chorus-codes 0.7.0

package/templates/architect-review.yaml

@@ -0,0 +1,52 @@
+id: architect-review
+name: Architecture Review
+description: Two-phase decision workflow. Anthropic drafts a design; 3 cross-lineage reviewers critique it before any coding starts.
+author: chorus
+agreementThreshold: 0.5
+onThresholdMet: ask
+maxRounds: 3
+yoloDefault: false
+# 2-phase decision workflow with denser plan scaffolding. Cockpit-side floor.
+estimatedBaselineTokens: 900
+
+phases:
+  - id: plan
+    kind: plan
+    title: Architecture Proposal
+    description: Anthropic (Opus) drafts an architecture proposal or design decision.
+    doer:
+      lineage: anthropic
+      models:
+        - claude-opus-4-7
+    reviewer:
+      require: 0
+      crossLineage: false
+      candidates: []
+    inputs:
+      include: []
+      exclude: []
+    iterate:
+      maxRounds: 1
+      onDisagreement: continue
+
+  - id: design-review
+    kind: review
+    title: Design Critique
+    description: Three reviewers from different lineages critique the proposal. Lower threshold (50%) encourages surfacing disagreement.
+    doer:
+      lineage: any
+      models: []
+    reviewer:
+      require: 3
+      crossLineage: true
+      candidates:
+        - lineage: anthropic
+        - lineage: openai
+        - lineage: google
+    inputs:
+      include:
+        - plan
+      exclude: []
+    iterate:
+      maxRounds: 2
+      onDisagreement: continue

package/templates/bug-diagnose.yaml

@@ -0,0 +1,31 @@
+id: bug-diagnose
+name: Bug Diagnosis
+description: Adversarial diagnosis. Anthropic produces a hypothesis; Gemini independently challenges it. Good for complex bugs where multiple perspectives uncover the root cause.
+author: chorus
+agreementThreshold: 0.5
+onThresholdMet: ask
+maxRounds: 2
+yoloDefault: false
+# 1 doer + 1 reviewer; multi-phase reasoning. Cockpit-side floor.
+estimatedBaselineTokens: 700
+
+phases:
+  - id: diagnose
+    kind: plan
+    title: Diagnose Bug
+    description: Anthropic (Opus) produces a hypothesis about what's broken and why. Gemini independently challenges it.
+    doer:
+      lineage: anthropic
+      models: []
+    reviewer:
+      require: 1
+      crossLineage: true
+      candidates:
+        - lineage: google
+          models: []
+    inputs:
+      include: []
+      exclude: []
+    iterate:
+      maxRounds: 2
+      onDisagreement: continue

package/templates/code-review.yaml

@@ -0,0 +1,46 @@
+id: code-review
+name: Code Review
+description: Single-phase review template. Claude Opus 4.7 produces; Codex (gpt-5.5) and Gemini 3.1 Pro Preview review independently. Both reviewers must agree.
+author: chorus
+agreementThreshold: 0.66
+onThresholdMet: ask
+maxRounds: 3
+yoloDefault: false
+# Single-phase doer + 2 reviewers. Cockpit-side floor for cost preview.
+estimatedBaselineTokens: 700
+
+# Ship phase: when chat is created with `repoPath`, after reviewers agree
+# chorus branches off main, commits the doer's diff, pushes, and opens a PR
+# via `gh pr create`. No auto-merge — human clicks Merge in GitHub.
+# Skip behavior: when chat has no repoPath, ship is auto-skipped and chat
+# ends `approved`.
+ship:
+  enabled: true
+  branchPattern: "chorus/{chatId}"
+  titleTemplate: "chorus: code-review via #{chatId}"
+
+phases:
+  - id: review
+    kind: review
+    title: Code Review
+    description: Claude Opus 4.7 produces the implementation; Codex (gpt-5.5) and Gemini 3.1 Pro Preview critique it independently. Quorum is 2 of 2.
+    doer:
+      lineage: anthropic
+      models:
+        - claude-opus-4-7
+    reviewer:
+      require: 2
+      crossLineage: true
+      candidates:
+        - lineage: openai
+          models:
+            - gpt-5.5
+        - lineage: google
+          models:
+            - gemini-3.1-pro-preview
+    inputs:
+      include: []
+      exclude: []
+    iterate:
+      maxRounds: 2
+      onDisagreement: continue

