From 7c662a0b02b694e09afac8711980477e5e918fd1 Mon Sep 17 00:00:00 2001 From: David Bednarczyk Date: Sun, 11 Oct 2026 22:49:19 -0400 Subject: [PATCH 1/2] Evals: unwrapped reward details and the orchestrator's nit guidance --- evals/tasks/vo-t01/tests/finish.py | 7 +++++-- evals/tasks/vo-t02/tests/finish.py | 7 +++++-- evals/tasks/vo-t03/tests/finish.py | 7 +++++-- evals/tasks/vo-t04/tests/finish.py | 7 +++++-- evals/tasks/vo-t05/tests/finish.py | 7 +++++-- evals/tasks/vo-t06/tests/finish.py | 7 +++++-- .../vo-t01/environment/skills/review-orchestrator/SKILL.md | 2 +- evals/team-tasks/vo-t01/steps/validator/tests/finish.py | 7 +++++-- .../vo-t02/environment/skills/review-orchestrator/SKILL.md | 2 +- evals/team-tasks/vo-t02/steps/validator/tests/finish.py | 7 +++++-- .../vo-t03/environment/skills/review-orchestrator/SKILL.md | 2 +- evals/team-tasks/vo-t03/steps/validator/tests/finish.py | 7 +++++-- .../vo-t04/environment/skills/review-orchestrator/SKILL.md | 2 +- evals/team-tasks/vo-t04/steps/validator/tests/finish.py | 7 +++++-- .../vo-t05/environment/skills/review-orchestrator/SKILL.md | 2 +- evals/team-tasks/vo-t05/steps/validator/tests/finish.py | 7 +++++-- .../vo-t06/environment/skills/review-orchestrator/SKILL.md | 2 +- evals/team-tasks/vo-t06/steps/validator/tests/finish.py | 7 +++++-- src/plugin/skills/review-orchestrator/SKILL.md | 2 +- 19 files changed, 67 insertions(+), 31 deletions(-) diff --git a/evals/tasks/vo-t01/tests/finish.py b/evals/tasks/vo-t01/tests/finish.py index c18c553..d81ba92 100644 --- a/evals/tasks/vo-t01/tests/finish.py +++ b/evals/tasks/vo-t01/tests/finish.py @@ -69,7 +69,9 @@ def finish( if not findings: reward = {"reward": 0.0, "recall": 0.0, "format": 0.0 if findings is None else 1.0, "findings": 0.0} # fmt: skip - _write(out, reward, {"judged": False, "reason": reason or "empty findings list"}) + # No judge ran: an empty criteria list marks the zero as unjudged, with the reason. + unjudged = {"score": 0.0, "criteria": [], "note": reason or "empty findings list"} + _write(out, reward, {DIMENSION: unjudged}) return 0 # Clear only the judge's workspace: the work folder may be the process's working directory, # and deleting it leaves imports that call os.getcwd() (LiteLLM's) to fail. @@ -93,7 +95,8 @@ def finish( return 1 recall = float(scores[DIMENSION]) reward = {"reward": recall, "recall": recall, "format": 1.0, "findings": float(len(findings))} - _write(out, reward, {"judged": True, "rubric": details}) + # Reward Kit's own document, unwrapped: graders and the platform read it as written. + _write(out, reward, details) return 0 diff --git a/evals/tasks/vo-t02/tests/finish.py b/evals/tasks/vo-t02/tests/finish.py index c18c553..d81ba92 100644 --- a/evals/tasks/vo-t02/tests/finish.py +++ b/evals/tasks/vo-t02/tests/finish.py @@ -69,7 +69,9 @@ def finish( if not findings: reward = {"reward": 0.0, "recall": 0.0, "format": 0.0 if findings is None else 1.0, "findings": 0.0} # fmt: skip - _write(out, reward, {"judged": False, "reason": reason or "empty findings list"}) + # No judge ran: an empty criteria list marks the zero as unjudged, with the reason. + unjudged = {"score": 0.0, "criteria": [], "note": reason or "empty findings list"} + _write(out, reward, {DIMENSION: unjudged}) return 0 # Clear only the judge's workspace: the work folder may be the process's working directory, # and deleting it leaves imports that call os.getcwd() (LiteLLM's) to fail. @@ -93,7 +95,8 @@ def finish( return 1 recall = float(scores[DIMENSION]) reward = {"reward": recall, "recall": recall, "format": 1.0, "findings": float(len(findings))} - _write(out, reward, {"judged": True, "rubric": details}) + # Reward Kit's own document, unwrapped: graders and the platform read it as written. + _write(out, reward, details) return 0 diff --git a/evals/tasks/vo-t03/tests/finish.py b/evals/tasks/vo-t03/tests/finish.py index c18c553..d81ba92 100644 --- a/evals/tasks/vo-t03/tests/finish.py +++ b/evals/tasks/vo-t03/tests/finish.py @@ -69,7 +69,9 @@ def finish( if not findings: reward = {"reward": 