Roadmap

All three variants structurally valid, Cast E2E-verified in a live Claude Code session, Alloy and Temper passed load verification.

• Three installable plugins

  anneal-cast, anneal-alloy, anneal-temper — each a complete Claude Code plugin with its own manifest, commands, skills, agents, hooks, and validation scripts.
• Eight-invariant contract

  Red Team Trinity always runs, Hephaestus validation always runs, dual output, skill enrichment, unbounded re-loop, parallelization by default, category routing, dual prompts by model family.
• Greek-god agent roster

  Metis, Prometheus (per-variant), Synthesizer (Alloy), Deepen-Loop (Temper), Momus, Red-Team Trinity, Oracle, Hephaestus, Atlas.
• Umbrella marketplace

  anneal-umbrella-dev lists all three plugins; install any subset.
• Opus 4.7 semantic-XML output schema

  _shared/opus-47-xml-schema.md is the canonical format; validator ships as scripts/validate-xml.py.
• Per-variant validation scripts

  scripts/validate-plugin.py in each variant; runs at install time and on-demand.

Known issues at v0.1.0

Stop making users type anneal-cast:anneal, anneal-alloy:anneal, anneal-temper:anneal. One command, three backends, auto-select.

Single /anneal command

# Old (v0.1.0) — three separate command namespaces
/anneal-cast:anneal "task"
/anneal-alloy:anneal "task" --versions 5
/anneal-temper:anneal "task" --depth 3

# New (v0.2.0) — one command, three backends
/anneal "task"                       # auto-select variant from task shape
/anneal --cast "task"               # force Cast
/anneal --alloy "task" --versions 5  # force Alloy, N=5
/anneal --temper "task" --depth 3    # force Temper, depth cap 3

The auto-select heuristic reads the probe report and Metis's preliminary classification:

Shared .anneal/runs/ state directory

.anneal/
├── runs/
│   ├── {variant}-{run_id}/         # one dir per run, all variants
│   └── ...
├── manifest.json               # index of all runs with metadata
└── comparisons/                # cross-variant comparisons
    └── {task-slug}/
        ├── cast-output/
        ├── alloy-output/
        └── temper-output/

Different projects have different plan-quality standards. A fintech codebase needs security-weighted scoring; a greenfield startup needs minimalism-weighted scoring. v0.3.0 lets users override Momus's scoring rubric and Alloy's bias set per project.

.anneal/rubric.yaml — project-local override

# Project-local rubric — overrides global defaults

momus:
  anchors:
    safe_min: 85        # default: 85
    caution_min: 70     # default: 70
    risky_min: 50       # default: 50
  weights:
    correctness: 1.0
    evidence_plan: 1.5  # this project rewards strong evidence planning
    sequencing: 0.8     # this project is ok with parallel phases

alloy:
  bias_set:
    - correctness
    - minimalist
    - defensive
    - domain-security    # custom domain bias
    - domain-compliance  # custom domain bias

temper:
  depth_cap_default: 4
  convergence:
    variance_threshold: 0.25    # tighter than default 0.30
    delta_threshold: 0.10       # tighter than default 0.15

Custom Alloy biases as skill files

Ship biases as SKILL-like files under ~/.claude/skills/anneal-biases/{bias-name}/SKILL.md. Anneal picks them up at tournament time. Enables domain-specific biases: security-first for a fintech codebase, compliance-first for healthcare, performance-first for gamedev.

Anneal emits plans. ValidationForge runs them and reports back. v1.0.0 wires the round-trip: VF verdict folds into Anneal's next run as a new Metis directive, and the loop terminates when VF reports overall PASS or Anneal emits three consecutive ABORTs.

Plan-to-verdict round-trip

vf_verdict:
  run_id: <anneal-run-id>
  vf_run_id: <vf-run-id>
  plan_path: .anneal/runs/.../plan/
  phase_verdicts:
    phase-00-preflight:           PASS
    phase-01-database-schema:     PASS
    phase-02-sign-in-flow:        FAIL
    phase-03-magic-link-fallback: SKIPPED
    phase-04-session-persistence: SKIPPED
    phase-05-logout-invalidation: SKIPPED
    phase-06-functional-validation: BLOCKED_BY_UPSTREAM
  overall_verdict: FAIL
  first_failing_phase: phase-02-sign-in-flow
  failure_evidence:
    - e2e-evidence/auth/phase-02-oauth-callback.png
    - e2e-evidence/auth/phase-02-sign-in-response.json

anneal learn — closing the loop

anneal learn --from-verdict <vf-verdict.yaml> replays the Anneal run with the phase-level failure folded in as a new Metis directive. The resulting re-run is constrained to avoid the failure mode.

The full circle: Anneal plans → VF validates → Anneal learns from VF's verdict → VF re-validates the new plan. The loop terminates when VF reports overall_verdict: PASS or Anneal emits three consecutive ABORT decisions.

CLI-first invocation

# Non-interactive CI-friendly invocation
anneal cast "task" --output .anneal/runs/custom-id/
anneal alloy "task" --versions 5 --output .anneal/runs/custom-id/
anneal temper "task" --depth 3 --output .anneal/runs/custom-id/

# Full round-trip: plan → validate → learn if fail
anneal cast "task" | vf-validate --auto-learn

# Cost estimator (pre-flight, no run)
anneal estimate "task"
# → estimated 9 agents, ~$2.50–3.20, ~10-12 min

Directional arcs, soft dates

These are directional targets, not commitments. Priority is demand-driven.

Open requests as of 2026-04-23

Priority is rough; exact order depends on implementation effort and requester volume. To influence order, open an issue at github.com/krzemienski/anneal with a concrete use case.

Explicit non-goals

These are recurring asks that don't fit the design. The decisions are deliberate, not deferrals.

Staying up to date