Files
correx/examples/workflows/role_pipeline.toml
T
kami e46777e29f kernel: deterministic static_check stage (§B-§5) + critique-outcome producer (§B-§6)
Two reviewer-reliability tracks, both deterministic and unit-verified (no model/network).

§B-§5 — static_check stage seam. The StaticAnalysisRunner + StaticFindingsRecordedEvent +
reviewer-context filter already existed; what was missing was a stage that runs the tool and
emits the event. Added:
- StaticCheckStageExecutor: reads stage metadata (static_tool/static_argv), runs the configured
  command via StaticAnalysisRunner, returns a StaticFindingsRecordedEvent. No-op (empty findings)
  when no runner is wired or no command is configured — safe to carry unconfigured.
- A deterministic-stage seam in DefaultSessionOrchestrator.enterStage: any stage with
  metadata["stage_type"] == "static_check" is run by the executor instead of the LLM subagent,
  then advances on its (unconditional) exit edge.
- TomlWorkflowLoader: stage_type/static_tool/static_argv fields → StageConfig.metadata.
- role_pipeline.toml: a static_check stage between implementer and reviewer (no-op until
  static_argv + a CommandRunner are set; activation is live-QA-gated, see StaticAnalysisRunner doc).

§B-§6 — critique-outcome producer. The CritiqueFinding type + CriticCalibrationProjection existed
but nothing fed them. Added:
- CritiqueFindingsRecordedEvent (+ CritiqueVerdict): the producing side — a critic's findings +
  verdict for one review iteration, carrying modelHash for per-model calibration.
- CritiqueOutcomeCorrelator: pure loop-resolution logic deciding UPHELD (fixed between rounds) /
  DISMISSED (persisted into an approved final) / INCONCLUSIVE (open at a non-approved terminal),
  per critic (role + modelHash) and per finding id.
- A hook in completeWorkflow/failWorkflow that correlates recorded findings into
  CritiqueOutcomeCorrelatedEvents at loop resolution — no-op when none recorded, idempotent.
  (LLM-side finding emission stays a separate model-gated activation.)

Tests: StaticCheckStageExecutorTest (4), StaticCheckStageTest integration (1, fake runner →
event + transition), CritiqueOutcomeCorrelatorTest (8), CritiqueCalibrationWiringTest integration
(1, seeded findings → outcomes at completion), updated RolePipelineWorkflowTest. Full suites for
core:events/kernel/critique, infrastructure:workflow, testing:integration green; detekt clean.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-22 21:14:00 +00:00

