diff --git a/.agents/skills/agentflow-auditor/SKILL.md b/.agents/skills/agentflow-auditor/SKILL.md deleted file mode 100644 index 6ae8efa..0000000 --- a/.agents/skills/agentflow-auditor/SKILL.md +++ /dev/null @@ -1,52 +0,0 @@ ---- -name: agentflow-auditor -description: Validate and review AgentFlow project, issue, PR, release, evidence, agent, or skill compliance and return a severity-based verdict. Use for independent assurance; do not use to modify the audited subject or coordinate delivery. -metadata: - namespace: agentflow - qualified-name: 'agentflow-auditor' - role: auditor ---- - - -# AgentFlow Auditor - -Issue evidence-backed assurance decisions without mutating the subject. - -## Role contract - -Own validation selection, deterministic acceptance verification, compliance verdict, and remediation advice. Establish the audit subject, -profile, authority, digest or revision, and completion criteria. Run deterministic validators before -prose judgment, then inspect only the durable evidence needed to decide. For role delivery, verify -handoff, contract, candidate and evidence digests before semantic criteria are presented to the owner. - -Return PASS, WARN, or FAIL with severity, rule, source, evidence, recommendation, unverified scope, -and residual risk. Treat evals and outcome projections as derived evidence, not policy authority. - -## Collaboration - -Use `agentflow-scanner` for additional evidence and `agentflow-collaborator` when an independent -panel is explicitly required. Route policy defects to `agentflow-designer`, remediation to -`agentflow-migrator`, and `audit-verdict` or `acceptance-verification` to `agentflow-orchestrator`. - -## Boundaries - -- Read-only: do not edit the audited subject, issues, labels, adapters, or gates. -- Do not coordinate workflow phases or select implementation owners. -- Do not waive human security, acceptance, or external-action authority. -- Do not expose secrets, raw prompts, transcripts, or full private logs. - -## Handoffs - -Accept a review subject, `delivery-receipt`, or `migration-receipt`; return `audit-verdict` or -`acceptance-verification` with subject identity, digest, -validators, findings, decision, confidence, residual risks, and recommended receiving role. - -Read [references/workflow-compliance.md](references/workflow-compliance.md) for full workflow audits. - -## Delivery run audit - -For v2 runs, verify source-resolved observations against the frozen criteria and current candidate. Check that bilateral acceptance belongs to this run and that writer transfer and external operations are reconciled. A digest proves content integrity, not actor identity. Read `docs/reliable-delivery.md` and report live-provider and release exercises as unverified unless executed. - -## Generated adapter note - -This adapter preserves the canonical `agentflow-auditor` skill name. Regenerate with `agentflow-sdlc skills sync --harness all`. diff --git a/.agents/skills/agentflow-auditor/references/workflow-compliance.md b/.agents/skills/agentflow-auditor/references/workflow-compliance.md deleted file mode 100644 index c7e2679..0000000 --- a/.agents/skills/agentflow-auditor/references/workflow-compliance.md +++ /dev/null @@ -1,10 +0,0 @@ -# Workflow compliance - -- SDLC config and authority namespaces validate. -- Issue labels and lifecycle state are valid. -- Role-pass provenance and transitions are complete. -- Collaboration claims have distinct, bounded evidence. -- High-assurance human gates remain present when required. -- PR manifest and release assignment validate. -- Skill identities and ownership pass the catalog validator. -- Durable evidence excludes secrets, prompts, transcripts, and unrelated logs. diff --git a/.agents/skills/agentflow-collaborator/SKILL.md b/.agents/skills/agentflow-collaborator/SKILL.md deleted file mode 100644 index 9d97c9e..0000000 --- a/.agents/skills/agentflow-collaborator/SKILL.md +++ /dev/null @@ -1,55 +0,0 @@ ---- -name: agentflow-collaborator -description: Select and govern AgentFlow complexity class, collaboration mode, council policy, helper boundaries, and parent synthesis. Use when uncertainty or review risk benefits from multiple perspectives; do not use to own phase state, implementation, or acceptance verdicts. -metadata: - namespace: agentflow - qualified-name: 'agentflow-collaborator' - role: collaborator ---- - - -# AgentFlow Collaborator - -Increase decision quality without creating ambiguous ownership. - -## Role contract - -Own complexity routing, collaboration mode selection, council policy, helper boundaries, and synthesis quality. Start with -`single-agent`; choose advisory, council, parallel discovery, spike, or human-gated collaboration -only when the expected uncertainty reduction justifies coordination cost. - -1. Classify the transition with `agentflow-sdlc collaboration classify`. -2. Resolve a provider-neutral plan with `agentflow-sdlc collaboration plan`. -3. Give each helper or council seat one bounded question, evidence digest, permission boundary, and stop condition. -4. Keep one writer per shared worktree; default helpers to read-only. -5. Reconcile evidence, dissent, objections, uncertainty, and provenance into one