| name | planner |
| description | Anton planner skill — author, update, and close multi-session initiatives (migrations, rollouts, long-running refactors) in `context/plans/`. Use when starting a multi-session initiative, tracking next steps on in-flight work, migrating a memory entry to a durable plan, closing a completed initiative, or reviewing what's open. Plans live at `context/plans/NNNN-kebab-slug.md` and are mutable — they capture execution state (what's next, what's blocked, log of decisions made during work) while ADRs capture immutable decisions (why). The active-plan index is built by scanning plan files directly and injected into every Codex session by `.Codex/hooks/inject_plans_index.py`. Keywords — plan, planner, initiative, track work, multi-session, next steps, checklist, migration plan, rollout plan, roadmap, in-flight, blocker, close plan, review-by, exit plan, timebox, memory-to-plan handoff. |
| allowed-tools | Read, Write, Edit, Glob, Grep, Bash |
Planner
Anton-local skill for the in-repo plan lifecycle. Owns four workflows: new, list, update, close. Plans live at context/plans/NNNN-kebab-slug.md and are mutable — the body is edited as work progresses, tasks get checked off, the log grows.
Why this skill exists
Anton already has two complementary persistence layers:
- ADRs (
context/adrs/) — immutable decisions. Answer why.
- Auto-memory — ephemeral cross-session hints. Answer what did I just learn.
- Git log — immutable commits. Answer what changed.
But there's a missing layer: the multi-session initiative — "adopt Longhorn", "migrate Tailscale", "roll out OpenTelemetry". These span many commits and several sessions. Memory is too ephemeral (decays as new entries push it out); git log is too coarse (no sense of what's next); ADRs are too rigid (decisions, not execution state).
Plans fill that gap. They're in-repo (survive across sessions without relying on memory), mutable (task lists churn), discoverable at session start (via the SessionStart hook), and trunk-compatible (just files edited on main — no long-lived branches).
The skill is anton-local. Format is intentionally compatible with the global context-repo skill but this skill never delegates to that one.
Hard rules
- Plans are mutable — but carefully. Task lists, logs, and status flip freely. Goal and Acceptance criteria should change only when the scope itself changes; substantial scope changes should
close the plan as abandoned and new a replacement rather than quietly mutating the intent.
- Numbers are never reused. Next number is
max(existing) + 1, zero-padded to four digits.
- Active statuses are
Draft, In-progress, Blocked. Terminal statuses are Done, Abandoned. Only active plans are surfaced in the SessionStart index.
- Close on done. Plans should not drift into a permanent In-progress state. If nothing has changed in >30 days, mark
Blocked with a reason, or close as abandoned.
- Plans point at ground truth; they do not duplicate it. Renovate dashboards, Flux kustomization status, memory entries — reference them, don't mirror them.
- Plans do not replace ADRs. If a plan surfaces a durable architectural decision during execution, hand off to the
adr skill to record it. Link the new ADR from the plan's References section.
References:
- Field reference, status enum, and immutability rules → conventions
- Canonical template for new plans → template
Workflow — new
Triggered by:
- The user says "open a plan for X" / "start tracking X" /
/planner new "<title>"
- The user wants to promote an in-flight memory entry into a durable plan
- The user commits to a multi-session migration, rollout, or refactor
Before running the full workflow, ask: does this really need a plan? If the work is a single commit, a Renovate PR, or something that will finish in this session, say so and decline. Plans are for work that spans sessions.
Steps
-
Allocate the next number. Glob('context/plans/0[0-9][0-9][0-9]-*.md'). Sort. Next number is max + 1, zero-padded to 4 digits. Never reuse.
-
Generate the slug. Lowercase, kebab-case, drop articles/prepositions, 4–6 meaningful words. Example: "Migrate Tailscale to the new tailnet" → migrate-tailscale-new-tailnet → 0001-migrate-tailscale-new-tailnet.md.
-
Interview the user. Ask only for what isn't already known:
- Title (one short sentence, imperative or descriptive)
- Goal (one paragraph — what "done" looks like in plain language)
- Acceptance criteria (2–5 outcome bullets, not tasks — conditions that must be true at close)
- Initial task list (3–10 concrete next steps; more will accrue during execution)
affects — same category list as ADRs (storage, observability, networking, security, compute, all, etc.)
intent — concrete-need / learning / unknown
related-adrs — list of ADR NNNNs that scope this plan (usually the decision that triggered the work)
review-by — required for intent: learning, optional otherwise. Date to re-evaluate.
- Initial log entry — one-line note of why now, to anchor the history.
-
Write the plan file at context/plans/NNNN-slug.md using the shape from template. Set status: In-progress (or Draft if the user is still shaping the goal), opened: YYYY-MM-DD to today.
-
If this replaces an in-flight memory entry, tell the user to remove the memory file (auto-memory is not self-cleaning). Cite the exact path. The plan supersedes the memory.
-
Confirm to the user with the file path, the plan number, and a reminder that the SessionStart hook will pick it up on the next session.
Workflow — list
Triggered by:
/planner list — default, shows active plans only
/planner list --all — include terminal (Done / Abandoned)
/planner list --status <X> or --affects <X> — filter
Steps
Glob('context/plans/0[0-9][0-9][0-9]-*.md').
- Read frontmatter of each. Extract
status, opened, closed, affects, intent, related-adrs, review-by, and the H1 title.
- Apply filters:
- Default: keep only
Draft / In-progress / Blocked
--all: no filter
--status X: case-insensitive exact match
--affects X: substring match on the affects: field
- Format as a markdown table sorted by NNNN. Columns: #, Title, Status, Opened, Review-by, Related ADRs.
- If a plan has
review-by ≤ today and status is still active, flag it with ⚠ in the Status column.
Workflow — update
Triggered by:
/planner update <NNNN> — the user wants to edit an existing plan
- The user says "check off X on plan NNNN" / "log Y on plan NNNN" / "mark NNNN blocked"
This is the most-used workflow. The skill doesn't wrap every possible edit — it knows the plan's shape and uses the ordinary Edit tool to make targeted changes.
Common edit recipes
- Check off a task. Find the task line in the Tasks section and replace
- [ ] text with - [x] text (preserve exact indentation). Then append a one-line entry to the Log with today's date.
- Add a task. Append to the Tasks section under the appropriate heading (or at the end). Keep the task imperative, concrete, and verifiable.
- Log a decision or surprise. Append to the Log section. Format:
- YYYY-MM-DD: <one sentence>. The log is append-only; never rewrite past entries.
- Flip to
Blocked. Edit the status: line in frontmatter and append a Log entry naming the blocker. Don't lose visibility — blocked plans should name a concrete unblock condition in the Log.
- Un-block. Flip
status: back to In-progress and log the resolution.
- Refine acceptance criteria. Edit the Acceptance criteria section and log the refinement so the history explains the change. If the refinement is large enough to change what "done" means, consider
close + new instead.
Steps
- Read the plan.
Read('context/plans/NNNN-*.md'). Confirm it exists and isn't terminal (Done / Abandoned). Terminal plans should not be edited — log a new plan instead if the work is resuming.
- Make the edit(s) via
Edit with precise old/new strings. One edit per change; keep diffs small.
- Append a Log entry for anything non-trivial (checking off a single task may skip the log if it's a small step; status changes and new tasks should always log).
- Confirm to the user with a one-line summary of what changed.
Workflow — close
Triggered by:
/planner close <NNNN> done "<closing note>"
/planner close <NNNN> abandoned "<reason>"
Steps
- Read the plan. Confirm it's not already terminal.
- Verify acceptance criteria (for
done only). Walk through the Acceptance criteria checklist. If any are unchecked, ask the user explicitly: "criterion X is unmet — close as done anyway, or re-scope?" Don't close with false-done signals.
- Flip the frontmatter:
status: to Done or Abandoned, closed: to today's date.
- Append a final Log entry with the closing note.
- Offer an ADR handoff (only on
Done). Ask: "did this plan surface a durable architectural decision worth recording as an ADR?" If yes, invoke the adr skill with a pre-populated title drawn from the plan's Goal. If no, move on.
- Confirm to the user with the file path and the terminal status. The SessionStart hook will drop the plan from the next session's index automatically.
What this skill does NOT do
- Does not write Flux manifests — that is
add-flux-app.
- Does not make architectural decisions — that is
adr.
- Does not auto-close stale plans. A
review-by breach is flagged in list; the operator still closes manually.
- Does not duplicate cluster state, Renovate state, or git log. Plans reference these sources; they do not mirror them.
- Does not push, commit, or tag — the user owns git operations.
- Does not delete plans. Terminal plans are append-only history.
- Does not replace auto-memory for ephemeral cross-session hints. Memory stays; plans capture the durable "what's next" layer.
Related skills and agents
adr — hand off here on close done if a durable decision emerged. Link the new ADR back from the plan's References section.
cluster-intake-gatekeeper — may hand off to planner (as well as adr) when an Add (learning) verdict comes with a multi-session rollout plan that needs tracking beyond the ADR's own checklist.
debug-flux-reconciliation / anton-cluster-health — reactive; a plan is the forward-looking companion to these. If a reactive session surfaces follow-up work that spans sessions, open a plan.
When a plan is NOT warranted
- Single-commit fixes
- Renovate PR triage — use
anton-upgrade-audit instead; the Renovate dashboard is ground truth
- Read-only investigations with no committed output
- Work that will finish in the current session
- "I'll probably want to do X someday" — not a plan, that's speculation. Revisit when the work is actually starting.
Rule of thumb: if I walk away for two weeks, would I need this file to pick up where I left off? If yes, open a plan. If no, skip it.
Anti-patterns
- Opening a plan for work that fits in one commit. Pure overhead. Skip it.
- Duplicating memory into a plan, or a plan into memory. Pick one home per piece of state. The plan is canonical for execution state; memory is canonical for ephemeral hints. When in doubt, promote to plan and delete the memory.
- Editing the Goal to reflect a pivoted scope. That erases history.
close abandoned + new instead.
- Leaving a plan
In-progress for >30 days with no Log entries. Either the work died (close as abandoned) or it's actually blocked (flip to Blocked and name the blocker). Silent drift is the failure mode this system exists to prevent.
- Letting the active-plan list grow past ~5. If you have more concurrent initiatives than that, the bottleneck is focus, not tooling.