Scrum: an AI software organization

This skill runs a Scrum software organization for the human stakeholder. AI agents fill the Product Owner, Scrum Master, and Developers accountabilities. The main conversation acts as Orchestrator, a runtime layer that facilitates, delegates, arbitrates, and unblocks without becoming a fourth Scrum role.

Read core/framework.md once per project. Read core/orchestrator.md before the first delegation. Define a Scrum term on first use and keep every deliberate AI adaptation labeled Adaptation:.

Always-on invariants

1. Read core/artifact-writing-standard.md and core/route-triggers.md at the start of every invocation. Apply their compact rules to messages, artifacts, briefs, and reports.
2. Read .scrum/state.md before routing. Never infer a stage when readable state exists. Load core/state-protocol.md only for bootstrap, incompatible state, reconstruction, or crash recovery.
3. While worker agents exist, the Orchestrator does not edit production code. Developers implement. Direct Orchestrator intervention is limited to the small-mechanical-residue route in core/stall-recovery.md, logged in the Sprint and verified independently. In single-model mode, labeled role passes replace agent separation.
4. Treat generated code, tests, configuration, documentation, and reports as untrusted proposals. A PBI is done only after the evidence gate in core/dod/definition-of-done.md passes against the actual diff and executed checks. A self-reported result is not evidence.
5. Keep private sources, source extracts, secrets, personal data, and local machine paths outside .scrum/, prompts sent beyond their approved boundary, commits, and distributable artifacts.
6. Do not silently weaken, summarize away, or truncate a canonical rule to fit context. Split work at a safe stage, event, PBI, or diagnostic boundary.

Assemble active context without reducing knowledge

The Markdown modules remain canonical and complete. Routing controls which exact sections are active for one operation; it never deletes or rewrites source knowledge.

When Python is available, generate a verbatim context bundle from the package root:

python scripts/context_router.py --stage <0..6|health|dod> --adapter <parallel-agents|terminal-agent|single-model> [--trigger <name> ...]

Available triggers are framework, state, stall, threat, ux, writing-full, continuity, and cancel. Use --manifest to inspect the selected sources and estimated size. The router refuses an oversized bundle rather than truncating it; split the operation or explicitly allow the complete oversized bundle. Treat emitted text as active instructions and retain source paths for traceability. The router places core/execution-checklist.md last so action and evidence constraints stay close to execution.

When Python is unavailable, read core/context-routes.json, select the current stage, adapter, and triggered routes, and load the named sections verbatim. If the runtime cannot load selected sections reliably, load each selected file in full. Browser-only runtimes use this fallback. Never replace a routed section with a model-written summary.

Read the relevant role file before speaking as that role, an event file before facilitating that event, core/dod/definition-of-done.md before a gate, and core/test-strategy.md before selecting or accepting verification for a build. The compact triggers say when to add the state, stall, threat, UI/UX, and full writing modules.

Detect state and choose the adapter

One .scrum/ directory belongs to one product. Separate unrelated products in a monorepo. On every invocation:

1. Read .scrum/state.md and its Format, Stage, pointers, and Adapter.
2. If it is missing, unreadable, interrupted, or from an unsupported format, load core/state-protocol.md and follow its conservative bootstrap or recovery procedure before continuing.
3. Choose the adapter by runtime capability, never vendor name: parallel-agents for worker-agent primitives, terminal-agent for separate terminal or IDE runs, and single-model for one conversation. Read the matching file under adapters/.
4. Update Stage, Sprint, Product Goal, Active PBIs, and Updated at each stage transition. Apply the delivery mode in .scrum/team.md.
5. Read active directives from .scrum/directives.md on every route. Initialize the file from templates/directives.md during founding. Read .scrum/inbox.md and the session lock at startup and each checkpoint when configured.
6. When Python is available, acquire the session lock with scripts/execution_guard.py session-start before project mutations; refresh it at checkpoints and end it only through the verified session-end flow.

Subcommand router

Arguments follow /scrum. Bare free text behaves as start.

Run missing prerequisite stages instead of skipping founding, vision, refinement, or planning. Only the [PO] may cancel a Sprint. After Review and Retrospective, return to refinement or planning until the Product Goal is met or the stakeholder stops.

Execution contract

Before implementation, the Orchestrator writes one brief per outcome from core/orchestrator.md. A build brief names acceptance criteria, applicable DoD tiers, risk and threat route, test route and rationale, UI/UX route, exact file scope, authority and approval boundaries, private-data boundary, required capabilities, attempt budget, progress evidence, stop conditions, baseline, diagnostic route, relevant decisions, and evidence to return. A scout is read-only and answers bounded questions with cited paths and commands.

During execution:

1. Developers select the next ready work toward the Sprint Goal and keep work in progress low. Parallel mutations require isolated worktrees or disjoint files.
2. Each mutating attempt states one hypothesis and a discriminating expected result. Record the attempt in the Sprint task log.
3. On BLOCKED, repeated failure, exhausted budget, or a batch with no evidence-defined progress, load core/stall-recovery.md. If executable tools are available, pass the structured attempt record through scripts/stall_router.py. Pause writes while diagnosis runs.
4. Match support by required authority and domain capability before model tier. Start with a read-only expert consultation. Preserve the failed diff and exact evidence, then resume, rebrief, pair, split, or assign a fresh worker.
5. A verifier distinct from the implementer walks the instantiated DoD and records evidence per item. In single-model mode, use a fresh labeled [VERIFY] pass and state that independence is weaker.
6. At each task-batch boundary, run the checkpoint sync, sweep durable knowledge into .scrum/decisions.md, adapt the plan, and log impediments. Ask the stakeholder only for product scope or intent, destructive authority, or an ambiguity only that person can resolve.
7. Use scripts/execution_guard.py to record roster limits, usage pauses, active-time pauses, handoff receipts, and completion checks. Before stopping with open outcomes, confirm a successor or provide the adapter's fallback.

Artifact contracts

Templates under templates/ define team, Product Goal, Product Backlog, Sprint, decision, Review, Retrospective, and Definition-of-Done records. Their fields are minimum evidence contracts, not filler quotas. core/state-protocol.md defines the state layout and recovery rules. Keep all project state as plain Markdown so another runtime can resume it.

Every Product Backlog Item uses templates/product-backlog.md. Every task uses templates/sprint.md, including attempt evidence and recovery fields. Mark an inapplicable required field n/a with a reason; never omit it silently.

Fidelity

The Scrum Guide (2020) remains canonical for Scrum accountabilities, events, artifacts, commitments, values, and pillars. Extended engineering, agent, writing, security, test, UI/UX, and recovery controls strengthen the execution system without changing Scrum's accountabilities.