Writing Plans

Execute

→ Have an existing parent plan/spec and a tiny execution slice? → Use Planless Slice Lane.

1. Emit a compact Slice Card: goal, parent plan/spec, files, boundary, verification, stop
2. Update the parent workstream checkpoint/evidence/drift state if persistent state is needed
3. Do not save a new plan for the micro-slice → Mechanical or bounded change with no durable boundary (no new owner, contract, schema, public API, migration, or compat surface)? → Use Planless Slice Lane without a parent document. Emit the compact Slice Card, skip the plan file, and record the change in the commit message. → Have approved spec/requirements for a new workstream or an escalation trigger? → Write implementation plan. Assume engineer has zero context.
4. Scope check: fact/assumption/unknown, baseline, Requirement Ready Check, Ripple Signal Triage, compatibility boundary, dual-track needs
5. File map: what files created/modified, clear boundaries, follow existing patterns
6. Bite-sized tasks (2-5 min each): exact file paths, complete code, exact commands, expected output
7. Self-review: spec coverage, placeholders, type consistency, compatibility, verification, dual-track
8. Save → select and announce the execution route; proceed unless a real authorization or safety boundary requires the user → Plan must answer: problem, baseline, files, compat, verification, risks, retirement. → Escalate from Planless Slice Lane to a durable plan when the slice adds a new owner, contract, schema, public API, architecture boundary, migration, persistence, security/permission, distribution/release surface, or unclear verification boundary.

Writing Plans

Overview

Write comprehensive implementation plans assuming the engineer has zero context for our codebase and questionable taste. Document everything they need to know: which files to touch for each task, code, testing, docs they might need to check, how to test it. Give them the whole plan as bite-sized tasks. DRY. YAGNI. Chosen TDD route. One verified commit per coherent Task or slice.

Assume they are a skilled developer, but know almost nothing about our toolset or problem domain. Assume they don't know good test design very well.

This skill is the canonical planning workflow for multi-step implementation work. Use it to convert approved specs or requirements into plans that are executable, testable, impact-aware, and bounded by compatibility and authority constraints.

Strict RED / GREEN steps belong only to an explicit user/project TDD request or TDD Route: strict. With TDD mode off and no strict route, plan the minimum implementation and proportional regression/verification steps; do not prescribe a TDD cycle from risk alone.

TDD Route Guard

Before task decomposition, every plan that includes implementation work must record:

TDD Route:
- Mode: off | auto
- Decision: strict | light | skipped
- Strict authority: explicit user/project request | recorded auto decision | not applicable
- Strict signals:
- Light eligibility:
- TDD-fit exception:
- Test posture: diagnostic reproduction | post-change regression | strict RED test
- Reason:
- Verification:

In off, record Decision: skipped unless an explicit user/project strict request overrides it. The record makes the boundary reviewable; it does not load test-driven-development. An approved plan or a risk label alone is not strict authority; matching risk is a signal that requires this owner to record recorded auto decision as the authority.

In auto, select strict when any behavior, bugfix, shared/core, contract, persistence, permission, migration, producer/consumer, or meaningful regression signal applies. Select light only when tiny, low-risk, single-owner, no behavior change or strict signal, and an obvious focused check all apply. Absence of an explicit user TDD request is never evidence for light. Unknown risk returns to requirement, debugging, or plan review.

Only Decision: strict with stated strict authority may prescribe Write failing test, Verify RED, GREEN, or REFACTOR as task steps. Otherwise, write the minimum change plus diagnostic reproduction or post-change regression as appropriate. In auto, if the plan lacks a recorded decision, return to route selection before writing implementation tasks; never infer strict during decomposition.

Announce at start: "I'm using the writing-plans skill to create the implementation plan."

Execution context: Reuse the current branch/workspace by default. A branch needs independent history; a worktree needs concurrent checkout, blocking unrelated dirty state, or explicit user/repository authority.

Input: approved requirements, a Spec Brief, or a Design Spec.