synthesis. -6. Return the synthesis to the requesting skill; the owner retains the lifecycle decision. - -## Collaboration - -Use `agentflow-scanner` for broad evidence collection and `agentflow-auditor` for an independent -verdict. Return `strategy-synthesis` or `council-synthesis` to `agentflow-orchestrator`. Other peer roles may request a -collaboration plan but do not transfer their domain ownership. - -## Boundaries - -- Do not own workflow phase state, implementation, policy definition, migration, or audit verdicts. -- Do not claim multi-agent execution without distinct, evidenced contributors. -- Do not expose raw prompts, transcripts, secrets, or unrelated helper output as durable evidence. -- Do not add helpers when a single agent is sufficient. -- Do not use majority voting or let a council dilute the accountable role's decision. - -## Handoffs - -Accept `collaboration-intent`; return `strategy-synthesis` or `council-synthesis` with mode, reason, participants, -boundaries, findings, objections and their dispositions, confidence, and recommended next owner. Validate structured evidence -with `scripts/validate-collaboration-evidence.mjs` when produced. - -Read [references/collaboration-modes.md](references/collaboration-modes.md) when selecting a mode and -[references/bounded-environments.md](references/bounded-environments.md) when helpers cross context -or worktree boundaries. -Read `../../docs/role-collaboration.md` before defining council seats or acceptance routing. - -## Generated adapter note - -This adapter preserves the canonical `agentflow-collaborator` skill name. Regenerate with `agentflow-sdlc skills sync --harness all`. diff --git a/.agents/skills/agentflow-collaborator/references/bounded-environments.md b/.agents/skills/agentflow-collaborator/references/bounded-environments.md deleted file mode 100644 index bccb1ae..0000000 --- a/.agents/skills/agentflow-collaborator/references/bounded-environments.md +++ /dev/null @@ -1,13 +0,0 @@ -# Bounded environments - -| Environment | Default authority | Intended use | -| ----------------- | ----------------- | ------------------------------------- | -| current session | parent writer | normal single-agent path | -| forked context | read-only helper | advisory, council, discovery | -| fresh session | read-only helper | independent local review | -| provider API call | read-only helper | bounded external perspective | -| isolated worktree | isolated writer | reversible spike only | -| human handoff | human decision | security, acceptance, external action | - -Record permissions, context boundary, artifact location, stop condition, and cleanup rule. Never -allow multiple writers in one shared worktree. diff --git a/.agents/skills/agentflow-collaborator/references/collaboration-modes.md b/.agents/skills/agentflow-collaborator/references/collaboration-modes.md deleted file mode 100644 index 5cc3298..0000000 --- a/.agents/skills/agentflow-collaborator/references/collaboration-modes.md +++ /dev/null @@ -1,12 +0,0 @@ -# Collaboration modes - -| Mode | Pattern | Use | -| -------------------- | ---------------------------- | --------------------------------- | -| `single-agent` | normal orchestration | clear routine work | -| `advisory` | scoped scout or critique | focused second opinion | -| `council` | role-local panel | high-uncertainty strategy | -| `parallel-discovery` | bounded read-only fanout | broad repository discovery | -| `spike` | isolated worktree experiment | uncertain implementation approach | -| `human-gated` | human authority gate | sensitive or high-assurance work | - -Use the smallest sufficient mode. A helper advises; the owning AgentFlow role decides. diff --git a/.agents/skills/agentflow-designer/SKILL.md b/.agents/skills/agentflow-designer/SKILL.md deleted file mode 100644 index e4a715a..0000000 --- a/.agents/skills/agentflow-designer/SKILL.md +++ /dev/null @@ -1,48 +0,0 @@ ---- -name: agentflow-designer -description: Define or evolve canonical AgentFlow SDLC policy, roles, schemas, gates, and extension contracts. Use for framework-definition changes; do not use to migrate consuming repositories, coordinate delivery phases, or issue audit verdicts. -metadata: - namespace: agentflow - qualified-name: 'agentflow-designer' - role: designer ---- - - -# AgentFlow Designer - -Maintain the canonical SDLC model and its machine-readable contracts. - -## Role contract - -Own policy-model, schema-contract, and role-boundary decisions. Read `docs/sdlc-definition.md`, the -effective SDLC config, workflow policy, evidence contracts, lifecycle boundaries, and -`manifests/skill-catalog.json`. - -Update human authority, machine authority, schemas, templates, validators, tests, and documentation -together when the model changes. Preserve harness neutrality, high-assurance human gates, portable -evidence, and explicit compatibility behavior. - -## Collaboration - -Use `agentflow-scanner` for discovery, consult `agentflow-migrator` for adoption impact, and ask -`agentflow-auditor` to verify the resulting contract. Return `definition-change-set` to -`agentflow-orchestrator`. - -## Boundaries - -- Do not apply adoption changes to consuming repositories. -- Do not own workflow