0.0, "recall": 0.0, "format": 0.0 if findings is None else 1.0, "findings": 0.0} # fmt: skip - _write(out, reward, {"judged": False, "reason": reason or "empty findings list"}) + # No judge ran: an empty criteria list marks the zero as unjudged, with the reason. + unjudged = {"score": 0.0, "criteria": [], "note": reason or "empty findings list"} + _write(out, reward, {DIMENSION: unjudged}) return 0 # Clear only the judge's workspace: the work folder may be the process's working directory, # and deleting it leaves imports that call os.getcwd() (LiteLLM's) to fail. @@ -93,7 +95,8 @@ def finish( return 1 recall = float(scores[DIMENSION]) reward = {"reward": recall, "recall": recall, "format": 1.0, "findings": float(len(findings))} - _write(out, reward, {"judged": True, "rubric": details}) + # Reward Kit's own document, unwrapped: graders and the platform read it as written. + _write(out, reward, details) return 0 diff --git a/evals/tasks/vo-t04/tests/finish.py b/evals/tasks/vo-t04/tests/finish.py index c18c553..d81ba92 100644 --- a/evals/tasks/vo-t04/tests/finish.py +++ b/evals/tasks/vo-t04/tests/finish.py @@ -69,7 +69,9 @@ def finish( if not findings: reward = {"reward": 0.0, "recall": 0.0, "format": 0.0 if findings is None else 1.0, "findings": 0.0} # fmt: skip - _write(out, reward, {"judged": False, "reason": reason or "empty findings list"}) + # No judge ran: an empty criteria list marks the zero as unjudged, with the reason. + unjudged = {"score": 0.0, "criteria": [], "note": reason or "empty findings list"} + _write(out, reward, {DIMENSION: unjudged}) return 0 # Clear only the judge's workspace: the work folder may be the process's working directory, # and deleting it leaves imports that call os.getcwd() (LiteLLM's) to fail. @@ -93,7 +95,8 @@ def finish( return 1 recall = float(scores[DIMENSION]) reward = {"reward": recall, "recall": recall, "format": 1.0, "findings": float(len(findings))} - _write(out, reward, {"judged": True, "rubric": details}) + # Reward Kit's own document, unwrapped: graders and the platform read it as written. + _write(out, reward, details) return 0 diff --git a/evals/tasks/vo-t05/tests/finish.py b/evals/tasks/vo-t05/tests/finish.py index c18c553..d81ba92 100644 --- a/evals/tasks/vo-t05/tests/finish.py +++ b/evals/tasks/vo-t05/tests/finish.py @@ -69,7 +69,9 @@ def finish( if not findings: reward = {"reward": 0.0, "recall": 0.0, "format": 0.0 if findings is None else 1.0, "findings": 0.0} # fmt: skip - _write(out, reward, {"judged": False, "reason": reason or "empty findings list"}) + # No judge ran: an empty criteria list marks the zero as unjudged, with the reason. + unjudged = {"score": 0.0, "criteria": [], "note": reason or "empty findings list"} + _write(out, reward, {DIMENSION: unjudged}) return 0 # Clear only the judge's workspace: the work folder may be the process's working directory, # and deleting it leaves imports that call os.getcwd() (LiteLLM's) to fail. @@ -93,7 +95,8 @@ def finish( return 1 recall = float(scores[DIMENSION]) reward = {"reward": recall, "recall": recall, "format": 1.0, "findings": float(len(findings))} - _write(out, reward, {"judged": True, "rubric": details}) + # Reward Kit's own document, unwrapped: graders and the platform read it as written. + _write(out, reward, details) return 0 diff --git a/evals/tasks/vo-t06/tests/finish.py b/evals/tasks/vo-t06/tests/finish.py index c18c553..d81ba92 100644 --- a/evals/tasks/vo-t06/tests/finish.py +++ b/evals/tasks/vo-t06/tests/finish.py @@ -69,7 +69,9 @@ def finish( if not findings: reward = {"reward": 0.0, "recall": 0.0, "format": 0.0 if findings is None else 1.0, "findings": 0.0} # fmt: skip - _write(out, reward, {"judged": False, "reason": reason or "empty findings list"}) + # No judge ran: an empty criteria list marks the zero as unjudged, with the reason. + unjudged = {"score": 0.0, "criteria": [], "note": reason or "empty findings list"} + _write(out, reward, {DIMENSION: unjudged}) return 0 # Clear only the judge's workspace: the work folder may be the process's working