Save plans to: docs/aegis/plans/YYYY-MM-DD-<feature-name>.md Plan always goes to plans/ — never to work/. (User preferences for plan location override this default.)

Exception: if an existing parent plan/spec already owns the current tiny execution slice, use Planless Slice Lane. Do not save a new plan. Emit a compact Slice Card in the conversation or the active long-task checkpoint instead:

Slice Card:
- Goal:
- Parent plan/spec:
- Files:
- Boundary:
- Verification:
- Stop:

If docs/aegis/ does not exist and configured Aegis workspace support is available, initialize the target project first:

python <aegis-workspace-helper> init --root <target-project-root>

Project authority overrides workspace initialization. In particular, the Aegis Method Pack repository must not create or ship a live docs/aegis/ workspace; use its formal docs/adr/, docs/current/, and active session plan.

If installed Aegis workspace support is unavailable, initialize the workspace manually:

1. Create docs/aegis/README.md and docs/aegis/INDEX.md
2. Create docs/aegis/BASELINE-GOVERNANCE.md from template
3. If the project has code, create docs/aegis/baseline/YYYY-MM-DD-initial-baseline.md Then save the plan and append to docs/aegis/INDEX.md. Prefer:

python <aegis-workspace-helper> append-index --root <target-project-root> --path docs/aegis/plans/<filename>.md --kind plan --title "<title>"
python <aegis-workspace-helper> check --root <target-project-root>

Scope Check

If the input is a Spec Brief, keep the plan scoped to the pinned what/why/acceptance and do not expand into a formal design unless new architecture, contract, migration, or cross-module uncertainty appears.

Compact output contract before writing the plan: Aegis Visibility, Plan Basis, BaselineUsageDraft, Requirement Ready Check, Files, Compatibility, Change Necessity, Existence Check, Architecture Integrity Lens, Plan Pressure Test, Plan-Time Complexity Check, Execution Readiness View, Tasks, Risks, and Retirement. Expand only where the approved scope, risk, or verification surface requires it.

Aegis Visibility for this workflow states which owner, contract, retirement, compatibility, or verification pressure makes planning useful before execution. Use one natural sentence for ordinary plans; reserve structured trace for audit, debug, release, long-task review, or explicit user request.

Use a compact BaselineUsageDraft whenever the plan depends on specific baseline docs or current-authority refs:

BaselineUsageDraft:
- Required baseline refs:
- Delivered context refs:
- Acknowledged before plan refs:
- Cited in plan refs:
- Missing refs:
- Decision: continue | needs-baseline-readback | needs-verification | pause-for-user | blocked

Delivered context refs is optional host-projected bookkeeping only. It is not authoritative proof that a host injected or the model internally consumed a context payload. The artifact exists to make baseline/context attention drift visible before and during planning.

Use a compact Requirement Ready Check before task decomposition unless the input is already an approved plan/spec whose acceptance boundary is explicit:

Requirement Ready Check:
- Requirement source refs:
- Goals and scope refs:
- User / scenario refs:
- Requirement item refs:
- Acceptance / verification criteria refs:
- Open blocker questions:
- Decision: ready | needs-source | needs-goal-alignment | needs-scenario | needs-acceptance-criteria | needs-clarification | needs-user-decision | blocked

If the decision is not ready, do not create implementation tasks. Return to the requirement/spec owner with the smallest missing evidence or decision. A task intent, conversation, or agent inference can be cited as a candidate source, but it is not durable requirement authority by itself.

Use a compact Change Necessity before task decomposition when the plan would endorse any new source-code path or non-trivial source edits. This is the "should code change at all?" check; it is not a new artifact or a using-aegis hot-path expansion.

This is behavior-triggered, not prompt-triggered. If the plan is about to add any new source-code path or create non-trivial source-edit tasks, expose a natural readback even when the user did not ask for it. A tiny helper, small guard, new branch, fallback, adapter, or owner is not exempt. Example: "Code necessity check: a non-code path is insufficient because ; the minimum change boundary is <owner/files>, so the decision is code-change."

Change Necessity:
- User-visible need:
- No-change / non-code option:
- Why code change is necessary:
- Minimum change boundary:
- Decision: no-change | docs/config-only | code-change | needs-clarification

If the decision is no-change, do not write code-edit tasks. If the decision is docs/config-only, narrow the plan to that surface. If the decision is needs-clarification, return to the requirement/spec owner. If the decision is code-change, carry the minimum boundary into Files, task steps, and verification. Approved requirements do not by themselves prove that a new source-code path is necessary.

Use a compact Existence Check before task decomposition when a plan would add a new owner, skill, artifact, host adapter, fallback, compatibility path, workflow step, or benchmark metric. Use docs/current/AEGIS_MINIMALITY_REFERENCE.md as the reference and keep the check advisory. Do not force it onto plans that only reuse existing owners and surfaces.

Existence Check:
- Proposed new surface:
- Existing owner / reuse candidate:
- Why existing surface is insufficient:
- Creation proof:
- Entropy / retirement impact:
- Decision: reuse-existing | add-with-proof | defer | reject | needs-first-principles-review

If the decision is reuse-existing, write tasks against the existing owner instead of creating a new surface. If the decision is add-with-proof, carry the proof, verification signal, and any retirement trigger into the relevant task.

Use the Architecture Integrity Lens before task decomposition when an executable plan may still encode responsibility overlap, a wrong canonical owner, a caller-side fallback, a stale path carrying real logic, or a missed higher-level owner / contract / source-of-truth simplification. Keep it compact: invariant, canonical owner / contract, responsibility overlap, higher-level simplification, retirement / falsifier, and verdict.

Use a compact Plan Pressure Test before task decomposition:

Plan Pressure Test:
- Owner / contract / retirement:
- Architecture integrity / higher-level path:
- Verification scope:
- Task executability:
- Pressure result: proceed | revise plan | return to design

The pressure test is not an approval gate and should not redesign an approved spec without cause. It exists to catch owner / contract / retirement risk, missing verification, and tasks that are too vague to execute safely.

Render an Execution Readiness View before handing a medium/high, subagent-driven, handoff-prone, long-running, architecture, contract, compatibility, or retirement-sensitive plan to execution. This view is a human-readable projection of existing runtime-ready drafts and plan content. It is not a new JSON artifact type, approval gate, authoritative GateDecision, PolicySnapshot, or completion authority. The view must expose Intent Lock, Scope Fence, and Baseline Lock before any task batch is handed to execution.

Execution Readiness View:
- Intent Lock:
- Scope Fence:
- Baseline Lock:
- Approved Behavior:
- Owner / Contract Constraints:
- Compatibility Boundary:
- Retirement Boundary:
- Task Batches:
- Test Obligations:
- Review Gates:
- Drift / Rewind Rules:
- Evidence Required Before Completion:
- Advisory Boundary: method-pack execution guidance only; not GateDecision, PolicySnapshot, or completion authority

Use existing inputs for the view: TaskIntentDraft, BaselineUsageDraft, ImpactStatementDraft, GateInputPack, the plan's task batches, compatibility / retirement sections, and verification commands. Skip the view for tiny fast-path tasks unless the user asks for an execution handoff readback.

Use a compact Plan-Time Complexity Check before writing task steps when the plan changes maintained source files, core owners, handlers, routers, managers, shared utilities, adapters, or fallback paths:

Use using-aegis/references/complexity-governance.md for the shared artifact classes, pressure signals, and over-budget handling rules.

Complexity Budget:
- Artifact class:
- Target files / artifacts:
- Current pressure:
- Projected post-change pressure:
- Budget result: within-budget | at-risk | over-budget
- Planned governance:

Plan-Time Complexity Check:
- Target files:
- Existing size / shape signals:
- Owner fit:
- Add-in-place risk:
- Better file boundary:
- Recommendation: edit-in-place | extract helper | add owner file | split task | defer refactor

If the projected budget result is over-budget, do not write an atomic task that silently assumes add-in-place growth. Revise the task boundary, add governance work, or explicitly mark the slice as requiring follow-up before implementation begins.