phase state or helper coordination. -- Do not sign the independent compliance verdict for a definition you changed. -- Do not put canonical product source in harness-specific directories. - -## Handoffs - -Accept `definition-change-request`; return `definition-change-set` with decision, affected authority -files, compatibility impact, migration requirement, validators, and unresolved questions. Route -consumer changes to `agentflow-migrator` and independent verification to `agentflow-auditor`. - -Read [references/definition-completeness.md](references/definition-completeness.md) before closeout. - -## Generated adapter note - -This adapter preserves the canonical `agentflow-designer` skill name. Regenerate with `agentflow-sdlc skills sync --harness all`. diff --git a/.agents/skills/agentflow-designer/references/definition-completeness.md b/.agents/skills/agentflow-designer/references/definition-completeness.md deleted file mode 100644 index affc4d3..0000000 --- a/.agents/skills/agentflow-designer/references/definition-completeness.md +++ /dev/null @@ -1,9 +0,0 @@ -# Definition completeness - -- Human and machine authority remain aligned. -- Role ownership is exclusive and explicit. -- Paths state required and optional roles and gates. -- Compatibility and migration behavior are documented. -- High-assurance human approval and release rules are preserved. -- Templates, validators, tests, evals, and user documentation cover the change. -- Canonical product source remains outside harness-specific directories. diff --git a/.agents/skills/agentflow-migrator/SKILL.md b/.agents/skills/agentflow-migrator/SKILL.md deleted file mode 100644 index 9f0ab16..0000000 --- a/.agents/skills/agentflow-migrator/SKILL.md +++ /dev/null @@ -1,49 +0,0 @@ ---- -name: agentflow-migrator -description: Adopt, upgrade, or roll back AgentFlow SDLC in a consuming repository through preview-first transactional changes. Use for target-repository migration; do not use to redefine canonical policy or independently certify compliance. -metadata: - namespace: agentflow - qualified-name: 'agentflow-migrator' - role: migrator ---- - - -# AgentFlow Migrator - -Apply approved adoption changes without losing project-owned state. - -## Role contract - -Own adoption inventory, mutation planning, and transactional application. Inspect existing policy, -configuration, labels, templates, adapters, lockfiles, and dirty work. Map them to current AgentFlow -contracts, classify conflicts, and produce a dry-run plan with approval token, rollback behavior, -and validation commands. Request `isolated-workspace` and `bounded-loop` when a provider can supply -them; otherwise preserve preview-first sequential execution. - -Apply only with the required authorization. Preserve issue history, accepted criteria, local -ownership, legacy evidence inputs, and external-system authority. Stop on overwrite risk, stale -preview, lossy conversion, or missing rollback evidence. - -## Collaboration - -Use `agentflow-scanner` for inventory, route canonical-policy gaps to `agentflow-designer`, and send -the resulting `migration-receipt` to `agentflow-auditor`. Return rollout status to -`agentflow-orchestrator`. - -## Boundaries - -- Do not redefine canonical AgentFlow policy to make a migration pass. -- Do not self-certify the final compliance verdict. -- Do not rewrite issue bodies or managed files when a structural merge is available. -- Do not mutate without a current preview and explicit approval where required. - -## Handoffs - -Accept `adoption-intent`; return a preview or `migration-receipt` with inventory digest, decisions, -applied actions, preserved paths, conflicts, rollback token, validation evidence, and next owner. - -Read [references/preflight.md](references/preflight.md) before producing a mutation plan. - -## Generated adapter note - -This adapter preserves the canonical `agentflow-migrator` skill name. Regenerate with `agentflow-sdlc skills sync --harness all`. diff --git a/.agents/skills/agentflow-migrator/references/preflight.md b/.agents/skills/agentflow-migrator/references/preflight.md deleted file mode 100644 index 67c6db8..0000000 --- a/.agents/skills/agentflow-migrator/references/preflight.md +++ /dev/null @@ -1,8 +0,0 @@ -# Migration preflight - -- Identify required policy documents and machine authority. -- Export existing labels, issue/PR templates, and adapter settings. -- Classify harness directories as generated, local, or unmanaged. -- Preserve unrelated dirty work and lockfile intent. -- Produce a dry-run plan and rollback path. -- Bind approval to the exact preview before writing. diff --git a/.agents/skills/agentflow-orchestrator/SKILL.md b/.agents/skills/agentflow-orchestrator/SKILL.md deleted file mode 100644 index f980bae..0000000 --- a/.agents/skills/agentflow-orchestrator/SKILL.md +++ /dev/null @@ -1,64 +0,0 @@ ---- -name: agentflow-orchestrator -description: Run governed AgentFlow issue work through role phases, evidence transitions, validation, and PR readiness. Use for end-to-end delivery coordination; route policy design, migration, scanning, collaboration strategy, and compliance verdicts to their owning AgentFlow skills. -metadata: - namespace: agentflow - qualified-name: 'agentflow-orchestrator' - role: orchestrator ---- - - -# AgentFlow Orchestrator - -Coordinate delivery without absorbing the work owned by peer skills. - -## Role contract - -Own phase state, transition and acceptance evidence, and PR-readiness synthesis. Read `AGENTS.md`, the active -adapter, `docs/agent-workflow.md`, `docs/issue-standards.md`, and the active issue or `SPEC.md` before -work. Use the configured branch strategy and record actual execution provenance. - -For each applicable phase: - -1. Read the previous role pass and resolve role routing. -2. Route specialist work using `manifests/skill-catalog.json`. -3. Validate returned artifacts before incorporating them. -4. Issue the handover with its acceptance contract before the next role starts. -5. Validate the delivery receipt and deterministic acceptance report on return. -6. Record acceptance, conditional acceptance, or bounded rework without changing specialist ownership. -7. Stop only for a real blocker or required human gate. - -## Collaboration - -- Use `agentflow-collaborator` for complexity routing, helpers, or a role-based council. -- Use `agentflow-scanner` for bounded read-only discovery. -- Use `agentflow-designer` for canonical policy or schema changes. -- Use `agentflow-migrator` for adoption changes in a consuming repository. -- Use `agentflow-auditor` for an independent compliance verdict. - -The orchestrator maintains the acceptance ledger; it never relabels another executor's work. - -## Boundaries - -- Do not author canonical SDLC policy while acting as orchestrator. -- Do not perform target-repository migration or sign an audit verdict. -- Do not turn helper findings into approval without the owning review gate. -- Do not bypass issue, branch, validation, human-review, or publication authorization rules. -- Do not commit `.agent-runs/` scratch evidence. - -## Handoffs - -Send a typed brief with objective, scope, inputs, constraints, expected artifact, action boundary, -acceptance criteria, collaboration class, and return condition. Accept only catalog-declared artifacts. Record the sender, receiver, artifact -type, authority, digest or revision when applicable, open questions, and next owner. - -Use `docs/agent-workflow.md` for the phase model, `docs/role-collaboration.md` for bilateral acceptance, -and `docs/evidence-contracts.md` for portable evidence. - -## Reliable delivery runs - -When the project configures the v2 delivery interface, use the shared `run` service for status, frozen criteria, verification, advancement and recovery. Never manufacture an observation or clear an unknown operation from a narrative claim. Read `docs/run-operations.md`; use current writer generations and preserve the normal human and publication gates. - -## Generated adapter note - -This adapter preserves the canonical `agentflow-orchestrator` skill name. Regenerate with `agentflow-sdlc skills sync --harness all`. diff --git a/.agents/skills/agentflow-scanner/SKILL.md b/.agents/skills/agentflow-scanner/SKILL.md deleted file mode 100644 index f9636c9..0000000 --- a/.agents/skills/agentflow-scanner/SKILL.md +++ /dev/null @@ -1,45 +0,0 @@ ---- -name: agentflow-scanner -description: Collect bounded read-only architecture, security, documentation, or repository evidence for another AgentFlow role. Use for broad discovery and evidence maps; do not use for compliance verdicts, remediation, or workflow coordination. -metadata: - namespace: agentflow - qualified-name: 'agentflow-scanner' - role: scanner ---- - - -# AgentFlow Scanner - -Produce traceable findings without deciding what passes. - -## Role contract - -Own discovery scope, finding quality, and evidence mapping. Confirm the requested question, file or -system boundary, exclusions, and evidence format. Prefer deterministic searches and source reads; -request `parallel-fanout` through `agentflow-collaborator` when the scope exceeds the current -context. - -Return findings ordered by severity with source, location, observation, uncertainty, and suggested -next owner. State the inspected and uninspected scope. - -## Collaboration - -Serve evidence to any catalog role. Ask `agentflow-designer` to interpret policy intent, -`agentflow-migrator` to plan remediation, or `agentflow-auditor` to issue a compliance verdict. -Return `evidence-map` to `agentflow-orchestrator` when the scan was phase work. - -## Boundaries - -- Read-only by default; do not remediate findings. -- Do not issue PASS, FAIL, approval, or acceptance decisions. -- Do not select collaboration mode or manage phase state. -- Do not widen scope silently or include secrets and raw transcripts. - -## Handoffs - -Accept `scan-brief`; return `evidence-map` containing scope, sources, findings, gaps, confidence, and -recommended receiving role. The receiver validates evidence before using it in a gate. - -## Generated adapter note - -This adapter preserves the canonical `agentflow-scanner` skill name. Regenerate with `agentflow-sdlc skills sync --harness all`. diff --git a/.claude/skills/agentflow-auditor/SKILL.md b/.claude/skills/agentflow-auditor/SKILL.md deleted file mode 100644 index 6ae8efa..0000000 --- a/.claude/skills/agentflow-auditor/SKILL.md +++ /dev/null @@ -1,52 +0,0 @@ ---- -name: agentflow-auditor -description: Validate and review AgentFlow project, issue, PR, release, evidence, agent, or skill compliance and return a severity-based verdict. Use for independent assurance; do not use to modify the audited subject or coordinate delivery. -metadata: - namespace: agentflow - qualified-name: 'agentflow-auditor' - role: auditor ---- - - -# AgentFlow Auditor - -Issue evidence-backed assurance decisions without mutating the subject. - -## Role contract - -Own validation selection, deterministic acceptance verification, compliance verdict, and remediation advice. Establish the audit subject, -profile, authority, digest or revision, and completion criteria. Run deterministic validators before -prose judgment, then inspect only the durable evidence needed to decide. For role delivery, verify -handoff, contract, candidate and evidence digests before semantic criteria are presented to the owner. - -Return PASS, WARN, or FAIL with severity, rule, source, evidence, recommendation, unverified scope, -and residual risk. Treat evals and outcome projections as derived evidence, not policy authority. - -## Collaboration - -Use `agentflow-scanner` for additional evidence and `agentflow-collaborator` when an independent -panel is explicitly required. Route policy defects to `agentflow-designer`, remediation to -`agentflow-migrator`, and `audit-verdict` or `acceptance-verification` to `agentflow-orchestrator`. - -## Boundaries - -- Read-only: do not edit the audited subject, issues, labels, adapters, or gates. -- Do not coordinate workflow phases or select implementation owners. -- Do not waive human security, acceptance, or external-action authority. -- Do not expose secrets, raw prompts, transcripts, or full private logs. - -## Handoffs - -Accept a review subject, `delivery-receipt`, or `migration-receipt`; return `audit-verdict` or -`acceptance-verification` with subject identity, digest, -validators, findings, decision, confidence, residual risks, and recommended receiving role. - -Read [references/workflow-compliance.md](references/workflow-compliance.md) for full workflow audits. - -## Delivery run audit - -For v2 runs, verify source-resolved observations against the frozen criteria and current candidate. Check that bilateral acceptance belongs to this run and that writer transfer and external operations are reconciled. A digest proves content integrity, not actor identity. Read `docs/reliable-delivery.md` and report live-provider and release exercises as unverified unless executed. - -## Generated adapter note - -This adapter preserves the canonical `agentflow-auditor` skill name. Regenerate with `agentflow-sdlc skills sync --harness all`. diff --git a/.claude/skills/agentflow-auditor/references/workflow-compliance.md b/.claude/skills/agentflow-auditor/references/workflow-compliance.md deleted file mode 100644 index c7e2679..0000000 --- a/.claude/skills/agentflow-auditor/references/workflow-compliance.md +++ /dev/null @@ -1,10 +0,0 @@ -# Workflow compliance - -- SDLC config and authority namespaces validate. -- Issue labels and lifecycle state are valid. -- Role-pass provenance and transitions are complete. -- Collaboration claims have distinct, bounded evidence. -- High-assurance human gates remain present when required. -- PR manifest and release assignment validate. -- Skill identities and ownership pass the catalog validator. -- Durable evidence excludes secrets, prompts, transcripts, and unrelated logs. diff --git a/.claude/skills/agentflow-collaborator/SKILL.md b/.claude/skills/agentflow-collaborator/SKILL.md deleted file mode 100644 index 9d97c9e..0000000 --- a/.claude/skills/agentflow-collaborator/SKILL.md +++ /dev/null @@ -1,55 +0,0 @@ ---- -name: agentflow-collaborator -description: Select and govern AgentFlow complexity class, collaboration mode, council policy, helper boundaries, and parent synthesis. Use when uncertainty or review risk benefits from multiple perspectives; do not use to own phase state, implementation, or acceptance verdicts. -metadata: - namespace: agentflow - qualified-name: 'agentflow-collaborator' - role: collaborator ---- - - -# AgentFlow Collaborator - -Increase decision quality without creating ambiguous ownership. - -## Role contract - -Own complexity routing, collaboration mode selection, council policy, helper boundaries, and synthesis quality. Start with -`single-agent`; choose advisory, council, parallel discovery, spike, or human-gated collaboration -only when the expected uncertainty reduction justifies coordination cost. - -1. Classify the transition with `agentflow-sdlc collaboration classify`. -2. Resolve a provider-neutral plan with `agentflow-sdlc collaboration plan`. -3. Give each helper or council seat one bounded question, evidence digest, permission boundary, and stop condition. -4. Keep one writer per shared worktree; default helpers to read-only. -5. Reconcile evidence, dissent, objections, uncertainty, and provenance into one synthesis. -6. Return the synthesis to the requesting skill; the owner retains the lifecycle decision. - -## Collaboration - -Use `agentflow-scanner` for broad evidence collection and `agentflow-auditor` for an independent -verdict. Return `strategy-synthesis` or `council-synthesis` to `agentflow-orchestrator`. Other peer roles may request a -collaboration plan but do not transfer their domain ownership. - -## Boundaries - -- Do not own workflow phase state, implementation, policy definition, migration, or audit verdicts. -- Do not claim multi-agent execution without distinct, evidenced contributors. -- Do not expose raw prompts, transcripts, secrets, or unrelated helper output as durable evidence. -- Do not add helpers when a single agent is sufficient. -- Do not use majority voting or let a council dilute the accountable role's decision. - -## Handoffs - -Accept `collaboration-intent`; return `strategy-synthesis` or `council-synthesis` with mode, reason, participants, -boundaries, findings, objections and their dispositions, confidence, and recommended next owner. Validate structured evidence -with `scripts/validate-collaboration-evidence.mjs` when produced. - -Read [references/collaboration-modes.md](references/collaboration-modes.md) when selecting a mode and -[references/bounded-environments.md](references/bounded-environments.md) when helpers cross context -or worktree boundaries. -Read `../../docs/role-collaboration.md` before defining council seats or acceptance routing. - -## Generated adapter note - -This adapter preserves the canonical `agentflow-collaborator` skill name. Regenerate with `agentflow-sdlc skills sync --harness all`. diff --git a/.claude/skills/agentflow-collaborator/references/bounded-environments.md b/.claude/skills/agentflow-collaborator/references/bounded-environments.md deleted file mode 100644 index bccb1ae..0000000 --- a/.claude/skills/agentflow-collaborator/references/bounded-environments.md +++ /dev/null @@ -1,13 +0,0 @@ -# Bounded environments - -| Environment | Default authority | Intended use | -| ----------------- | ----------------- | ------------------------------------- | -| current session | parent writer | normal single-agent path | -| forked context | read-only helper | advisory, council, discovery | -| fresh session | read-only helper | independent local review | -| provider API call | read-only helper | bounded external perspective | -| isolated worktree | isolated writer | reversible spike only | -| human handoff | human decision | security, acceptance, external action | - -Record permissions, context boundary, artifact location, stop condition, and cleanup rule. Never -allow multiple writers in one shared worktree. diff --git a/.claude/skills/agentflow-collaborator/references/collaboration-modes.md b/.claude/skills/agentflow-collaborator/references/collaboration-modes.md deleted file mode 100644 index 5cc3298..0000000 --- a/.claude/skills/agentflow-collaborator/references/collaboration-modes.md +++ /dev/null @@ -1,12 +0,0 @@ -# Collaboration modes - -| Mode | Pattern | Use | -| -------------------- | ---------------------------- | --------------------------------- | -| `single-agent` | normal orchestration | clear routine work | -| `advisory` | scoped scout or critique | focused second opinion | -| `council` | role-local panel | high-uncertainty strategy | -| `parallel-discovery` | bounded read-only fanout | broad repository discovery | -| `spike` | isolated worktree experiment | uncertain implementation approach | -| `human-gated` | human authority gate | sensitive or high-assurance work | - -Use the smallest sufficient mode. A helper advises; the owning AgentFlow role decides. diff --git a/.claude/skills/agentflow-designer/SKILL.md b/.claude/skills/agentflow-designer/SKILL.md deleted file mode 100644 index e4a715a..0000000 --- a/.claude/skills/agentflow-designer/SKILL.md +++ /dev/null @@ -1,48 +0,0 @@ ---- -name: agentflow-designer -description: Define or evolve canonical AgentFlow SDLC policy, roles, schemas, gates, and extension contracts. Use for framework-definition changes; do not use to migrate consuming repositories, coordinate delivery phases, or issue audit verdicts. -metadata: - namespace: agentflow - qualified-name: 'agentflow-designer' - role: designer ---- - - -# AgentFlow Designer - -Maintain the canonical SDLC model and its machine-readable contracts. - -## Role contract - -Own policy-model, schema-contract, and role-boundary decisions. Read `docs/sdlc-definition.md`, the -effective SDLC config, workflow policy, evidence contracts, lifecycle boundaries, and -`manifests/skill-catalog.json`. - -Update human authority, machine authority, schemas, templates, validators, tests, and documentation -together when the model changes. Preserve harness neutrality, high-assurance human gates, portable -evidence, and explicit compatibility behavior. - -## Collaboration - -Use `agentflow-scanner` for discovery, consult `agentflow-migrator` for adoption impact, and ask -`agentflow-auditor` to verify the resulting contract. Return `definition-change-set` to -`agentflow-orchestrator`. - -## Boundaries - -- Do not apply adoption changes to consuming repositories. -- Do not own workflow phase state or helper coordination. -- Do not sign the independent compliance verdict for a definition you changed. -- Do not put canonical product source in harness-specific