directory, # and deleting it leaves imports that call os.getcwd() (LiteLLM's) to fail. @@ -93,7 +95,8 @@ def finish( return 1 recall = float(scores[DIMENSION]) reward = {"reward": recall, "recall": recall, "format": 1.0, "findings": float(len(findings))} - _write(out, reward, {"judged": True, "rubric": details}) + # Reward Kit's own document, unwrapped: graders and the platform read it as written. + _write(out, reward, details) return 0 diff --git a/evals/team-tasks/vo-t01/environment/skills/review-orchestrator/SKILL.md b/evals/team-tasks/vo-t01/environment/skills/review-orchestrator/SKILL.md index e01511f..5d6e189 100644 --- a/evals/team-tasks/vo-t01/environment/skills/review-orchestrator/SKILL.md +++ b/evals/team-tasks/vo-t01/environment/skills/review-orchestrator/SKILL.md @@ -9,7 +9,7 @@ You run a code review by directing other agents. You do not review code yourself or edit any findings file, and you do not edit the repository. Your hands are limited to starting sub-agents, waiting for them, running three commands and writing one short summary of the change. Merging, dropping and ranking findings is the script's job and checking them is the validator's. You -do neither, and one reviewer finding what another did is not evidence that it is real. +do neither, and one reviewer finding what another did is not evidence that it is real. Keep nits to a minimum This skill is an explicit request for sub-agents and parallel agent work. Start the sub-agents below even if your default instructions say not to spawn them unless asked. Start no others: diff --git a/evals/team-tasks/vo-t01/steps/validator/tests/finish.py b/evals/team-tasks/vo-t01/steps/validator/tests/finish.py index c18c553..d81ba92 100644 --- a/evals/team-tasks/vo-t01/steps/validator/tests/finish.py +++ b/evals/team-tasks/vo-t01/steps/validator/tests/finish.py @@ -69,7 +69,9 @@ def finish( if not findings: reward = {"reward": 0.0, "recall": 0.0, "format": 0.0 if findings is None else 1.0, "findings": 0.0} # fmt: skip - _write(out, reward, {"judged": False, "reason": reason or "empty findings list"}) + # No judge ran: an empty criteria list marks the zero as unjudged, with the reason. + unjudged = {"score": 0.0, "criteria": [], "note": reason or "empty findings list"} + _write(out, reward, {DIMENSION: unjudged}) return 0 # Clear only the judge's workspace: the work folder may be the process's working directory, # and deleting it leaves imports that call os.getcwd() (LiteLLM's) to fail. @@ -93,7 +95,8 @@ def finish( return 1 recall = float(scores[DIMENSION]) reward = {"reward": recall, "recall": recall, "format": 1.0, "findings": float(len(findings))} - _write(out, reward, {"judged": True, "rubric": details}) + # Reward Kit's own document, unwrapped: graders and the platform read it as written. + _write(out, reward, details) return 0 diff --git a/evals/team-tasks/vo-t02/environment/skills/review-orchestrator/SKILL.md b/evals/team-tasks/vo-t02/environment/skills/review-orchestrator/SKILL.md index e01511f..5d6e189 100644 --- a/evals/team-tasks/vo-t02/environment/skills/review-orchestrator/SKILL.md +++ b/evals/team-tasks/vo-t02/environment/skills/review-orchestrator/SKILL.md @@ -9,7 +9,7 @@ You run a code review by directing other agents. You do not review code yourself or edit any findings file, and you do not edit the repository. Your hands are limited to starting sub-agents, waiting for them, running three commands and writing one short summary of the change. Merging, dropping and ranking findings is the script's job and checking them is the validator's. You -do neither, and one reviewer finding what another did is not evidence that it is real. +do neither, and one reviewer finding what another did is not evidence that it is real. Keep nits to a minimum This skill is an explicit request for sub-agents and parallel agent work. Start the sub-agents below even if your default instructions say not to spawn them unless asked. Start no others: diff --git a/evals/team-tasks/vo-t02/steps/validator/tests/finish.py b/evals/team-tasks/vo-t02/steps/validator/tests/finish.py