Skip to content

strat-creator

strat-creator turns approved RFEs into reviewed, feature-ready strategies in the RHAISTRAT Jira project. Where an RFE captures the what and why, a strategy adds the how — technical approach, affected components, impacted teams, dependencies, and non-functional requirements — grounded in the platform's actual architecture.

The plugin implements a two-loop pipeline. The CI loop runs unattended: it clones approved RFEs into RHAISTRAT (strategy-create), refines each into a full strategy (strategy-refine), and subjects it to an adversarial quality gate that combines deterministic rubric scoring with four independent forked reviewers (strategy-review). The gate labels each strategy strat-creator-rubric-pass or strat-creator-needs-attention. The human loop then lets a staff engineer pull a post-CI strategy into a local workspace (strategy-pull), iterate on it with the same refine/review skills in local mode, and either push it back for re-evaluation (strategy-push) or sign it off as feature-ready (strategy-signoff).

Every skill is designed to be safe and idempotent: a shared set of pipeline label gates prevents reprocessing already-handled strategies, a --dry-run mode skips all external Jira writes, and strict section ownership guarantees the RFE business need is copied verbatim and human-authored input is never overwritten. Jira access works through the Atlassian MCP server when available, falling back to a REST API script driven by JIRA_SERVER/JIRA_USER/JIRA_TOKEN environment variables.

Plugin Details

Pipeline

strat-creator pipeline

Dependencies

Skills

Skill Description Invocable
/strategy-create Create strategies from approved RFEs by cloning them to RHAISTRAT in Jira
/strategy-refine Refine a strategy with technical HOW, dependencies, and NFRs
/strategy-review Adversarial review with rubric scoring and independent forked reviewers
/strategy-pull Pull a post-CI strategy from Jira into local workspace for human review
/strategy-push Push a locally-refined strategy back to Jira and resubmit to CI
/strategy-signoff Sign off on a CI-approved strategy with human sign-off label
/export-rubric Export the scoring rubric to artifacts/strat-rubric.md
/strategy-feasibility-review Reviews strategy for technical feasibility and effort estimate credibility internal
/strategy-testability-review Reviews strategy for testability and measurable acceptance criteria internal
/strategy-scope-review Reviews strategy for right-sizing and bounded scope internal
/strategy-architecture-review Reviews strategy for architectural correctness and integration patterns internal

Installation

Claude Code

/plugin install strat-creator@opendatahub-skills

OpenAI Codex — add the marketplace, then enable strat-creator from the /plugins browser:

codex plugin marketplace add opendatahub-io/skills-registry

Architecture

Artifact conventions. All CI-mode skills read and write under artifacts/: strat-tasks/ (strategy files with YAML frontmatter), strat-reviews/ (per-strategy review files), strat-originals/ (frozen RFE/STRAT snapshots), and strat-skipped.md (audit trail of gated-out RFEs). The human loop mirrors this structure under local/ with workflow: local frontmatter. Skills auto-detect local mode by checking for local/strat-tasks/ and prefer it when both exist.

Structured frontmatter. Task and review files carry validated YAML frontmatter (strat_id, title, source_rfe, jira_key, priority, status, reviewers.*, recommendation). Skills never hand-write YAML — they go through scripts/frontmatter.py (schema/read/set). Long-running skills persist progress via scripts/state.py so they survive context compression.

Section ownership. Each strategy file has three top-level sections with strict rules: ## Business Need (from RFE) is copied character-for-character and never modified; ## Strategy (AI Generated by Agentic SDLC Pipeline) is the only section the pipeline writes and is fully regenerated each run; ## Staff Engineer / SME Input is human-owned and read-only for the agent, taking highest priority among refinement inputs.

Pipeline label gates. Gate logic (status checks, required-label checks, already-processed skips) is duplicated across strategy-create, strategy-refine, and strategy-review so each is independently safe to run. Strategies already carrying strat-creator-rubric-pass or strat-creator-needs-attention are skipped by the CI-mode gate; local mode bypasses the gate because the human is deliberately iterating.

Review orchestration. strategy-review is the most complex skill: it bootstraps the assess-strat plugin, spawns a background scorer agent against the rubric, runs deterministic scripts (parse_results.py, apply_scores.py, summarize_run.py) to compute the verdict with no LLM judgment, then invokes the four reviewer skills (strategy-feasibility-review, strategy-testability-review, strategy-scope-review, strategy-architecture-review) in parallel via the Skill tool. Each reviewer runs in an isolated context: fork so no reviewer sees another's output, preserving independent adversarial judgment. Prose reviewer verdicts are informational only — the gate decision comes solely from the numeric rubric scores.

Architecture grounding. Refinement and review fetch opendatahub-io/architecture-context into .context/architecture-context/ (via fetch-architecture-context.sh). Component docs, PLATFORM.md, and human-authored overlays/ (recent corrections that supersede the generated docs) ground technical claims and let reviewers flag outdated assumptions.