| name | cosim-gpu-flow-plan |
| description | Use as the first step for non-trivial cosim testing, debugging, review, or iterative implementation tasks. Do not use for ordinary repository maintenance such as commit splitting, submodule pointer updates, author config, or documentation-only cleanup. |
Cosim Flow Plan
Use as the first step for non-trivial cosim tasks: testing, bug localization,
code review, or iterative implementation. It creates a standardized task
workspace and hands off to the domain skill.
Do not use this skill for ordinary repository maintenance: commit splitting,
submodule pointer updates, history rewriting requested only for commit hygiene,
author configuration, .gitignore changes, documentation-only cleanup, or
skill-directory refactoring that does not require cosim execution evidence.
Hard policy
- For cosim test, debug, review, and RLCR tasks, do not modify source files
before the plan is written and the evidence ledger has confirmed the target.
- Review tasks require explicit human confirmation of the target before any
reviewer delegation, test execution, or source edit. The confirmed target must
name the reviewed change, base reference, file or subsystem scope, allowed
source changes, required evidence, and stopping rule if one is supplied.
- All agent-created artifacts stay under
artifacts/<task-slug>/.
- Do not add AI roles such as Codex or Claude to contribution trailers. When a commit needs
Signed-off-by, use the human author identity from git config user.name and git config user.email.
- Formatting must be diff-scoped. Whole-file or project-wide formatter output
mixed with behavior changes is a hard failure unless the confirmed task is
explicitly formatting-only. Prefer
clang-format-diff, gem5's modified-range
style tools, or an equivalent patch-range formatter; otherwise format only
the changed hunk and its immediate context. Before commit, compare normal
diff stats with whitespace-ignored diff stats and remove unrelated formatter
churn.
- Optimization edits must be based on current
HEAD. A historical commit may
be the reviewed change, but source changes are applied to checked-out HEAD
after recording git rev-parse HEAD and whether the reviewed commit is an
ancestor of HEAD. If later commits already changed the same area, classify
findings against current HEAD before editing.
Workspace
artifacts/<task-slug>/
├── plan.md
├── evidence.md
├── commands.md
├── patch/
│ ├── base-commits.txt # repo path, branch, hash, and commit message
│ └── binary-provenance.txt
├── logs/
├── state/
├── tests/ # per-run test artifacts
├── reviews/ # independent review briefs and outputs
├── matrix.tsv # test tasks only
├── review-findings.md # review tasks only
├── rlcr/ # iterative implementation/debug/validation tasks only
└── scratch/
Task types
Choose one template based on the task.
For authorization boundaries, goal status, artifact relocation, or checkpoint
continuation, read references/policy-and-checkpoint.md.
Test (--type test)
Goal: execute programs against cosim and classify outcomes.
Template fields:
- program list, source repo/path, build target
- interrupt mode (
HSA_ENABLE_INTERRUPT=0 or 1)
- execution mode (
single, repeated, multi)
- timeout strategy (two-phase: probe → precise)
- expected baseline (reference working case)
Output: matrix.tsv with per-program outcome rows.
Hand off to: cosim-gpu-test.
Debug (--type bug)
Goal: prepare a workspace for log-driven debug investigation.
Template fields:
- initial symptom and outcome class
- failing program or workload
- expected nearest working baseline, if known
- evidence paths already available
- allowed investigation scope
Hand off to: cosim-gpu-debug.
Review (--type review)
Goal: review local changes before pushing through the independent reviewer
workflow.
Template fields:
- commit or PR reference
- current
HEAD used for any optimization edits
- whether the reviewed commit is an ancestor of
HEAD
- review scope (subsystem, files)
- explicit user-confirmed review target
- requested round count, if specified
- stopping rule, if specified
- required evidence gates
- allowed automated tests and fixed environment rows
Do not classify review tasks into quick/full modes. cosim-gpu-review owns the
independent review workflow; the user controls only the round count and stop
condition when they specify them.
Hand off to: cosim-gpu-review.
RLCR (--type rlcr)
Goal: run iterative implementation, debugging, or validation rounds.
Template fields:
- round objective and acceptance criteria
- owning domain or operational skill
- verification gates
- independent review path
- stop condition
Hand off to: cosim-gpu-rlcr-loop.
Plan Template
Read references/plan-template.md when creating plan.md.
Composition
After cosim-gpu-flow-plan, invoke the domain skill:
- Test task →
cosim-gpu-test
- Bug task →
cosim-gpu-debug
- Review task →
cosim-gpu-review
- RLCR task →
cosim-gpu-rlcr-loop
Domain skills read plan.md from the active workspace. Use cosim-gpu-build for
all build and provenance decisions.