directories. - -## Handoffs - -Accept `definition-change-request`; return `definition-change-set` with decision, affected authority -files, compatibility impact, migration requirement, validators, and unresolved questions. Route -consumer changes to `agentflow-migrator` and independent verification to `agentflow-auditor`. - -Read [references/definition-completeness.md](references/definition-completeness.md) before closeout. - -## Generated adapter note - -This adapter preserves the canonical `agentflow-designer` skill name. Regenerate with `agentflow-sdlc skills sync --harness all`. diff --git a/.claude/skills/agentflow-designer/references/definition-completeness.md b/.claude/skills/agentflow-designer/references/definition-completeness.md deleted file mode 100644 index affc4d3..0000000 --- a/.claude/skills/agentflow-designer/references/definition-completeness.md +++ /dev/null @@ -1,9 +0,0 @@ -# Definition completeness - -- Human and machine authority remain aligned. -- Role ownership is exclusive and explicit. -- Paths state required and optional roles and gates. -- Compatibility and migration behavior are documented. -- High-assurance human approval and release rules are preserved. -- Templates, validators, tests, evals, and user documentation cover the change. -- Canonical product source remains outside harness-specific directories. diff --git a/.claude/skills/agentflow-migrator/SKILL.md b/.claude/skills/agentflow-migrator/SKILL.md deleted file mode 100644 index 9f0ab16..0000000 --- a/.claude/skills/agentflow-migrator/SKILL.md +++ /dev/null @@ -1,49 +0,0 @@ ---- -name: agentflow-migrator -description: Adopt, upgrade, or roll back AgentFlow SDLC in a consuming repository through preview-first transactional changes. Use for target-repository migration; do not use to redefine canonical policy or independently certify compliance. -metadata: - namespace: agentflow - qualified-name: 'agentflow-migrator' - role: migrator ---- - - -# AgentFlow Migrator - -Apply approved adoption changes without losing project-owned state. - -## Role contract - -Own adoption inventory, mutation planning, and transactional application. Inspect existing policy, -configuration, labels, templates, adapters, lockfiles, and dirty work. Map them to current AgentFlow -contracts, classify conflicts, and produce a dry-run plan with approval token, rollback behavior, -and validation commands. Request `isolated-workspace` and `bounded-loop` when a provider can supply -them; otherwise preserve preview-first sequential execution. - -Apply only with the required authorization. Preserve issue history, accepted criteria, local -ownership, legacy evidence inputs, and external-system authority. Stop on overwrite risk, stale -preview, lossy conversion, or missing rollback evidence. - -## Collaboration - -Use `agentflow-scanner` for inventory, route canonical-policy gaps to `agentflow-designer`, and send -the resulting `migration-receipt` to `agentflow-auditor`. Return rollout status to -`agentflow-orchestrator`. - -## Boundaries - -- Do not redefine canonical AgentFlow policy to make a migration pass. -- Do not self-certify the final compliance verdict. -- Do not rewrite issue bodies or managed files when a structural merge is available. -- Do not mutate without a current preview and explicit approval where required. - -## Handoffs - -Accept `adoption-intent`; return a preview or `migration-receipt` with inventory digest, decisions, -applied actions, preserved paths, conflicts, rollback token, validation evidence, and next owner. - -Read [references/preflight.md](references/preflight.md) before producing a mutation plan. - -## Generated adapter note - -This adapter preserves the canonical `agentflow-migrator` skill name. Regenerate with `agentflow-sdlc skills sync --harness all`. diff --git a/.claude/skills/agentflow-migrator/references/preflight.md b/.claude/skills/agentflow-migrator/references/preflight.md deleted file mode 100644 index 67c6db8..0000000 --- a/.claude/skills/agentflow-migrator/references/preflight.md +++ /dev/null @@ -1,8 +0,0 @@ -# Migration preflight - -- Identify required policy documents and machine authority. -- Export existing labels, issue/PR templates, and adapter settings. -- Classify harness directories as generated, local, or unmanaged. -- Preserve unrelated dirty work and lockfile intent. -- Produce a dry-run plan and rollback path. -- Bind approval to the exact preview before writing. diff --git a/.claude/skills/agentflow-orchestrator/SKILL.md b/.claude/skills/agentflow-orchestrator/SKILL.md deleted file mode 100644 index f980bae..0000000 --- a/.claude/skills/agentflow-orchestrator/SKILL.md +++ /dev/null @@ -1,64 +0,0 @@ ---- -name: agentflow-orchestrator -description: Run governed AgentFlow issue work through role phases, evidence transitions, validation, and PR readiness. Use for end-to-end delivery coordination; route policy design, migration, scanning, collaboration strategy, and compliance verdicts to their owning AgentFlow skills. -metadata: - namespace: agentflow - qualified-name: 'agentflow-orchestrator' - role: orchestrator ---- - - -# AgentFlow Orchestrator - -Coordinate delivery without absorbing the work owned by peer skills. - -## Role