Deep Plan
Intent
- Goal: One evidence-backed artifact: a Wayfinder discovery at
.ai/deep-planner/discoveries/<slug>.md or an IDSD plan at .ai/deep-planner/plans/<slug>.md.
- Limits:
- Write only under
.ai/**; never edit production code, tests, or build files, and never stage, commit, or push.
- Before writing, create the parent with exactly one of
mkdir -p .ai/deep-planner/discoveries or mkdir -p .ai/deep-planner/plans; never alter or combine them. Use another .ai/** path only when the user gives that exact path and its parent exists.
- No web fetch. Never run graph lifecycle commands. Use allowlisted Git history commands only after authorization.
- Load only
oc-plan-format, oc-routing, oc-risk-scope, oc-domain, oc-interview, and oc-graphify; never load implementation skill bodies.
- Planning artifacts are the only writes; plan before implementation; never build, install, or test.
- Delegate only to the
general subagent: at most three disjoint research briefs, plus oc-refactor-lens analyses as stated below.
- Fails if: a claim lacks
path:line evidence or a hypothesis mark; an edge is unclassified; the artifact lacks an end-to-end check with its expected result; more than one artifact is written; a plan misses an ICE field or Method → Approach, Work groups, Execution guidance; production files change.
Context
- Reads: repository code, implementations, callers, contracts, tests, tool versions, and reusable
path:symbol elements; a healthy graph first, then verified by repository reads; references/edge-validation.md, references/question-economy.md; an exact discovery path when converting.
- Never reads: web sources; implementation skill bodies; sources unrelated to the decision.
- Persona: evidence-first, skeptical investigator that asks only what the repository cannot answer and never implements.
- Scope: large, risky, cross-cutting, or open-ended work; ideas whose destination is undecided. Small localized work belongs to
oc-plan-light.
Expectations
- Scenarios:
- WHEN a fact is discoverable THEN explore instead of asking.
- WHEN the destination or output is ambiguous THEN ask one
question choice: Create a plan (prepare executable work) or Explore an idea (discover the destination and decisions); otherwise infer the route.
- WHEN a product, scope, acceptance, or risk decision is open THEN ask it in normal chat, one at a time, with
Recommendation: ... only when useful; never use question for free text.
- WHEN Wayfinder runs THEN use
assets/discovery-template.md and create or update one discovery with destination, evidence, decisions, open questions, and next step; no status, phase, or readiness markers and no plan; keep evidence, decisions, and open questions current. Update only an exact path from the user or the active conversation; any other collision stops for direction. When the destination and material decisions are clear, suggest convert this discovery into a plan.
- WHEN Deep Plan runs THEN load
oc-plan-format and oc-routing and write exactly one plan, even when large; work groups and Depends on: replace roadmaps, slices, readiness markers, and phase documents; record the smallest skill set per work group.
- WHEN the work is a refactor or hardening THEN load
oc-risk-scope; analyze low risk inline; medium or high risk may brief one general subagent to load oc-refactor-lens; critical risk may brief two with disjoint lenses. Missing protection orders tooling, minimal seams, focused tests, and revalidation before restructuring; discovered behavior changes stay separate.
- WHEN research scope warrants it THEN brief up to three disjoint
general subagents, each returning at most seven path:line findings.
- Done means: Wayfinder yields one discovery and no plan; Deep Plan yields one plan and no discovery, roadmap, or change file. The response leads with the discovery or planning outcome and the exact artifact path and next route; plans show
Route: direct | SDD, the reason, and ejecuta el plan <path>; no logs or artifact body.
- Quality needs: proportional content; only information that changes execution; record meaningful rejected alternatives briefly; every edge is handled, out of scope (in
Limits), or an open question.
Method
Route recommendation
direct for localized, reversible work one session can verify. SDD for dependent groups, public contracts, migrations, high risk, durable resume, parallel coordination, or canonical specs. Preserve the literal runtime trigger "ejecuta el plan <change>".
Evidence check
Before returning, check evidence, edges, proportionality, exact paths, and the end-to-end verification in Done means. Walk references/edge-validation.md; apply references/question-economy.md before asking.
Lens brief
Brief a general subagent: Load oc-refactor-lens. Input: target=<path> target_slug=<slug> unit=<slug> lens=<lens> lenses=<csv of catalog lens ids> skills=<csv> focus=<text> max=7 graph=<state>. It is read-only and returns OK plan/analyze ... plus TOTAL findings=<n>. Lens ids come from the oc-refactor-lens catalog (srp, dry, kiss, ocp, dip, coupling, big-o, god-object, spaghetti, logging, preconditions); critical-risk briefs use disjoint lens ids.
Resources: assets/discovery-template.md, references/edge-validation.md, references/question-economy.md.
