odin-loop

Use when the user says "audit my code", "find all the bugs", "review until clean", or "grill my changes". Runs an iterative multi-agent review that resolves every finding at or above a configurable severity floor to zero, or stops at a gate or cap. Not for remote, credential, or irreversible changes.

Use when product copy needs buyer-objection evidence collected through approved outreach. Conducts structured interviews via approved channels with consent, then synthesizes a copy recommendation grounded in anonymized exact language. Not for unsolicited outreach or survey design.

Use when a measured cost surface needs one-change-at-a-time reduction under frozen guardrails. Each proposed change is adopted with N=5 gate evidence or recorded dead with the measurement that killed it.

Use when a long agentic project needs each cycle to produce a learning memo with lessons, a keep/iterate/restart decision, and version labels only on quality-cleared artifacts. Drives each slice through the real surface and captures evidence before iterating. Not for single-pass builds.

Use when environment-dependent code, templates, or deployment configuration changes, or when runtime configuration is missing. Produces a bidirectionally reconciled environment-variable contract.

Use when recent resolved feedback may reveal a broader recurring defect pattern: sweep resolved feedback for shared root causes, search the project surface for generalized recurrence, and verify no recurrence remains. Not for source-level feedback collection; use feedback-sweep for that.

Use when a first screen needs lower transfer bytes without visual or behavioral change: measure compressed bytes, reduce assets and bundles to a budget, prove pixel identity, and pass the test suite. Not for visual redesign or behavioral changes.

Use when a visual needs repeatable fixed-view rendering and independent rubric scoring: render the fixed view through a specified interface, score it against a frozen rubric, and prove the saved render clears the threshold. Not for free-form visual review or subjective critique.

Use when a rendering path needs stable frame-time, CPU, GPU, and memory evidence against fixed targets: stabilize the configuration, define a sampling window, and prove every target with two consecutive comparable runs. Not for one-shot profiling or visual quality review.

Use when an authorized reference surface needs a clean-room frontend reconstruction across static, motion, and responsive fidelity. Captures the reference at declared viewports and timelines, reconstructs, and measures fidelity against tolerances. Not for styling or component work without a reference surface.

Use when a complete product, rather than one component, needs production-like acceptance evidence against its documented acceptance criteria. Traverses the capability inventory, records per-capability evidence, and stops at declared success, non-success, or bound. Not for single-component evaluation or evaluation without documented criteria.

Use when a genealogical identity or relationship needs correlation, conflict analysis, and negative-search proof. Applies the Genealogical Proof Standard to classify the proposition as proved, disproved, likely, or possible. Not for non-genealogical identity proof.

Use when a test surface needs coverage raised to a configured target with behavior-guarding assertions, not vacuous execution-only tests. Binds coverage metric and tool, mutation tool and kill threshold, measures baseline, adds guarding tests, and re-measures. Not for line-coverage inflation without mutation proof.

Use when agent strategies need a reproducible finite tournament under a frozen evaluation protocol. Produces a scorecard with move records, scores, and hidden-identity validation. Not for open-ended or infinite matchups — use a continuous benchmark for those.

Use when one named review viewpoint must run fix cycles until a fresh reviewer finds nothing. Leaves verified machine-owned fixes and a written verdict. Not for multi-viewpoint review or remote, credential, publish, deploy, or irreversible changes.

Use when a user asks to iteratively improve a skill, fix skill-quality findings, or resolve scanner alerts (W001, W011, W012) on a skill. Resolves the target, runs structured reviews or scanner-backed remediation, applies fixes with per-edit re-scan, and repeats until clean. Not for agent grading; use skill-doctor for that.

Use when a bug, failure, flake, regression, review finding, or ticket needs the core fixed so it cannot break the same way again. Reproduces the failure, proves the root cause, repairs at the root, and pins with a regression test. Not for greenfield features — use tdd; not for style-only review or typo-class one-liners.

Use when a test suite is too slow and must be accelerated without weakening behavior or coverage. Also handles CI parallelization when the bottleneck is serial execution. Not for deleting tests — use tests-purge-unneeded; not for remote, credential, publish, deploy, or irreversible changes.

Use when thumbnail concepts need real-size, accuracy-first scoring without misleading claims. Renders at real platform dimensions, scores against a fixed rubric with weighted accuracy dimensions, and writes an immutable receipt. Not for generating thumbnails or declaring winners that fail the accuracy rubric.

Use when a human explicitly invokes a watcher cycle for an open pull request that must be watched until mergeable or blocked. Classifies the PR and addresses bounded review findings on its own branch. Don't use for merge, force-push, or work outside the invoked cycle.

Use when authorized workflow history may contain a repeated process worth extracting and replay-testing. Produces a contradiction-tested workflow mined from authorized history.