package/templates/red-green.yaml

@@ -0,0 +1,169 @@
+id: red-green
+name: Red / Green (Adversarial)
+description: TheDailyClaude's 7-phase flow. Cross-lineage doer/reviewer pairs at every phase; implementer blind to tests; deterministic verify loop. Catches subtle bugs that self-review misses.
+author: '@TheDailyClaude (reddit)'
+agreementThreshold: 0.66
+onThresholdMet: ask
+maxRounds: 3
+yoloDefault: false
+# 7-phase TheDailyClaude flow; densest scaffolding. Cockpit-side floor.
+estimatedBaselineTokens: 1100
+
+# Ship phase: same shape as code-review. Auto-skipped when chat has no
+# repoPath (e.g. a paste-a-snippet quick check). v0.5 stops at `gh pr create`
+# — no auto-merge regardless of how comprehensive red-green's verification
+# was. Human gate at the merge button is intentional.
+ship:
+  enabled: true
+  branchPattern: "chorus/{chatId}"
+  titleTemplate: "chorus: red-green via #{chatId}"
+
+phases:
+  - id: plan
+    kind: plan
+    title: Plan
+    description: Anthropic drafts the implementation approach; OpenAI reviews.
+    doer:
+      lineage: anthropic
+      models:
+        - claude-opus-4-7
+    reviewer:
+      require: 1
+      crossLineage: true
+      candidates:
+        - lineage: openai
+    inputs:
+      include: []
+      exclude: []
+    iterate:
+      maxRounds: 2
+      onDisagreement: continue
+
+  - id: spec
+    kind: spec
+    title: API Specification
+    description: OpenAI derives the API contract and specification from the plan; Anthropic reviews.
+    doer:
+      lineage: openai
+      models: []
+    reviewer:
+      require: 1
+      crossLineage: true
+      candidates:
+        - lineage: anthropic
+    inputs:
+      include:
+        - plan
+      exclude: []
+    iterate:
+      maxRounds: 2
+      onDisagreement: continue
+
+  - id: tests
+    kind: tests
+    title: Test Suite
+    description: Anthropic writes failing tests against the specification; OpenAI reviews.
+    doer:
+      lineage: anthropic
+      models:
+        - claude-opus-4-7
+    reviewer:
+      require: 1
+      crossLineage: true
+      candidates:
+        - lineage: openai
+    inputs:
+      include:
+        - spec
+      exclude: []
+    iterate:
+      maxRounds: 2
+      onDisagreement: continue
+
+  - id: implement
+    kind: implement
+    title: Implementation
+    description: OpenCode (or fallback) implements based ONLY on spec—cannot see tests. This info asymmetry prevents overfitting to test literals.
+    doer:
+      lineage: opencode
+      models: []
+    reviewer:
+      require: 1
+      crossLineage: true
+      candidates:
+        - lineage: anthropic
+    inputs:
+      include:
+        - spec
+      exclude:
+        - tests
+    iterate:
+      maxRounds: 3
+      onDisagreement: continue
+
+  - id: verify
+    kind: verify
+    title: Verification
+    description: Run tests against implementation. Feed back failing test names only (no bodies/expected results) to preserve spec-driven discipline.
+    doer:
+      lineage: any
+      models: []
+    reviewer:
+      require: 0
+      crossLineage: false
+      candidates: []
+    inputs:
+      include:
+        - implement
+        - tests
+      exclude: []
+    iterate:
+      maxRounds: 3
+      onDisagreement: continue
+
+  - id: final-review
+    kind: review
+    title: Final Review
+    description: OpenAI reviews the complete implementation against plan, spec, tests, and code; Anthropic arbitrates.
+    doer:
+      lineage: openai
+      models: []
+    reviewer:
+      require: 1
+      crossLineage: true
+      candidates:
+        - lineage: anthropic
+    inputs:
+      include:
+        - plan
+        - spec
+        - tests
+        - implement
+      exclude: []
+    iterate:
+      maxRounds: 2
+      onDisagreement: continue
+
+  - id: divergence
+    kind: divergence
+    title: Divergence Check
+    description: If implementation diverges from spec, flag it. Optionally loop back to spec for refinement.
+    doer:
+      lineage: anthropic
+      models:
+        - claude-opus-4-7
+    reviewer:
+      require: 1
+      crossLineage: true
+      candidates:
+        - lineage: openai
+    inputs:
+      include:
+        - plan
+        - spec
+        - tests
+        - implement
+      exclude: []
+    iterate:
+      maxRounds: 1
+      onDisagreement: continue