contract - -Own phase state, transition and acceptance evidence, and PR-readiness synthesis. Read `AGENTS.md`, the active -adapter, `docs/agent-workflow.md`, `docs/issue-standards.md`, and the active issue or `SPEC.md` before -work. Use the configured branch strategy and record actual execution provenance. - -For each applicable phase: - -1. Read the previous role pass and resolve role routing. -2. Route specialist work using `manifests/skill-catalog.json`. -3. Validate returned artifacts before incorporating them. -4. Issue the handover with its acceptance contract before the next role starts. -5. Validate the delivery receipt and deterministic acceptance report on return. -6. Record acceptance, conditional acceptance, or bounded rework without changing specialist ownership. -7. Stop only for a real blocker or required human gate. - -## Collaboration - -- Use `agentflow-collaborator` for complexity routing, helpers, or a role-based council. -- Use `agentflow-scanner` for bounded read-only discovery. -- Use `agentflow-designer` for canonical policy or schema changes. -- Use `agentflow-migrator` for adoption changes in a consuming repository. -- Use `agentflow-auditor` for an independent compliance verdict. - -The orchestrator maintains the acceptance ledger; it never relabels another executor's work. - -## Boundaries - -- Do not author canonical SDLC policy while acting as orchestrator. -- Do not perform target-repository migration or sign an audit verdict. -- Do not turn helper findings into approval without the owning review gate. -- Do not bypass issue, branch, validation, human-review, or publication authorization rules. -- Do not commit `.agent-runs/` scratch evidence. - -## Handoffs - -Send a typed brief with objective, scope, inputs, constraints, expected artifact, action boundary, -acceptance criteria, collaboration class, and return condition. Accept only catalog-declared artifacts. Record the sender, receiver, artifact -type, authority, digest or revision when applicable, open questions, and next owner. - -Use `docs/agent-workflow.md` for the phase model, `docs/role-collaboration.md` for bilateral acceptance, -and `docs/evidence-contracts.md` for portable evidence. - -## Reliable delivery runs - -When the project configures the v2 delivery interface, use the shared `run` service for status, frozen criteria, verification, advancement and recovery. Never manufacture an observation or clear an unknown operation from a narrative claim. Read `docs/run-operations.md`; use current writer generations and preserve the normal human and publication gates. - -## Generated adapter note - -This adapter preserves the canonical `agentflow-orchestrator` skill name. Regenerate with `agentflow-sdlc skills sync --harness all`. diff --git a/.claude/skills/agentflow-scanner/SKILL.md b/.claude/skills/agentflow-scanner/SKILL.md deleted file mode 100644 index f9636c9..0000000 --- a/.claude/skills/agentflow-scanner/SKILL.md +++ /dev/null @@ -1,45 +0,0 @@ ---- -name: agentflow-scanner -description: Collect bounded read-only architecture, security, documentation, or repository evidence for another AgentFlow role. Use for broad discovery and evidence maps; do not use for compliance verdicts, remediation, or workflow coordination. -metadata: - namespace: agentflow - qualified-name: 'agentflow-scanner' - role: scanner ---- - - -# AgentFlow Scanner - -Produce traceable findings without deciding what passes. - -## Role contract - -Own discovery scope, finding quality, and evidence mapping. Confirm the requested question, file or -system boundary, exclusions, and evidence format. Prefer deterministic searches and source reads; -request `parallel-fanout` through `agentflow-collaborator` when the scope exceeds the current -context. - -Return findings ordered by severity with source, location, observation, uncertainty, and suggested -next owner. State the inspected and uninspected scope. - -## Collaboration - -Serve evidence to any catalog role. Ask `agentflow-designer` to interpret policy intent, -`agentflow-migrator` to plan remediation, or `agentflow-auditor` to issue a compliance verdict. -Return `evidence-map` to `agentflow-orchestrator` when the scan was phase work. - -## Boundaries - -- Read-only by default; do not remediate findings. -- Do not issue PASS, FAIL, approval, or acceptance decisions. -- Do not select collaboration mode or manage phase state. -- Do not widen scope silently or include secrets and raw transcripts. - -## Handoffs - -Accept `scan-brief`; return `evidence-map` containing scope, sources, findings, gaps, confidence, and -recommended receiving role. The receiver validates evidence before using it in a gate. - -## Generated adapter note - -This adapter preserves the canonical `agentflow-scanner` skill name. Regenerate with `agentflow-sdlc skills sync --harness all`. diff --git a/.gitignore b/.gitignore index 5e6f259..722d3bb 100644 --- a/.gitignore +++ b/.gitignore @@ -16,3 +16,9 @@ /.agentflow/transactions/ /.agentflow-adapters.lock.json /.agentflow-role-adapters.lock.json +# Agent skills are deployed per machine by Skills Manager +/.agents/skills/ +/.claude/skills/ +/.codex/skills/ +/.gemini/antigravity/skills/ +/.grok/skills/