155 lines
5.3 KiB
TOML

# Role pipeline: analyst → architect → planner → implementer → static_check ⇄ reviewer
#
# Each stage produces a typed artifact that the next stage `needs`, so work flows forward
# without a human relaying notes. The decision journal (pinned into every stage's context)
# carries steering/approvals/verdicts across the whole run, so the reviewer sees the same
# ground truth as the planner.
#
# The implementer⇄reviewer loop is gated by review_report.verdict:
# approved → done
# changes_requested → back to implementer (capped by implementer.max_retries, then escalates)
#
# Requires these artifact kinds in ~/.config/correx/config.toml (schemas under docs/schemas/):
# [[artifacts]]
# id = "analysis"; schema_path = "schemas/analysis.json"; llm_emitted = true
# [[artifacts]]
# id = "design"; schema_path = "schemas/design.json"; llm_emitted = true
# [[artifacts]]
# id = "impl_plan"; schema_path = "schemas/impl_plan.json"; llm_emitted = true
# [[artifacts]]
# id = "review_report"; schema_path = "schemas/review_report.json"; llm_emitted = true
#
# Prompt files (prompts/*.md, relative to this workflow) must exist for a real run.
id = "role_pipeline"
start = "analyst"
# 1. Understand the request and the relevant code. Read-only.
# ground_references: every workspace-relative file path the analysis names is checked
# for existence; a hallucinated path fails the stage and retries with the misses fed back.
[[stages]]
id = "analyst"
prompt = "prompts/analyst.md"
produces = [{ name = "analysis", kind = "analysis" }]
allowed_tools = ["file_read", "ShellTool"]
ground_references = true
token_budget = 16384
max_retries = 2
# 2. Decide the approach and component boundaries.
[[stages]]
id = "architect"
prompt = "prompts/architect.md"
needs = ["analysis"]
produces = [{ name = "design", kind = "design" }]
token_budget = 16384
max_retries = 2
# 3. Break the design into ordered, verifiable steps.
[[stages]]
id = "planner"
prompt = "prompts/planner.md"
needs = ["design"]
produces = [{ name = "impl_plan", kind = "impl_plan" }]
token_budget = 16384
max_retries = 2
# 4. Implement the plan. Writes files (jailed to the workspace). The loop target —
# max_retries here caps how many review→implement refinement rounds are allowed.
# Optional: a `writes` manifest (workspace-relative globs) hard-bounds where this
# stage may write — a FILE_WRITE outside it is blocked as scope creep. Left open
# here because the targets are task-specific; a task-scoped workflow would set e.g.
# writes = ["core/sessions/**", "testing/sessions/**"]
[[stages]]
id = "implementer"
prompt = "prompts/implementer.md"
needs = ["impl_plan"]
produces = [{ name = "patch", kind = "file_written" }]
allowed_tools = ["file_read", "file_write", "file_edit", "ShellTool"]
token_budget = 32768
max_retries = 3
# 4b. Deterministic static analysis (NO LLM). Runs a configured tool (compiler/ktlint/detekt)
# over the patch and records its findings as a StaticFindingsRecordedEvent, which the
# reviewer-context filter then strips from the reviewer's context so the LLM reviewer spends
# its attention on semantic review rather than re-finding what the build already reports.
# No-op until `static_argv` is set (and a CommandRunner is wired) — safe to carry unconfigured.
# Activate e.g. with: static_argv = "./gradlew detekt --console=plain"
[[stages]]
id = "static_check"
stage_type = "static_check"
static_tool = "detekt"
token_budget = 0
max_retries = 0
# 5. Review the patch against the plan AND the analyst's acceptance criteria. The reviewer
# needs `analysis` so it judges the diff against concrete, pre-stated criteria (§5 narrow
# question) rather than whole files against taste.
[[stages]]
id = "reviewer"
prompt = "prompts/reviewer.md"
needs = ["patch", "impl_plan", "analysis"]
produces = [{ name = "review_report", kind = "review_report" }]
token_budget = 32768
max_retries = 2
# --- forward edges ---
[[transitions]]
id = "analyst-to-architect"
from = "analyst"
to = "architect"
condition_type = "artifact_validated"
condition_artifact_id = "analysis"
[[transitions]]
id = "architect-to-planner"
from = "architect"
to = "planner"
condition_type = "artifact_validated"
condition_artifact_id = "design"
[[transitions]]
id = "planner-to-implementer"
from = "planner"
to = "implementer"
condition_type = "artifact_validated"
condition_artifact_id = "impl_plan"
# implementer → static_check (once the patch is validated) → reviewer. The static_check stage
# produces no artifact, so its exit edge is unconditional (always_true).
[[transitions]]
id = "implementer-to-static_check"
from = "implementer"
to = "static_check"
condition_type = "artifact_validated"
condition_artifact_id = "patch"
[[transitions]]
id = "static_check-to-reviewer"
from = "static_check"
to = "reviewer"
condition_type = "always_true"
# --- verdict-gated loop exit / re-entry ---
[[transitions]]
id = "review-approved"
from = "reviewer"
to = "done"
condition_type = "artifact_field_equals"
condition_artifact_id = "review_report"
condition_field = "verdict"
condition_value = "approved"
condition_operator = "eq"
[[transitions]]
id = "review-changes-requested"
from = "reviewer"
to = "implementer"
condition_type = "artifact_field_equals"
condition_artifact_id = "review_report"
condition_field = "verdict"
condition_value = "approved"
condition_operator = "neq"