Task instructions

Task Instructions

Agent instruction protocol: 1.0.

Behavioral compliance depends on the connected agent or integration. The Day Shift CLI exposes repository policy and evidence, but it cannot force an unsupported agent to obey them.

The verified CLI automation protocol version 2 policy, evidence, next-action, and run-continuation examples are frozen in apps/docs/content-agent/agent-reference-examples.json.

Use these instructions only for one selected task-definition.md. The task definition is the work agreement; its paired implementation-summary.md is the canonical technical evidence artifact.

Before implementation

1. Confirm the current-work selection bundle and inherited run directory.
2. Read the task objective, dependencies, target paths, current and desired behavior, acceptance criteria, validation commands, rollback, constraints, and open questions.
3. Run the revision-bound link gate. Apply only safe selected-artifact candidates with the exact preview evidence.
4. Obtain a current read-only readiness review with zero blockers.
5. Use the separate readiness-authorization transition when required.
6. Open one stable implementation attempt. For Runtime or Hybrid evidence, capture its baseline after authorization and before the first target write.

Warnings remain evidence to carry forward. Errors, stale revisions, missing dependencies, later-sibling dependencies, or unresolved planning questions stop implementation.

During implementation

• Change only declared target paths. Preserve unrelated user work and protected paths.
• Do not let agent mode broaden authority. Autonomous, Guided, and Interactive agents share the same stop conditions.
• Record discovered work instead of silently expanding scope.
• Build compiled output before validation when the declared command consumes compiled artifacts.
• Execute each declared validation through its selected command-owned, manual, or imported evidence posture. Do not claim that a read-only planning validation executed task commands.
• Keep expected-versus-actual changed-path accounting and investigate suspicious deletion, file-type, or test-coverage changes.

Machine-readable policy and evidence example

examples.agentContext comes from agent context --artifact <path> --intent <text> --format json --detail full. It keeps explicit fresh selection, observed lifecycle state, declared and effective mode, winning policy source, all seven invariant safety stops, evidence profile, required selected-artifact/workflow/prompt evidence lanes, current blockers, and one descriptor-backed recommended action in the same complete envelope.

The fixture intentionally pairs effectiveAgentMode: "guided" with evidenceProfile: "runtime"; neither field derives from or overrides the other. The selected task’s declared target paths and exclusions remain the scope contract even though the bounded example projects only action-relevant context. Protected paths still include unrelated user work, secrets, release state, generated state outside the task, and every path the selected contract excludes.

examples.nextAction proves that navigation is read-only and resolves implementation_incomplete to the task implementation workflow without fabricating a write command. examples.runStatus and examples.runResume prove a fresh full-cycle-strict run with read-only metadata posture, a current active artifact and gate, manifest completeness, and a re-resolved callable next action. Run history remains evidence only; it cannot authorize the action it names.

Evidence and closeout

1. Build the canonical implementation summary.
2. Complete its body with changed paths, acceptance results, validation evidence or structured exceptions, scope classification, runtime boundary, deviations, risks, and discovered work.
3. Rebuild the summary so runtime attribution and task-scoped changes are derived from current evidence.
4. Apply the explicit implementation-completion transition when applicable.
5. Run read-only summary review, apply an explicit disposition from its current receipt, rerun review, and close the task only from a fresh close-authorizing receipt.
6. Re-open both task artifacts and verify terminal status, the closed attempt, accepted disposition, current required validations, and preserved runtime attribution.

Review, recommendation, disposition, and close are distinct steps. See Review boundaries.

Stop conditions

Stop before writing when scope is vague, readiness is missing, a target is unexpectedly dirty without an adoption contract, a baseline is stale, validation cannot meaningfully test acceptance, or a requested action needs destructive, external-system, release, or additional planning authority.

Use Stop and escalate for the required report shape. Return to Agent orientation if the selected task or policy source changes.