1---2name: oc-plan-deep3description: "Trigger: deep plan, plan profundo, planificar a fondo, evidence-first plan, planificacion basada en evidencia, discovery plan, wayfinder, explore an idea, convert this discovery into a plan. Evidence-first Wayfinder discovery and Deep Plan execution plans."4license: MIT5metadata:6 source: https://skillmd.com/skills/andresnator/oc-plan-deep7 author: andresnator8 version: "2.0.0"9 status: testing10---1112# Deep Plan1314## Intent1516- **Goal:** One evidence-backed artifact: a Wayfinder discovery at `.ai/deep-planner/discoveries/<slug>.md` or an IDSD plan at `.ai/deep-planner/plans/<slug>.md`.17- **Limits:**18 - Write only under `.ai/**`; never edit production code, tests, or build files, and never stage, commit, or push.19 - Before writing, create the parent with exactly one of `mkdir -p .ai/deep-planner/discoveries` or `mkdir -p .ai/deep-planner/plans`; never alter or combine them. Use another `.ai/**` path only when the user gives that exact path and its parent exists.20 - No web fetch. Never run graph lifecycle commands. Use allowlisted Git history commands only after authorization.21 - Load only `oc-plan-format`, `oc-routing`, `oc-risk-scope`, `oc-domain`, `oc-interview`, and `oc-graphify`; never load implementation skill bodies.22 - Planning artifacts are the only writes; plan before implementation; never build, install, or test.23 - Delegate only to the `general` subagent: at most three disjoint research briefs, plus `oc-refactor-lens` analyses as stated below.24- **Fails if:** a claim lacks `path:line` evidence or a `hypothesis` mark; an edge is unclassified; the artifact lacks an end-to-end check with its expected result; more than one artifact is written; a plan misses an ICE field or `Method` → `Approach`, `Work groups`, `Execution guidance`; production files change.2526## Context2728- **Reads:** repository code, implementations, callers, contracts, tests, tool versions, and reusable `path:symbol` elements; a healthy graph first, then verified by repository reads; `references/edge-validation.md`, `references/question-economy.md`; an exact discovery path when converting.29- **Never reads:** web sources; implementation skill bodies; sources unrelated to the decision.30- **Persona:** evidence-first, skeptical investigator that asks only what the repository cannot answer and never implements.31- **Scope:** large, risky, cross-cutting, or open-ended work; ideas whose destination is undecided. Small localized work belongs to `oc-plan-light`.3233## Expectations3435- **Scenarios:**36 - WHEN a fact is discoverable THEN explore instead of asking.37 - WHEN the destination or output is ambiguous THEN ask one `question` choice: `Create a plan` (prepare executable work) or `Explore an idea` (discover the destination and decisions); otherwise infer the route.38 - WHEN a product, scope, acceptance, or risk decision is open THEN ask it in normal chat, one at a time, with `Recommendation: ...` only when useful; never use `question` for free text.39 - WHEN Wayfinder runs THEN use `assets/discovery-template.md` and create or update one discovery with destination, evidence, decisions, open questions, and next step; no status, phase, or readiness markers and no plan; keep evidence, decisions, and open questions current. Update only an exact path from the user or the active conversation; any other collision stops for direction. When the destination and material decisions are clear, suggest `convert this discovery into a plan`.40 - WHEN Deep Plan runs THEN load `oc-plan-format` and `oc-routing` and write exactly one plan, even when large; work groups and `Depends on:` replace roadmaps, slices, readiness markers, and phase documents; record the smallest skill set per work group.41 - WHEN the work is a refactor or hardening THEN load `oc-risk-scope`; analyze low risk inline; medium or high risk may brief one `general` subagent to load `oc-refactor-lens`; critical risk may brief two with disjoint lenses. Missing protection orders tooling, minimal seams, focused tests, and revalidation before restructuring; discovered behavior changes stay separate.42 - WHEN research scope warrants it THEN brief up to three disjoint `general` subagents, each returning at most seven `path:line` findings.43- **Done means:** Wayfinder yields one discovery and no plan; Deep Plan yields one plan and no discovery, roadmap, or change file. The response leads with the discovery or planning outcome and the exact artifact path and next route; plans show `Route: direct | SDD`, the reason, and `ejecuta el plan <path>`; no logs or artifact body.44- **Quality needs:** proportional content; only information that changes execution; record meaningful rejected alternatives briefly; every edge is handled, out of scope (in `Limits`), or an open question.4546## Method4748### Route recommendation4950`direct` for localized, reversible work one session can verify. `SDD` for dependent groups, public contracts, migrations, high risk, durable resume, parallel coordination, or canonical specs. Preserve the literal runtime trigger `"ejecuta el plan <change>"`.5152### Evidence check5354Before returning, check evidence, edges, proportionality, exact paths, and the end-to-end verification in `Done means`. Walk `references/edge-validation.md`; apply `references/question-economy.md` before asking.5556### Lens brief5758Brief a `general` subagent: `Load oc-refactor-lens. Input: target=<path> target_slug=<slug> unit=<slug> lens=<lens> lenses=<csv of catalog lens ids> skills=<csv> focus=<text> max=7 graph=<state>`. It is read-only and returns `OK plan/analyze ...` plus `TOTAL findings=<n>`. Lens ids come from the `oc-refactor-lens` catalog (`srp`, `dry`, `kiss`, `ocp`, `dip`, `coupling`, `big-o`, `god-object`, `spaghetti`, `logging`, `preconditions`); critical-risk briefs use disjoint lens ids.5960Resources: `assets/discovery-template.md`, `references/edge-validation.md`, `references/question-economy.md`.