Coding Safety Plan Profile
Purpose
Define the minimum planning profile for a bounded coding Work Slice before implementation starts.
This profile specializes the existing Work Slice, Intent-to-Evidence, Agent Harness and readiness-gate contracts for code changes. It does not create a new lifecycle, runtime, command, policy engine or autonomous coding flow.
When to use
Use this profile for Standard or High-Assurance coding slices when the work changes source code,
configuration, tests, build behavior, schemas, permissions, runtime behavior or documentation that
could affect how a system is used.
Lite coding work may use a compact version when the change is obviously reversible, local and low risk.
Required plan fields
A coding plan MUST declare:
coding_safety_plan:
slice_id: string
planning_depth: Lite | Standard | High-Assurance
base_state:
branch: string
starting_commit: string
dirty_state: clean | tracked_changes | untracked_present | unknown
protected_paths: []
intent_refs:
intent_to_evidence_ref: not_needed | string
work_slice_ref: string
scope:
in_scope: []
out_of_scope: []
assumptions: []
human_decisions_needed: []
execution_boundary:
allowed_paths: []
blocked_paths: []
allowed_commands: []
blocked_actions: []
expected_changes:
files_or_areas: []
behavior_change: none | docs_only | test_only | code | config | schema | runtime | unknown
user_visible_change: yes | no | unknown
verification:
commands: []
expected_results: []
manual_checks: []
unavailable_checks: []
drift_control:
drift_signals: []
check_cadence: before_commit | before_pr | per_checkpoint | manual
response: tighten_scope | ask_human | stop | split_slice
rollback:
strategy: git_revert | restore_files | feature_flag | migration_rollback | not_applicable | unavailable
rollback_refs: []
irreversible_risks: []
stop_conditions: []
human_review:
required: true | false
reviewer_profile: non_engineer | engineer | domain_expert | security | owner
review_focus: []
Required invariants
- No coding without a base state. The plan must identify branch, starting commit and dirty-state posture. Untracked files may exist, but their treatment must be explicit.
- Scope is narrower than intent. The plan must list both
in_scopeandout_of_scopeso the implementer can detect drift. - Verification names expected results. Commands alone are insufficient; the reviewer must know what passing means.
- Rollback or stop is explicit. If rollback is
unavailable, a human review or stop condition must explain why work can safely proceed. - Human merge remains human. Passing tests, AI confidence, or an implementation agent cannot authorize merge or deployment.
- No hidden runtime expansion. The plan cannot introduce new runtime, backend, schema, collector, dependency, CLI, automation or policy-engine scope unless that scope is explicitly listed and authorized by the Work Slice.
Non-engineer review checklist
A non-engineer reviewer should be able to answer:
- What will change?
- What will explicitly not change?
- How will we know it worked?
- What evidence should exist before merge?
- What would make us stop or ask for help?
- How can the change be undone or contained?
- Which decisions still belong to a human?
If those answers are not visible, the slice is not ready for execution.
Relationship to Adaptive Skills
An implementation-planning-advisor skill is not promoted by this contract. A future skill proposal
must prove a recurring gap not covered by existing feature-planning, testing, debugging,
architecture-review, Lean Implementation and independent validation guidance.