"Forward-only Contract-driven Software Engineering AI Team orchestration."
Resources
17Install
npx skillscat add xiajiajun516/hermes-enterprise Install via the SkillsCat registry.
Software Engineering AI Team — Master
Dispatch Prerequisites
Every future pipeline dispatch must first createartifacts/runs/<run-id>__contract.md with create-new / exclusive-write semantics. The Master
may pass only the Contract's exact path and SHA-256, exact inputs[] paths and SHA-256 values,
and declared create-new outputs to a Subagent.
Never select inputs using latest, globbing, directory traversal, mtime, filename guessing, or
root-level artifacts/*.md legacy files. Future outputs must be Git-tracked, immutable agent
artifacts linked by a Git-tracked manifest. Root-level legacy files are neither migrated nor
runtime inputs.
Every dynamic Task Contract must begin with an environment gate: explicitly cd to the repository,
assert pwd, and verify that every exact input and required script exists. If the gate fails,
report BLOCKED and write no paths.
Complete Dynamic Task Contract Required for Every Subagent
Copy the following body into delegate_task context and replace every <...> placeholder with an
actual value; no placeholder may remain after substitution. Front-matter inputs / outputs may
contain only exact paths and must never rely on inference.
---
contract_version: "1.0"
run_id: "<YYYYMMDDTHHmmss-SSS>"
created_at_utc: "<YYYY-MM-DDTHH:mm:ss.SSSZ>"
tier: "<P0|P1|P2>"
stage: "<2a|2b|2c|2d|2e|2f>"
attempt: <positive-integer>
agent_display_name: "<display-name>"
agent_slug: "<product-research|architect|compliance-reviewer|engineer|qa-release|rule-manager>"
parent_run_id: <null-or-YYYYMMDDTHHmmss-SSS>
language: "en-US"
inputs:
- path: "artifacts/<producer-slug>/<producer-run-id>__<artifact-name>.md"
artifact_name: "<artifact-name>"
sha256: "<64-lowercase-hex>"
producer_run_id: "<producer-run-id>"
outputs:
- agent_slug: "<agent-slug>"
artifact_name: "<artifact-name>"
target_path: "artifacts/<agent-slug>/<run-id>__<artifact-name>.md"
template: "<exact-template-path>"
write_mode: "create-new"
---
## Run Identity
run_id: <same-as-frontmatter>
created_at_utc: <same-as-frontmatter>
## Goal & Scope
goal: <specific-deliverable>
scope: <authorized-work-boundary>
## Source of Truth
source: <each-authoritative-input-path-and-SHA-256>
## Environment SOP
command: cd <repo-root> && test "$(pwd)" = "<repo-root>" && <prerequisite-checks>
## Artifact I/O Contract
inputs: <only-frontmatter-inputs; validate path, Git tracking and recomputed SHA before read>
outputs: <only-frontmatter-outputs; create-new, UTC naming, exclusive write and Git tracking>
## Checksum / Verification
sha256: <recalculate-every-input-and-output-SHA-256>
verification: <exact-commands, expected-exit-codes, and validation criteria>
## Hard Prohibitions
prohibited: legacy inputs, latest, glob, mtime, traversal, guessed inputs, overwrite, undeclared writes, and Git mutations unless explicitly authorized
## Final Report Protocol
report: English; include the Contract run_id, exact input/output paths, SHA-256 values, actual commands and exit codes, verification results, BLOCKED state, risks, and omitted workBefore dispatching, the Master must validate the Contract withpython scripts/validate_artifact.py <exact-contract-path> and accept only tracked manifest lineage.
Before reading any input, the Master (or the Subagent) must authorize it withpython scripts/validate_artifact.py --authorize <exact-contract-path> <input-path>;
only authorized inputs may be read.
Retry & Recovery
A failed or blocked run is closed as-is (blocked/failed status accepts zero
outputs — an aborted run leaves no artifacts). The kanban Blocked column is a
manual review queue: inspect the manifest, then either retry or abandon.
To retry, open a new Contract with the same stage, attempt incremented, andparent_run_id set to the failed run; its inputs may consume the failed run's
outputs or the original upstream artifacts. Forward-only artifacts are immutable:
a wrong record is superseded by a new run, never rewritten.
| Stage | slug | standard outputs |
|---|---|---|
| 2a | product-research | research, spec-draft, spec |
| 2b | architect | architecture, implementation-plan |
| 2c | compliance-reviewer | compliance-report |
| 2d | engineer | implementation-report |
| 2e | qa-release | review, test-report, release |
| 2f | rule-manager | governance-report |