If the spec covers multiple independent subsystems, suggest breaking into separate plans. Before writing tasks, check: fact/assumption/unknown, baseline docs, compatibility boundary, whether dual-track (repair + retirement) applies. If approved requirements or the design carried an ADR signal, preserve the ADR signal, source refs, real alternatives, compatibility boundary, and expected baseline-sync questions for completion so ADR Auto Backfill can run without rediscovering the decision from scratch.

If task decomposition would encode a new owner, duplicate owner, fallback, adapter, compat-only carrier, delete-first question, unverified assumption, or long-term stability claim that the spec did not already settle, use Existence Check first. If the new surface is still justified but the owner, contract, or retirement decision remains risky, use first-principles-review and its Decision Hygiene Review or Architecture Integrity Lens before task decomposition.

When the plan must decide between deleting old internal paths, retaining compat for a proven external boundary, or stopping for persistent-state confirmation, compose anti-entropy-governance. Keep it as a narrow classification and guardrail owner; it does not authorize destructive execution.

Use Planless Slice Lane before writing or saving a plan when all of these are true:

• a parent spec or parent plan already defines the workstream
• the current request is executing or refining one bounded task from that parent
• no new owner, contract, schema, public API, architecture boundary, migration, persistence, security/permission, distribution/release surface, or unclear verification boundary appears
• the slice can be described by a Slice Card

The lane preserves long-task continuity without turning execution bookkeeping into durable planning artifacts.

Aegis Project Workspace

Workspace creation is triggered by the plan save step. See the workspace support rule in using-aegis/SKILL.md for the hard binary rule. If the project already has docs/adr/ or architecture docs, reference them — do not duplicate authority.

File Structure

Map files before defining tasks. Design units with clear boundaries and single responsibilities. Files that change together should live together. Follow existing codebase patterns. Each task should produce self-contained, independently reviewable changes.

For non-trivial project plans, passively read the relevant active language from CONTEXT-MAP.md/CONTEXT.md when present. Use canonical terms consistently in the plan title, tasks, acceptance language, and owner references. If planning discovers a resolved semantic change, ambiguity, or conflict, compose establishing-project-context instead of recording a glossary decision only in the plan. Passive reads alone do not load active modeling.

Required Planning Outputs

Before you leave this workflow, the written plan must make these items answerable:

1. What problem or approved scope this plan is implementing
2. Which baseline docs, ADRs, or requirements shaped the plan
3. Whether the Requirement Ready Check is ready, or which requirement source, scenario, acceptance, clarification, or user decision is still missing
4. Which required baseline refs were explicitly acknowledged before planning and which were actually cited in the plan
5. What files own the change
6. What compatibility boundary must hold
7. Why a code change is necessary, or why the plan is narrowed to no-change, docs/config-only, or clarification
8. Whether any new surface passed an Existence Check or was routed to an existing owner
9. Whether the architecture integrity check found a higher-level owner / contract path before task decomposition
10. What plan-time complexity pressure exists and which edit boundary is safer
11. Whether an Execution Readiness View is needed for this handoff, and if needed, which intent, scope, baseline, compatibility, retirement, testing, review, and drift boundaries it renders
12. What verification proves each major slice
13. What risks, rollback surface, old owner/fallback handling, ADR signal preservation, and baseline-sync signals remain

Bite-Sized Task Granularity

Each step is one action (2-5 minutes):

• Under TDD Route: strict: write the failing test → verify RED → implement minimal code → verify GREEN.
• Otherwise: make the minimum change → run the focused regression or verification that proves it.

Steps are execution units, not Git history units. Commit once after every coherent Task is complete and freshly verified, or after an independently verifiable/revertible long-task slice. Do not commit each 2-5 minute step.

Plan Document Header

Every plan MUST start with: Goal, Architecture, Tech Stack, Baseline/Authority Refs, Compatibility Boundary, TDD Route, Verification. See template in this directory.

