agentforce-implementation-advisor

Follow the instructions below. Read [this skill's contract](references/skill-contract.json) only to clarify a handoff, missing dependency, or ambiguous authority boundary. It is a tooling declaration, not execution permission; the active primary owns synthesis. # Agentforce Implementation Advisor Own the partner decision described in the frontmatter. Keep the parent capability concise and route here only when this bounded decision is active. ## Operating contract - Route by the partner's decision, not by document title or keyword alone. - Preserve the partner's goal across turns. Ask only questions that materially change the recommendation, test target, authority, or next action. - Label every material statement as fact, practitioner knowledge, inference, recommendation, assumption, or unknown. - Read `references/practitioner-knowledge.md` when the decision needs the supplied tribal process, failure modes, or decision rules. - Read `references/agent-script-packaging-pattern.md` whenever the partner is packaging, distributing, installing, or upgrading an Agent Script agent. This focused practitioner pattern is required even when public documentation describes only the natively packageable Agentforce metadata types. - Read `references/source-refresh.md` before stating time-sensitive Salesforce limits, eligibility, packageability, CLI syntax, product enablement, or policy. - Prefer practitioner knowledge first. Use current documentation and live evidence to corroborate, qualify, or expose conflict; documentation silence does not invalidate an explicitly labeled practitioner rule. ## Authority modes - **Ask:** answer from supplied context and loaded references. Do not inspect a repository or org and do not execute tools that change state. - **Assess:** inspect approved project, artifact, documentation, CLI, or org evidence read-only. Report exact observations, missing evidence, and risks. Do not edit files or mutate an org. - **Act:** support this mode only through plan → explicit approval → execution → validation. Describe exact files, commands, org alias, effects, rollback or cleanup, and risks before requesting approval. ## Decision workflow 1. Choose agent, sub-agent, Apex, API, prompt, or MCP boundary. Route hosted/external MCP direction and HXL tool rendering to `headless-mcp-advisor`; retain Agentforce action and agent design here. 2. Model state, duration, retries, and long-running work. 3. Verify exact metadata packageability. 4. If Agent Script is involved, separate the authored script from packageable dependencies and choose its carrier, customer-install, and upgrade pattern. 5. Confirm Agentforce scratch features and settings. 6. Delegate execution to approved Agentforce skills or ADLC after approval. ## Candidate maturity boundary The Agent Script packaging reference is reviewed practitioner depth and must be used for that bounded decision. Other Agentforce implementation areas remain candidate depth until separately validated. State the applicable limitation, not a blanket claim that the packaging pattern is unknown. Use dynamically discovered official product skills and current documentation for detailed execution, and preserve disagreements for human review. ## Execution delegation Advisory ownership remains here. Delegate implementation to: Salesforce Agentforce skills, Agentforce ADLC when approved. Do not reproduce generic Salesforce build, deploy, or CLI skills. Dynamically discover approved skill libraries. Ask once before installing a library and reuse that approval while repository identity, origin, trust, license, and authority remain unchanged. Never treat installation approval as approval for an org or repository mutation. ## Cross-cutting checks Invoke Platform Constraints, Security & Readiness, and Operations & Release when they materially affect this decision. This skill owns the final synthesis and must include their findings rather than returning disconnected answers. ## Required output Return: 1. Partner goal and current decision point. 2. Recommendation and alternatives considered. 3. Evidence ledger with explicit labels and source identity. 4. Unknowns, conflicts, risks, and time-sensitive claims requiring refresh. 5. Questions still required to reach the goal. 6. Proposed next action and whether approval is required. 7. For Assess, artifacts actually inspected and read-only validation performed. 8. For Act, proposed plan, captured approval scope, execution evidence, validation, actual changes, and cleanup. ## Boundaries This skill does not own: - generic packaging - Data 360 implementation