package/templates/review-only.yaml

@@ -0,0 +1,73 @@
+id: review-only
+name: Review Only
+description: |
+  Paste a diff, draft, or text blob. Six reviewers (Codex + Gemini + Claude
+  + opencode-go Kimi + opencode-go DeepSeek + opencode-go Qwen) critique it
+  independently. Single pass — revise yourself and resubmit for another round.
+author: chorus
+agreementThreshold: 0.66
+onThresholdMet: ask
+maxRounds: 1
+yoloDefault: false
+# Lightest template — no doer phase, single reviewer-only ask. Cockpit-side floor.
+estimatedBaselineTokens: 500
+
+# Ship is meaningless without a doer diff. The runner skips it for any
+# review-only template; the explicit `enabled: false` here keeps intent
+# obvious to template-readers.
+ship:
+  enabled: false
+
+phases:
+  - id: review
+    kind: review_only
+    title: External Review
+    description: Five lineages critique the supplied artifact independently.
+    reviewer:
+      require: 2
+      crossLineage: true
+      candidates:
+        - lineage: openai
+          models:
+            - gpt-5.5
+        - lineage: google
+          models:
+            - gemini-3.1-pro-preview
+        - lineage: anthropic
+          models:
+            - claude-opus-4-7
+        # opencode-go transports — both run through the opencode binary
+        # (lineage=opencode). Different model families (Moonshot's Kimi
+        # vs DeepSeek) carried on the voice's vendor_family for diversity
+        # scoring, but both are listed under the opencode lineage because
+        # that's the binary that actually runs them. Listing kimi as
+        # `lineage: moonshot` would imply the standalone kimi-CLI
+        # subscription — an impossible combo for users without it.
+        - lineage: opencode
+          models:
+            - opencode-go/kimi-k2.6
+        - lineage: opencode
+          models:
+            - opencode-go/deepseek-v4-pro
+        - lineage: opencode
+          models:
+            - opencode-go/qwen3.6-plus
+    artifact:
+      label: "Artifact to review"
+      hint: "Paste a unified diff, a markdown draft, code, or any text blob."
+      maxBytes: 1048576
+    inputs:
+      include: []
+      exclude: []
+
+# Reviewer fallback — when any reviewer slot exhausts (quota_exhausted,
+# auth_error, rate_limit, etc.) Chorus raises a Claude Sonnet 4.6
+# replacement so the cockpit shows a finished review instead of an
+# ERRORED card. Same-lineage dedupe means this won't fire if Claude Opus
+# is already running in this slot — it activates for the openai/google/
+# opencode reviewers when their primary model fails.
+fallback:
+  reviewer:
+    - lineage: anthropic
+      models:
+        - claude-sonnet-4-6

package/templates/tri-review.yaml

@@ -0,0 +1,41 @@
+id: tri-review
+name: Tri-Review
+description: Single-round, three-reviewer template. Claude Opus 4.7 produces; Codex (gpt-5.5), Gemini 3.1 Pro Preview, and Kimi K2.6 review independently. Quorum is 2 of 3.
+author: chorus
+agreementThreshold: 0.66
+onThresholdMet: ask
+maxRounds: 1
+yoloDefault: false
+# Single-phase doer + 3 reviewers. Cockpit-side floor for cost preview.
+estimatedBaselineTokens: 750
+
+phases:
+  - id: review
+    kind: review
+    title: Tri-Review
+    description: Claude Opus 4.7 produces; Codex, Gemini, and Kimi review independently. 2-of-3 quorum.
+    doer:
+      lineage: anthropic
+      models:
+        - claude-opus-4-7
+    reviewer:
+      require: 2
+      crossLineage: true
+      candidates:
+        - lineage: openai
+          models:
+            - gpt-5.5
+        - lineage: google
+          models:
+            - gemini-3.1-pro-preview
+        - lineage: moonshot
+          models:
+            - kimi-k2.6
+    inputs:
+      include: []
+      exclude: []
+    iterate:
+      maxRounds: 1
+      onDisagreement: continue
+      shareSessionAcrossRounds: false
+      shareSessionAcrossPhases: false