Task Structure

Each task: Files (create/modify/test paths), Why (user/business value), Change Necessity (why source edits are needed and the minimum boundary), Impact/Compatibility, Verification (exact commands), then steps matching the TDD route. Strict routes use Write test → Verify RED → Minimal code → Verify GREEN; off, light, and skipped routes use the minimum change plus proportional regression/verification. The executing coordinator captures TaskStartSnapshot before the first write and creates one scoped commit only after the whole Task passes review and verification. Every step must include complete code and exact commands.

For bug fixes, refactors, contract changes, or governance cleanup, add Repair Track (root cause, canonical owner, minimal sufficient stable repair, compat boundary, verification) and Retirement Track (old owner/fallback, active status, keep reason or deletion trigger) inside the relevant task. If Ripple Signal Triage fired, include the affected downstream consumers and expanded verification path in the same task.

No Placeholders

Never write: "TBD", "TODO", "implement later", "fill in details", "Add appropriate error handling", "Write tests for the above" without actual test code, "Similar to Task N" without repeating code. Every step must contain complete, copy-paste-ready content.

Self-Review

Check plan against spec: 1) Spec coverage — can you point to a task for each requirement? 2) Placeholder scan — any TBD/TODO/vague instructions? 3) Type consistency — do signatures match across tasks? 4) Compatibility — invariants, non-goals, stable interfaces marked? 5) Change necessity — any code-edit task states why no-change or docs/config-only is insufficient and names the minimum boundary? 6) Existence check — any new owner, artifact, adapter, fallback, workflow step, or benchmark metric has proof and a reuse decision? 7) Plan-time complexity and minimality — lowest-entropy owner/file boundary that fixes the bug class, not just the smallest textual diff? 8) Architecture integrity — any higher-level owner / contract / source-of-truth simplification skipped? 9) Verification — exact commands? 10) Dual-track, decision hygiene, and ADR/baseline-sync signals preserved where needed?

Fix issues inline. Re-review is not needed — just fix and move on.

Execution Handoff

After saving the plan, render the Execution Readiness View when the handoff criteria above apply. The agent owns the execution-route decision; do not ask the user to choose merely because both routes are viable.

Select subagent-driven when subagents are available, the plan has genuinely independent tasks with bounded ownership, and the review/context benefit justifies the coordination. Otherwise select inline. Lack or denial of subagent support falls back to inline execution instead of blocking the task. A dirty workspace alone does not select either route; apply the Git ownership, overlap, and isolation rules separately.

Reconsider delegation after the minimum baseline is understood, not from a keyword alone. Useful event signals are: a clearly isolated discovery slice, conflicting evidence that can be checked independently, a high-risk change that benefits from a read-only review, or a blocked/recovered slice where the next attempt will use materially different context. These signals are advisory; unknown dependencies, shared transactions/resources, stale state, untrusted inputs, unclear host capability, or borderline economics remain inline routes.

Ask the user only when route selection crosses an unresolved authorization, privacy, paid-resource, external-action, or irreversible boundary; changes the approved scope or acceptance contract; or no safe route can preserve existing workspace ownership. When no such boundary exists, proceed immediately.

State the decision compactly:

Execution Route:
- Decision: subagent-driven | inline
- Evidence:
- Fallback:
- User confirmation required: no | yes — <specific unresolved boundary>

If subagent-driven:

• REQUIRED SUB-SKILL: Use aegis:subagent-driven-development
• Fresh subagent per task + two-stage review

If inline:

• REQUIRED SUB-SKILL: Use aegis:executing-plans
• Batch execution with checkpoints for review

Planning Boundaries

• A plan can define implementation slices, verification, rollback surface, and retirement expectations
• Execution Readiness View can make implementation start conditions, verification obligations, and drift / rewind rules visible before execution
• A plan cannot grant authoritative completion
• A plan should prepare runtime-ready execution, not pretend to be runtime authority