index c18c553..d81ba92 100644 --- a/evals/team-tasks/vo-t02/steps/validator/tests/finish.py +++ b/evals/team-tasks/vo-t02/steps/validator/tests/finish.py @@ -69,7 +69,9 @@ def finish( if not findings: reward = {"reward": 0.0, "recall": 0.0, "format": 0.0 if findings is None else 1.0, "findings": 0.0} # fmt: skip - _write(out, reward, {"judged": False, "reason": reason or "empty findings list"}) + # No judge ran: an empty criteria list marks the zero as unjudged, with the reason. + unjudged = {"score": 0.0, "criteria": [], "note": reason or "empty findings list"} + _write(out, reward, {DIMENSION: unjudged}) return 0 # Clear only the judge's workspace: the work folder may be the process's working directory, # and deleting it leaves imports that call os.getcwd() (LiteLLM's) to fail. @@ -93,7 +95,8 @@ def finish( return 1 recall = float(scores[DIMENSION]) reward = {"reward": recall, "recall": recall, "format": 1.0, "findings": float(len(findings))} - _write(out, reward, {"judged": True, "rubric": details}) + # Reward Kit's own document, unwrapped: graders and the platform read it as written. + _write(out, reward, details) return 0 diff --git a/evals/team-tasks/vo-t03/environment/skills/review-orchestrator/SKILL.md b/evals/team-tasks/vo-t03/environment/skills/review-orchestrator/SKILL.md index e01511f..5d6e189 100644 --- a/evals/team-tasks/vo-t03/environment/skills/review-orchestrator/SKILL.md +++ b/evals/team-tasks/vo-t03/environment/skills/review-orchestrator/SKILL.md @@ -9,7 +9,7 @@ You run a code review by directing other agents. You do not review code yourself or edit any findings file, and you do not edit the repository. Your hands are limited to starting sub-agents, waiting for them, running three commands and writing one short summary of the change. Merging, dropping and ranking findings is the script's job and checking them is the validator's. You -do neither, and one reviewer finding what another did is not evidence that it is real. +do neither, and one reviewer finding what another did is not evidence that it is real. Keep nits to a minimum This skill is an explicit request for sub-agents and parallel agent work. Start the sub-agents below even if your default instructions say not to spawn them unless asked. Start no others: diff --git a/evals/team-tasks/vo-t03/steps/validator/tests/finish.py b/evals/team-tasks/vo-t03/steps/validator/tests/finish.py index c18c553..d81ba92 100644 --- a/evals/team-tasks/vo-t03/steps/validator/tests/finish.py +++ b/evals/team-tasks/vo-t03/steps/validator/tests/finish.py @@ -69,7 +69,9 @@ def finish( if not findings: reward = {"reward": 0.0, "recall": 0.0, "format": 0.0 if findings is None else 1.0, "findings": 0.0} # fmt: skip - _write(out, reward, {"judged": False, "reason": reason or "empty findings list"}) + # No judge ran: an empty criteria list marks the zero as unjudged, with the reason. + unjudged = {"score": 0.0, "criteria": [], "note": reason or "empty findings list"} + _write(out, reward, {DIMENSION: unjudged}) return 0 # Clear only the judge's workspace: the work folder may be the process's working directory, # and deleting it leaves imports that call os.getcwd() (LiteLLM's) to fail. @@ -93,7 +95,8 @@ def finish( return 1 recall = float(scores[DIMENSION]) reward = {"reward": recall, "recall": recall, "format": 1.0, "findings": float(len(findings))} - _write(out, reward, {"judged": True, "rubric": details}) + # Reward Kit's own document, unwrapped: graders and the platform read it as written. + _write(out, reward, details) return 0 diff --git a/evals/team-tasks/vo-t04/environment/skills/review-orchestrator/SKILL.md b/evals/team-tasks/vo-t04/environment/skills/review-orchestrator/SKILL.md index e01511f..5d6e189 100644 --- a/evals/team-tasks/vo-t04/environment/skills/review-orchestrator/SKILL.md +++ b/evals/team-tasks/vo-t04/environment/skills/review-orchestrator/SKILL.md @@ -9,7 +9,7 @@ You run a code review by directing other agents. You do not review code yourself or edit any findings file, and you do not