: _task_prompt has one caller, build_command; Harbor-backed arms pass {task_path} and read instruction.md themselves, and the repo profiles (repos/profiles/*.yaml) build their prompt elsewhere, so neither changes.
• Earlier nits: the changelog is scoped to command arms and says {prompt_escaped} doubles braces (dba386c); the test covers all three placeholders on a multi-line instruction with a trailing newline (cdcb9c8).
• Changelog: Changed is the right category for a prompt change that moves scores, and the entry says old command-arm runs are not comparable.