feat: add node repair operator for autonomous task failure recovery (#1016)

When a task fails verification, the executor now attempts structured
repair before interrupting the user:
- RETRY: adjusts approach and re-attempts
- DECOMPOSE: splits the task into smaller verifiable sub-tasks
- PRUNE: skips with justification when infeasible

Only escalates to the user when the repair budget is exhausted or an
architectural decision is required (Rule 4). Configurable via
workflow.node_repair (bool) and workflow.node_repair_budget (int).
Defaults to enabled with budget=2.

Inspired by the NODE_REPAIR operator in STRUCTUREDAGENT (arXiv:2603.05294).

Co-authored-by: buftar <buftar@gmail.com>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>

CHANGELOG.md

@@ -6,6 +6,9 @@ Format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ## [Unreleased]
 
+### Added
+- **Node repair operator** (`workflows/node-repair.md`) — autonomous recovery when task verification fails. Instead of immediately asking the user, the executor attempts structured repair: RETRY (different approach), DECOMPOSE (break into sub-tasks), or PRUNE (skip with justification). Only escalates to the user when the repair budget is exhausted or an architectural decision is needed. Repair budget defaults to 2 attempts per task; configurable via `workflow.node_repair_budget`. Disable entirely with `workflow.node_repair: false` to restore original behavior.
+
 ## [1.22.4] - 2026-03-03
 
 ### Added

get-shit-done/workflows/execute-plan.md

@@ -287,7 +287,22 @@ Orchestrator parses → presents to user → spawns fresh continuation with your
 </step>
 
 <step name="verification_failure_gate">
-If verification fails: STOP. Present: "Verification failed for Task [X]: [name]. Expected: [criteria]. Actual: [result]." Options: Retry | Skip (mark incomplete) | Stop (investigate). If skipped → SUMMARY "Issues Encountered".
+If verification fails:
+
+**Check if node repair is enabled** (default: on):
+```bash
+NODE_REPAIR=$(node "./.claude/get-shit-done/bin/gsd-tools.cjs" config-get workflow.node_repair 2>/dev/null || echo "true")
+```
+
+If `NODE_REPAIR` is `true`: invoke `@./.claude/get-shit-done/workflows/node-repair.md` with:
+- FAILED_TASK: task number, name, done-criteria
+- ERROR: expected vs actual result
+- PLAN_CONTEXT: adjacent task names + phase goal
+- REPAIR_BUDGET: `workflow.node_repair_budget` from config (default: 2)
+
+Node repair will attempt RETRY, DECOMPOSE, or PRUNE autonomously. Only reaches this gate again if repair budget is exhausted (ESCALATE).
+
+If `NODE_REPAIR` is `false` OR repair returns ESCALATE: STOP. Present: "Verification failed for Task [X]: [name]. Expected: [criteria]. Actual: [result]. Repair attempted: [summary of what was tried]." Options: Retry | Skip (mark incomplete) | Stop (investigate). If skipped → SUMMARY "Issues Encountered".
 </step>
 
 <step name="record_completion_time">

get-shit-done/workflows/node-repair.md

@@ -0,0 +1,92 @@
+<purpose>
+Autonomous repair operator for failed task verification. Invoked by execute-plan when a task fails its done-criteria. Proposes and attempts structured fixes before escalating to the user.
+</purpose>
+
+<inputs>
+- FAILED_TASK: Task number, name, and done-criteria from the plan
+- ERROR: What verification produced — actual result vs expected
+- PLAN_CONTEXT: Adjacent tasks and phase goal (for constraint awareness)
+- REPAIR_BUDGET: Max repair attempts remaining (default: 2)
+</inputs>
+
+<repair_directive>
+Analyze the failure and choose exactly one repair strategy:
+
+**RETRY** — The approach was right but execution failed. Try again with a concrete adjustment.
+- Use when: command error, missing dependency, wrong path, env issue, transient failure
+- Output: `RETRY: [specific adjustment to make before retrying]`
+
+**DECOMPOSE** — The task is too coarse. Break it into smaller verifiable sub-steps.
+- Use when: done-criteria covers multiple concerns, implementation gaps are structural
+- Output: `DECOMPOSE: [sub-task 1] | [sub-task 2] | ...` (max 3 sub-tasks)
+- Sub-tasks must each have a single verifiable outcome
+
+**PRUNE** — The task is infeasible given current constraints. Skip with justification.
+- Use when: prerequisite missing and not fixable here, out of scope, contradicts an earlier decision
+- Output: `PRUNE: [one-sentence justification]`
+
+**ESCALATE** — Repair budget exhausted, or this is an architectural decision (Rule 4).
+- Use when: RETRY failed more than once with different approaches, or fix requires structural change
+- Output: `ESCALATE: [what was tried] | [what decision is needed]`
+</repair_directive>
+
+<process>
+
+<step name="diagnose">
+Read the error and done-criteria carefully. Ask:
+1. Is this a transient/environmental issue? → RETRY
+2. Is the task verifiably too broad? → DECOMPOSE
+3. Is a prerequisite genuinely missing and unfixable in scope? → PRUNE
+4. Has RETRY already been attempted with this task? Check REPAIR_BUDGET. If 0 → ESCALATE
+</step>
+
+<step name="execute_retry">
+If RETRY:
+1. Apply the specific adjustment stated in the directive
+2. Re-run the task implementation
+3. Re-run verification
+4. If passes → continue normally, log `[Node Repair - RETRY] Task [X]: [adjustment made]`
+5. If fails again → decrement REPAIR_BUDGET, re-invoke node-repair with updated context
+</step>
+
+<step name="execute_decompose">
+If DECOMPOSE:
+1. Replace the failed task inline with the sub-tasks (do not modify PLAN.md on disk)
+2. Execute sub-tasks sequentially, each with its own verification
+3. If all sub-tasks pass → treat original task as succeeded, log `[Node Repair - DECOMPOSE] Task [X] → [N] sub-tasks`
+4. If a sub-task fails → re-invoke node-repair for that sub-task (REPAIR_BUDGET applies per sub-task)
+</step>
+
+<step name="execute_prune">
+If PRUNE:
+1. Mark task as skipped with justification
+2. Log to SUMMARY "Issues Encountered": `[Node Repair - PRUNE] Task [X]: [justification]`
+3. Continue to next task
+</step>
+
+<step name="execute_escalate">
+If ESCALATE:
+1. Surface to user via verification_failure_gate with full repair history
+2. Present: what was tried (each RETRY/DECOMPOSE attempt), what the blocker is, options available
+3. Wait for user direction before continuing
+</step>
+
+</process>
+
+<logging>
+All repair actions must appear in SUMMARY.md under "## Deviations from Plan":
+
+| Type | Format |
+|------|--------|
+| RETRY success | `[Node Repair - RETRY] Task X: [adjustment] — resolved` |
+| RETRY fail → ESCALATE | `[Node Repair - RETRY] Task X: [N] attempts exhausted — escalated to user` |
+| DECOMPOSE | `[Node Repair - DECOMPOSE] Task X split into [N] sub-tasks — all passed` |
+| PRUNE | `[Node Repair - PRUNE] Task X skipped: [justification]` |
+</logging>
+
+<constraints>
+- REPAIR_BUDGET defaults to 2 per task. Configurable via config.json `workflow.node_repair_budget`.
+- Never modify PLAN.md on disk — decomposed sub-tasks are in-memory only.
+- DECOMPOSE sub-tasks must be more specific than the original, not synonymous rewrites.
+- If config.json `workflow.node_repair` is `false`, skip directly to verification_failure_gate (user retains original behavior).
+</constraints>