Cairn planning
Skill eventually-consistent-code/cairn2/skills/cairn-planning
Use when a repo has cairn.json and the cairn MCP tools (plan_*, issue_*) are available — the cairn 2.0 planning lifecycle. Owns the depth dial (quick/standard/deep), the model-routing rubric for agent fan-out, artifact conventions under .cairn/plans/, and the precedence rule (git plan docs win over tracker issue text).From its SKILL.md
npx -y skills add eventually-consistent-code/cairn2 --skill cairn-planningAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
4.6 KB, as published. Nobody here has run it
cairn planning (lifecycle policy)
The server owns mechanism (scaffolding, frontmatter, mirroring, drift math). This skill owns judgment: how deep to plan, which models to fan out, what goes in which artifact, and what "verified" means.
Activation gate
Apply when cairn.json exists in the repo AND the plan_* MCP tools are
available. The tracker named in cairn.json is the single source of truth for
work items; git owns plans and memory.
Artifacts — what goes where
| File | Owns | Written by |
|---|---|---|
| PROJECT.md | vision, requirements | /cairn:new interview |
| roadmap.md | phase table + status | /cairn:new, updated at verify/ship |
| phases/NN/CONTEXT.md | locked decisions | planning discussion |
| phases/NN/RESEARCH.md | deep-mode research brief | research fan-out |
| SURVEY.md | project-wide research brief (survey markers) | /cairn:survey fan-out |
| phases/NN/PLAN.md | task breakdown + issues: frontmatter | /cairn:plan |
| phases/NN/VERIFICATION.md | goal-backward check results; its EXISTENCE marks the phase verified | /cairn:verify |
Frontmatter is the constrained flat form only (key: value, key: [a, b]).
Update issues: through plan_issues_set, not hand-editing, so validation runs.
Depth dial
Resolution order: command flag > PLAN.md depth: frontmatter > cairn.json
default > standard.
| quick | standard | deep | |
|---|---|---|---|
| research | none | 1 research subagent | parallel fan-out, multi-angle |
| plan | draft tasks directly | PLAN.md + CONTEXT.md | + plan-checker agent pass |
| verify | tests pass | + tracker cross-check (plan_drift) | + adversarial verification, VERIFICATION.md rigor |
MVP shaping (plan --mvp)
First tasks form ONE thin vertical slice exercising every layer end-to-end (walking skeleton); depth and breadth come only after the slice stands. A slice that can't demo is not a slice.
TDD eligibility (plan --tdd)
Eligible: behavior-testable code — logic, APIs, parsers, state machines.
Not eligible: config, docs, scaffolding, pure styling, generated code.
Eligibility is per task, decided at plan time, stored in PLAN.md tdd:
frontmatter via plan_meta_set; work-time enforcement is the RED/GREEN
commit pair in the ledger, checked at verify.
Model routing (research fan-out)
cairn.json → agents.model: auto | inherit | haiku | sonnet | opus.
inherit = session model everywhere; explicit value pins everything; auto
routes per work class:
| work class | signals | routed model |
|---|---|---|
| mechanical | enumerate/locate/scrape | small/fast (haiku-tier) |
| synthesis | research briefs, pattern analysis | session model (sonnet-tier) |
| judgment gates | plan-checker, adversarial verify, architecture trade-offs | strongest available (opus-tier) |
Blast radius rules: output that gates a lifecycle transition (verify/ship) routes UP, never down. Downgrade only mechanical work. Uncertain → inherit.
The rubric governs ALL research fan-out — scout, survey, and
plan --deep — not only deep mode. Research verbs dispatch one subagent
per pending topic, in parallel, classifying each TOPIC by work class to
pick its model; the main thread writes results and flips markers as each
agent completes.
Lifecycle
new → plan N → work N → verify N → ship. Each command's file defines its steps; this skill's conventions bind them all:
- Claim before work:
issue_update(id, state: "in_progress")+context_setbefore touching code for an issue. - Close only on verified done: tests passing, behavior confirmed. Never close to make a gate pass.
- Precedence: on conflict between a phase's CONTEXT.md/PLAN.md and tracker
issue text, the GIT DOC WINS — update the issue (
issue_update) with a dated note; never silently follow the stale issue. - Drift is a stop signal:
plan_driftflags mean reconcile before proceeding (missing → recreate; closed-unverified → verify the phase or reopen the issue). - Assignee courtesy: never claim an issue assigned to someone else without saying so.
Failure honesty
Report gates truthfully: a failed verify is a failed verify. Never write VERIFICATION.md for a phase that didn't pass — its existence is a machine-read signal (drift treats closed issues in verified phases as normal).