Guide

Execution, Review, Dissent, and Salvage

Execution is the inner loop. Review closes it. Dissent stress-tests it. Salvage preserves learning when the run drifts.

Source: docs/execution-review-salvage.md

Execution, Review, Dissent, and Salvage

Execution is the inner loop. Review closes it. Dissent stress-tests it. Salvage preserves learning when the run drifts.

Framework positionValue
Skills/execute, /review, /dissent, /salvage
Run after/solution-space and the agent brief
Producesexecution report, review findings, dissent memo, or salvage note
Feedsdurable knowledge extraction or a smaller restart

Execution keeps speed aligned: build the selected slice, check it against the brief, and stop when the work drifts.

Execute

Before editing, run a pre-flight check:

  • [ ] Aim is clear.
  • [ ] Constraints are known.
  • [ ] Context is loaded.
  • [ ] Scope is bounded.
  • [ ] Success criteria are defined.

During execution:

  1. Make the smallest coherent change for the selected solution level.
  2. Add or update evidence.
  3. Run the relevant checks.
  4. Watch for drift.
  5. Stop if the brief is wrong.

Drift signals:

  • scope is expanding;
  • files outside the plan keep appearing;
  • “while I am here” work is accumulating;
  • approach has changed direction more than once;
  • tests are being bent to fit the patch;
  • the agent is protecting code it already wrote.

Review

Review the work against the aim and brief, not the agent's summary.

Ask:

  • Is it still necessary?
  • Is it still aligned?
  • Is it sufficient without overbuilding?
  • Is the mechanism clear?
  • Are ripple effects complete?

A clean review still names residual risk.

Dissent

Dissent assumes the accepted-looking answer still fails.

Look for:

  • the chosen solution level was too low or too high;
  • the symptom moved elsewhere;
  • checks prove implementation, not behavior;
  • role boundaries blurred;
  • a new parallel path was introduced;
  • production conditions are not represented by tests;
  • future maintainers will misunderstand the boundary.

Dissent must be allowed to change the decision. Otherwise it is theater.

Salvage

Salvage when the run is no longer converging.

Keep:

  • better problem statement;
  • constraints discovered;
  • rejected solution paths;
  • checks that should survive;
  • role boundaries that mattered;
  • knowledge artifact candidates;
  • smaller restart plan.

Drop the draft if keeping it makes the system worse.

Artifact

Produce the artifact that matches where the run stopped. Each one must preserve observed evidence, residual risk, and the next action.

Stop pointArtifactMust preserve
Execution completesexecution reportchanged files, behavior changed, commands run, observed results, remaining risk
Review finds issuesreview findingsaccepted behavior, findings, evidence checked, required follow-up
Acceptance needs stress-testingdissent memosteel-man, contrary evidence, pre-mortem, hidden assumptions, recommendation
Run stops convergingsalvage note and restart planwhy salvaged, learnings, guardrails, missing context, smaller restart path

Review check

Reject the execution run if:

  • pre-flight cannot state the aim;
  • the patch ignores the selected solution level;
  • checks were weakened or invented after the fact;
  • review relies on the implementer's explanation;
  • dissent cannot name a real possible failure;
  • salvage preserves code only because time was spent on it.

Go deeper

← Authoring SubagentsKnowledge Extraction →