edit the repository. Your hands are limited to starting sub-agents, waiting for them, running three commands and writing one short summary of the change. Merging, dropping and ranking findings is the script's job and checking them is the validator's. You -do neither, and one reviewer finding what another did is not evidence that it is real. +do neither, and one reviewer finding what another did is not evidence that it is real. Keep nits to a minimum This skill is an explicit request for sub-agents and parallel agent work. Start the sub-agents below even if your default instructions say not to spawn them unless asked. Start no others: diff --git a/evals/team-tasks/vo-t04/steps/validator/tests/finish.py b/evals/team-tasks/vo-t04/steps/validator/tests/finish.py index c18c553..d81ba92 100644 --- a/evals/team-tasks/vo-t04/steps/validator/tests/finish.py +++ b/evals/team-tasks/vo-t04/steps/validator/tests/finish.py @@ -69,7 +69,9 @@ def finish( if not findings: reward = {"reward": 0.0, "recall": 0.0, "format": 0.0 if findings is None else 1.0, "findings": 0.0} # fmt: skip - _write(out, reward, {"judged": False, "reason": reason or "empty findings list"}) + # No judge ran: an empty criteria list marks the zero as unjudged, with the reason. + unjudged = {"score": 0.0, "criteria": [], "note": reason or "empty findings list"} + _write(out, reward, {DIMENSION: unjudged}) return 0 # Clear only the judge's workspace: the work folder may be the process's working directory, # and deleting it leaves imports that call os.getcwd() (LiteLLM's) to fail. @@ -93,7 +95,8 @@ def finish( return 1 recall = float(scores[DIMENSION]) reward = {"reward": recall, "recall": recall, "format": 1.0, "findings": float(len(findings))} - _write(out, reward, {"judged": True, "rubric": details}) + # Reward Kit's own document, unwrapped: graders and the platform read it as written. + _write(out, reward, details) return 0 diff --git a/evals/team-tasks/vo-t05/environment/skills/review-orchestrator/SKILL.md b/evals/team-tasks/vo-t05/environment/skills/review-orchestrator/SKILL.md index e01511f..5d6e189 100644 --- a/evals/team-tasks/vo-t05/environment/skills/review-orchestrator/SKILL.md +++ b/evals/team-tasks/vo-t05/environment/skills/review-orchestrator/SKILL.md @@ -9,7 +9,7 @@ You run a code review by directing other agents. You do not review code yourself or edit any findings file, and you do not edit the repository. Your hands are limited to starting sub-agents, waiting for them, running three commands and writing one short summary of the change. Merging, dropping and ranking findings is the script's job and checking them is the validator's. You -do neither, and one reviewer finding what another did is not evidence that it is real. +do neither, and one reviewer finding what another did is not evidence that it is real. Keep nits to a minimum This skill is an explicit request for sub-agents and parallel agent work. Start the sub-agents below even if your default instructions say not to spawn them unless asked. Start no others: diff --git a/evals/team-tasks/vo-t05/steps/validator/tests/finish.py b/evals/team-tasks/vo-t05/steps/validator/tests/finish.py index c18c553..d81ba92 100644 --- a/evals/team-tasks/vo-t05/steps/validator/tests/finish.py +++ b/evals/team-tasks/vo-t05/steps/validator/tests/finish.py @@ -69,7 +69,9 @@ def finish( if not findings: reward = {"reward": 0.0, "recall": 0.0, "format": 0.0 if findings is None else 1.0, "findings": 0.0} # fmt: skip - _write(out, reward, {"judged": False, "reason": reason or "empty findings list"}) + # No judge ran: an empty criteria list marks the zero as unjudged, with the reason. + unjudged = {"score": 0.0, "criteria": [], "note": reason or "empty findings list"} + _write(out, reward, {DIMENSION: unjudged}) return 0 # Clear only the judge's workspace: the work folder may be the process's working directory, # and deleting it leaves imports that call os.getcwd() (LiteLLM's) to fail. @@ -93,7 +95,8 @@ def finish( return 1 recall = float(scores[DIMENSION]) reward = {"reward": recall, "recall": recall, "format": 1.0, "findings": float(len(findings))} - _write(out, reward, {"judged": True, "rubric": details}) + # Reward Kit's own document, unwrapped: graders and the platform read it as written. + _write(out, reward, details) return 0 diff --git a/evals/team-tasks/vo-t06/environment/skills/review-orchestrator/SKILL.md b/evals/team-tasks/vo-t06/environment/skills/review-orchestrator/SKILL.md index e01511f..5d6e189 100644 --- a/evals/team-tasks/vo-t06/environment/skills/review-orchestrator/SKILL.md +++ b/evals/team-tasks/vo-t06/environment/skills/review-orchestrator/SKILL.md @@ -9,7 +9,7 @@ You run a code review by directing other agents. You do not review code yourself or edit any findings file, and you do not edit the repository. Your hands are limited to starting sub-agents, waiting for them, running three commands and writing one short summary of the change. Merging, dropping and ranking findings is the script's job and checking them is the validator's. You -do neither, and one reviewer finding what another did is not evidence that it is real. +do neither, and one reviewer finding what another did is not evidence that it is real. Keep nits to a minimum This skill is an explicit request for sub-agents and parallel agent work. Start the sub-agents below even if your default instructions say not to spawn them unless asked. Start no others: diff --git a/evals/team-tasks/vo-t06/steps/validator/tests/finish.py b/evals/team-tasks/vo-t06/steps/validator/tests/finish.py index c18c553..d81ba92 100644 --- a/evals/team-tasks/vo-t06/steps/validator/tests/finish.py +++ b/evals/team-tasks/vo-t06/steps/validator/tests/finish.py @@ -69,7 +69,9 @@ def finish( if not findings: reward = {"reward": 0.0, "recall": 0.0, "format": 0.0 if findings is None else 1.0, "findings": 0.0} # fmt: skip - _write(out, reward, {"judged": False, "reason": reason or "empty findings list"}) + # No judge ran: an empty criteria list marks the zero as unjudged, with the reason. + unjudged = {"score": 0.0, "criteria": [], "note": reason or "empty findings list"} + _write(out, reward, {DIMENSION: unjudged}) return 0 # Clear only the judge's workspace: the work folder may be the process's working directory, # and deleting it leaves imports that call os.getcwd() (LiteLLM's) to fail. @@ -93,7 +95,8 @@ def finish( return 1 recall = float(scores[DIMENSION]) reward = {"reward": recall, "recall": recall, "format": 1.0, "findings": float(len(findings))} - _write(out, reward, {"judged": True, "rubric": details}) + # Reward Kit's own document, unwrapped: graders and the platform read it as written. + _write(out, reward, details) return 0 diff --git a/src/plugin/skills/review-orchestrator/SKILL.md b/src/plugin/skills/review-orchestrator/SKILL.md index e01511f..5d6e189 100644 --- a/src/plugin/skills/review-orchestrator/SKILL.md +++ b/src/plugin/skills/review-orchestrator/SKILL.md @@ -9,7 +9,7 @@ You run a code review by directing other agents. You do not review code yourself or edit any findings file, and you do not edit the repository. Your hands are limited to starting sub-agents, waiting for them, running three commands and writing one short summary of the change. Merging, dropping and ranking findings is the script's job and checking them is the validator's. You -do neither, and one reviewer finding what another did is not evidence that it is real. +do neither, and one reviewer finding what another did is not evidence that it is real. Keep nits to a minimum This skill is an explicit request for sub-agents and parallel agent work. Start the sub-agents below even if your default instructions say not to spawn them unless asked. Start no others: From e24b757fe5527647025b3c4ae6cedf54aff3f439 Mon Sep 17 00:00:00 2001 From: David Bednarczyk Date: Sun, 11 Oct 2026 23:09:03 -0400 Subject: [PATCH 2/2] Drop the e2e test line from the orchestrator prompt --- .../vo-t01/environment/skills/review-orchestrator/SKILL.md | 2 +- .../vo-t02/environment/skills/review-orchestrator/SKILL.md | 2 +- .../vo-t03/environment/skills/review-orchestrator/SKILL.md | 2 +- .../vo-t04/environment/skills/review-orchestrator/SKILL.md | 2 +- .../vo-t05/environment/skills/review-orchestrator/SKILL.md | 2 +- .../vo-t06/environment/skills/review-orchestrator/SKILL.md | 2 +- src/plugin/skills/review-orchestrator/SKILL.md | 2 +- 7 files changed, 7 insertions(+), 7 deletions(-) diff --git a/evals/team-tasks/vo-t01/environment/skills/review-orchestrator/SKILL.md b/evals/team-tasks/vo-t01/environment/skills/review-orchestrator/SKILL.md index 5d6e189..e01511f 100644 --- a/evals/team-tasks/vo-t01/environment/skills/review-orchestrator/SKILL.md +++ b/evals/team-tasks/vo-t01/environment/skills/review-orchestrator/SKILL.md @@ -9,7 +9,7 @@ You run a code review by directing other agents. You do not review code yourself or edit any findings file, and you do not edit the repository. Your hands are limited to starting sub-agents, waiting for them, running three commands and writing one short summary of the change. Merging, dropping and ranking findings is the script's job and checking them is the validator's. You -do neither, and one reviewer finding what another did is not evidence that it is real. Keep nits to a minimum +do neither, and one reviewer finding what another did is not evidence that it is real. This skill is an explicit request for sub-agents and parallel agent work. Start the sub-agents below even if your default instructions say not to spawn them unless asked. Start no others: diff --git a/evals/team-tasks/vo-t02/environment/skills/review-orchestrator/SKILL.md b/evals/team-tasks/vo-t02/environment/skills/review-orchestrator/SKILL.md index 5d6e189..e01511f 100644 --- a/evals/team-tasks/vo-t02/environment/skills/review-orchestrator/SKILL.md +++ b/evals/team-tasks/vo-t02/environment/skills/review-orchestrator/SKILL.md @@ -9,7 +9,7 @@ You run a code review by directing other agents. You do not review code yourself or edit any findings file, and you do not edit the repository. Your hands are limited to starting sub-agents, waiting for them, running three commands and writing one short summary of the change. Merging, dropping and ranking findings is the script's job and checking them is the validator's. You -do neither, and one reviewer finding what another did is not evidence that it is real. Keep nits to a minimum +do neither, and one reviewer finding what another did is not evidence that it is real. This skill is an explicit request for sub-agents and parallel agent work. Start the sub-agents below even if your default instructions say not to spawn them unless asked. Start no others: diff --git a/evals/team-tasks/vo-t03/environment/skills/review-orchestrator/SKILL.md b/evals/team-tasks/vo-t03/environment/skills/review-orchestrator/SKILL.md index 5d6e189..e01511f 100644 --- a/evals/team-tasks/vo-t03/environment/skills/review-orchestrator/SKILL.md +++ b/evals/team-tasks/vo-t03/environment/skills/review-orchestrator/SKILL.md @@ -9,7 +9,7 @@ You run a code review by directing other agents. You do not review code yourself or edit any findings file, and you do not edit the repository. Your hands are limited to starting sub-agents, waiting for them, running three commands and writing one short summary of the change. Merging, dropping and ranking findings is the script's job and checking them is the validator's. You -do neither, and one reviewer finding what another did is not evidence that it is real. Keep nits to a minimum +do neither, and one reviewer finding what another did is not evidence that it is real. This skill is an explicit request for sub-agents and parallel agent work. Start the sub-agents below even if your default instructions say not to spawn them unless asked. Start no others: diff --git a/evals/team-tasks/vo-t04/environment/skills/review-orchestrator/SKILL.md b/evals/team-tasks/vo-t04/environment/skills/review-orchestrator/SKILL.md index 5d6e189..e01511f 100644 --- a/evals/team-tasks/vo-t04/environment/skills/review-orchestrator/SKILL.md +++ b/evals/team-tasks/vo-t04/environment/skills/review-orchestrator/SKILL.md @@ -9,7 +9,7 @@ You run a code review by directing other agents. You do not review code yourself or edit any findings file, and you do not edit the repository. Your hands are limited to starting sub-agents, waiting for them, running three commands and writing one short summary of the change. Merging, dropping and ranking findings is the script's job and checking them is the validator's. You -do neither, and one reviewer finding what another did is not evidence that it is real. Keep nits to a minimum +do neither, and one reviewer finding what another did is not evidence that it is real. This skill is an explicit request for sub-agents and parallel agent work. Start the sub-agents below even if your default instructions say not to spawn them unless asked. Start no others: diff --git a/evals/team-tasks/vo-t05/environment/skills/review-orchestrator/SKILL.md b/evals/team-tasks/vo-t05/environment/skills/review-orchestrator/SKILL.md index 5d6e189..e01511f 100644 --- a/evals/team-tasks/vo-t05/environment/skills/review-orchestrator/SKILL.md +++ b/evals/team-tasks/vo-t05/environment/skills/review-orchestrator/SKILL.md @@ -9,7 +9,7 @@ You run a code review by directing other agents. You do not review code yourself or edit any findings file, and you do not edit the repository. Your hands are limited to starting sub-agents, waiting for them, running three commands and writing one short summary of the change. Merging, dropping and ranking findings is the script's job and checking them is the validator's. You -do neither, and one reviewer finding what another did is not evidence that it is real. Keep nits to a minimum +do neither, and one reviewer finding what another did is not evidence that it is real. This skill is an explicit request for sub-agents and parallel agent work. Start the sub-agents below even if your default instructions say not to spawn them unless asked. Start no others: diff --git a/evals/team-tasks/vo-t06/environment/skills/review-orchestrator/SKILL.md b/evals/team-tasks/vo-t06/environment/skills/review-orchestrator/SKILL.md index 5d6e189..e01511f 100644 --- a/evals/team-tasks/vo-t06/environment/skills/review-orchestrator/SKILL.md +++ b/evals/team-tasks/vo-t06/environment/skills/review-orchestrator/SKILL.md @@ -9,7 +9,7 @@ You run a code review by directing other agents. You do not review code yourself or edit any findings file, and you do not edit the repository. Your hands are limited to starting sub-agents, waiting for them, running three commands and writing one short summary of the change. Merging, dropping and ranking findings is the script's job and checking them is the validator's. You -do neither, and one reviewer finding what another did is not evidence that it is real. Keep nits to a minimum +do neither, and one reviewer finding what another did is not evidence that it is real. This skill is an explicit request for sub-agents and parallel agent work. Start the sub-agents below even if your default instructions say not to spawn them unless asked. Start no others: diff --git a/src/plugin/skills/review-orchestrator/SKILL.md b/src/plugin/skills/review-orchestrator/SKILL.md index 5d6e189..e01511f 100644 --- a/src/plugin/skills/review-orchestrator/SKILL.md +++ b/src/plugin/skills/review-orchestrator/SKILL.md @@ -9,7 +9,7 @@ You run a code review by directing other agents. You do not review code yourself or edit any findings file, and you do not edit the repository. Your hands are limited to starting sub-agents, waiting for them, running three commands and writing one short summary of the change. Merging, dropping and ranking findings is the script's job and checking them is the validator's. You -do neither, and one reviewer finding what another did is not evidence that it is real. Keep nits to a minimum +do neither, and one reviewer finding what another did is not evidence that it is real. This skill is an explicit request for sub-agents and parallel agent work. Start the sub-agents below even if your default instructions say not to spawn them unless asked. Start no others: