devops-tune-rethink

name	devops-tune-rethink
version	0.2.0
description	Strategic reset for stuck development: iterations circle without reaching the goal, or a named part of the app (or the whole app) is functionally and visually not where it should be and incremental fixes stopped helping. Re-derives the original goals from repo + live-app evidence, validates them in a short question round (incl. jointly calibrating how much may be torn down), generates genuinely fresh approaches via code-blind lens agents, reconciles them against the codebase, lets the user decide on a devops-concept page, then hands the chosen approach to devops-run-autonomous for implementation with test-plan verification (deep-knowledge/test-plan.md). Triggers on: "festgefahren", "stuck", "unstuck", "wir drehen uns im Kreis", "neu denken", "rethink", "frischer Ansatz", "fresh approach", "komplett neu denken", "das führt zu nichts". Do NOT trigger for: debugging/errors (use /devops-fix), incremental consistency or UI passes (/devops-tune-harden, /devops-tune-polish), or normal feature work.
argument-hint	[app or section that is stuck, e.g. 'the onboarding flow']
allowed-tools	Read, Write, Edit, Glob, Grep, Bash, Agent, AskUserQuestion, CronCreate, CronDelete, mcp__Claude_Preview__*, mcp__plugin_playwright_playwright__*, mcp__plugin_devops_dotclaude-completion__*

Tune Rethink — Strategic Reset

Break out of a stuck development loop for $ARGUMENTS: rebuild the goal picture, think fresh without being anchored by the current implementation, then reconcile, decide, and implement autonomously.

Why this skill exists: when the same area has been iterated on many times without reaching the goal, the current implementation itself anchors every new attempt. This skill enforces un-anchored ideation structurally — the ideation agents cannot read the code — and re-introduces existing knowledge only in a second, explicit step.

Step 0 — Load Extensions

Check for optional overrides. Use Glob to verify each path exists before reading. Skip missing files silently.

1. Global: ~/.claude/skills/tune-rethink/SKILL.md + reference.md
2. Project: {project}/.claude/skills/tune-rethink/SKILL.md + reference.md
3. Merge order: project > global > plugin defaults

Step 1 — Intake & Scope

$ARGUMENTS names the target — the whole app or a section of it. The rethink is holistic for the named target (not a micro-polish of one control).

• If $ARGUMENTS is empty: ask ONE question (AskUserQuestion) to name the target before anything else.
• State the scope boundary explicitly: "In scope: X including Y. Out of scope: Z." This boundary travels with the brief through every later step.

Step 2 — Evidence Mining (self-served)

Build your own picture BEFORE asking the user anything. Two source classes:

1. Repo evidence: git log for the target area, open/closed issues, existing concept pages, docs, CHANGELOG, CLAUDE.md. Reconstruct: what was the intended goal, what was already tried, where did iterations circle.
2. Live evidence: look at the running app — this is the only honest source for "it doesn't look/feel right". Follow {PLUGIN_ROOT}/deep-knowledge/test-autonomy.md and preview-testing.md: start the app (Claude Preview preferred), view the target at the profile viewports (responsive-testing.md), click through its core flows, capture snapshots.

Output: a hypothesis dossier, explicitly marked as hypotheses — "this was the goal, this is today's state, here is the gap, this was tried".

Fallbacks: no repo evidence → the question round carries more weight (more questions allowed). App not startable → static evidence only; note "no live look possible" in the dossier.

Step 3 — Question Round (validate + calibrate)

AskUserQuestion, ONE question per round, ~3–5 rounds, multiple-choice preferred. Mandatory contents, in this order:

1. Confirm/correct the hypothesis dossier (goal, gap, tried-before).
2. Success criteria: what does good look like for this target?
3. Biggest frustration / most-missed outcome.
4. Hard no-gos: stack, data model, deadlines — non-negotiables.
5. Demolition corridor — calibrated together, never assumed: how much may be torn down at the large scale (nothing / the named section / everything). The corridor bounds what gets seriously developed; it is agreed jointly so radical options don't surprise the user later.

Output: the Rethink Brief — ONE completely code-free document containing scope, goal, gap, success criteria, no-gos, corridor. Persist it to .claude/rethink/<YYYY-MM-DD>-<slug>/brief.md. The brief is the ONLY project context the fresh-phase agents receive, so it must stand alone.

Step 4 — Fresh Phase (code-blind fan-out)

Spawn three devops:rethinker agents in parallel (one message, three Agent calls). Each prompt = the full Rethink Brief + ONE lens block from deep-knowledge/lenses.md:

Lens	Owns the question
product-value	Does the approach achieve the purpose — simplest thing that reaches the goal?
ux-design	Structure, flow, layout idea, visual direction
enduser-feel	Journey, friction, "does it feel good?"

For backend-only targets (no UI surface in scope), replace enduser-feel with the architecture lens.

The rethinker agent has no file tools — code-blindness is enforced by its tool set, not by a polite request. Do NOT paste code, file paths, or implementation details into its prompt; that would defeat the isolation.

Naming per deep-knowledge/agent-conventions.md: [role:rethinker · Ideation] <lens> approach for <target>.

Each agent returns one RETHINK_APPROACH block (format in lenses.md). If an agent dies, continue with the survivors (minimum 1) and note the gap on the concept page.

Step 5 — Reconciliation (the second step)

Back in the main context, now WITH codebase access, integrate what is already known:

• Evaluate each approach against reality: migration path, what survives, effort, risks, which success criteria it serves.
• Merge near-duplicate approaches.
• Label every approach with its blast radius relative to the corridor:
  • in-corridor — eligible for implementation.
  • over-corridor — shown as a flagged outlier for information; NEVER implemented without the user explicitly widening the corridor first.

Persist the reconciled set to .claude/rethink/<date>-<slug>/approaches.md.

Step 6 — Concept Page

Invoke devops-concept (decision template). Each approach is a variant with: pros/cons, blast-radius label, effort estimate, migration sketch, and which success criteria it serves. The page offers two decision actions:

• Iterate — feedback flows back: revise via Step 5, or run a new fresh round (Step 4) when the feedback changes direction fundamentally. Then re-render. Nothing is implemented.
• Implement — locks the chosen approach and proceeds to Step 7. Only this explicit action starts implementation.

Persist the decision to .claude/rethink/<date>-<slug>/decision.md.

Step 7 — Autonomous Handoff

Only after an explicit Implement decision. Assemble the task briefing:

• the chosen approach (full concept content),
• scope boundary, success criteria, no-gos, corridor,
• test mandate: pin the profile per deep-knowledge/test-plan.md and verify per its recommendation,
• pointer to the .claude/rethink/<date>-<slug>/ artifacts.

Then invoke devops-run-autonomous with this briefing as the task. Autonomous owns permission priming, confirmation, worktree, implementation, testing, and the report — do not re-implement any of that here.

If the user chose an over-corridor approach: ask ONE explicit corridor-widening question before the handoff.

Completion

• Handoff happened (Step 7) → devops-run-autonomous owns the completion card.
• Run ends earlier (no decision yet, or user stops at the concept page) → render a completion card yourself (analysis if nothing was written, ready if brief/approaches were persisted). Artifacts remain on disk for a later resume.

Error Handling

Situation	Behavior
Empty $ARGUMENTS	One intake question first
No repo evidence	Heavier question round
App not startable	Static evidence; dossier notes it
Lens agent failure	Continue with ≥1 approach, gap noted on page
User picks over-corridor approach	Explicit corridor-widening question before handoff
No decision on concept page	Nothing implemented; artifacts kept for resume

Rules

• Never skip the question round — evidence hypotheses are validated, not assumed correct.
• Never feed code, file paths, or current-implementation details to the fresh-phase agents.
• Never start implementation without the explicit Implement decision.
• The corridor is agreed with the user, never assumed.