Three defects the first sweep left behind, each caught by a rule the guard did not yet have. Statuses. RULES.md gives three; the tree had 25 values, and the free-text ones were carrying progress. "Implemented (wiring landed)" and "Ratified (rev. 2e). Rev. 2e adds the OpenTaskByPR lookup. Rev. 2d revised section 10..." were both changelogs living in the one field the other rules exempt. Every PRD now declares Draft, Approved or Finalised and nothing else; five that declared nothing now declare one. A revision number moves to **Revision:**, a pointer to the authoritative document to **Authority:**, and what a past revision changed to git. CLAUDE.md's citation of state-store-contract rev. 2e still resolves. Stale cross-references. "Supersedes the 2026-09-03 open question", "this bullet previously said", "an earlier draft of this document claimed" all point at text the reader cannot see, because it was replaced. A supersede marker survives only when it names a file, a section or a symbol that can be found. Temporal hedges. 100 uses of today, currently and right now across 31 files. Each dates its sentence to the day it was written: "[notify] today is one webhook URL, read in one place" stops being true the moment this design lands, while still reading as the premise it rests on. The header-field exemption was also too wide. "**Date:** 2026-09-03, Channel resolved 2026-09-22" smuggled a dated status into the one line the rules did not read; the exemption is now a Date holding only a date and a Status holding only one of the three.
5.6 KiB
Dynamic workflow triage
Status: Finalised
Epic: archie-core-enfj. Scope: give workflow.Route() a real, content-aware
fallback instead of unconditionally defaulting every unlabeled task to the
heaviest workflow.
Problem
Route() (internal/domain/workflow/workflow.go) has exactly two real
signals: an explicit t.Workflow, and label matches via
workflowForLabels (routing.go, bug→tdd, feature→feasibility,
bootstrap→bootstrap). Anything else — which is every chat-spawned task,
since task_spawn rarely sets labels — falls straight through to
reg["implement"], the full prepare → baseline → plan → build → commit-push → review → open-pr pipeline, regardless of what the task
actually asks for.
task_spawn's own tool schema overpromises: "workflow: … Omit to let the
daemon route it" (internal/gateway/task_tools.go) implies real
routing exists. It doesn't — "route" means "check labels, then
give up and run the heaviest workflow anyway."
Confirmed in production: a chat-spawned task whose
entire content was "this is just a test, close it when you receive it" ran
the full implement pipeline end to end — 669,421 tokens, ~7 minutes —
when nothing about the request needed a worktree-driven build at all. The
plan stage's own LLM output correctly concluded no code change was
needed, but that conclusion arrives two expensive stages (baseline,
plan) too late to save anything.
Decision
Add a triage workflow: cheap, classifies once, then either closes the
task immediately or hands off to the workflow the task actually needs.
Route()'s final fallback (reg["implement"]) becomes
reg["triage"] when a triage entry is registered, else the existing
implement fallback — a two-line change, fully backward compatible when
triage isn't wired up.
Triage does not replace label-based routing. A labeled task
(bug/feature/bootstrap) already has a free, reliable signal and goes
straight to its known workflow unchanged — spending a classification
call there would be pure waste. Triage only fires in the gap that
defaults to implement blindly: no explicit workflow, no label match.
Triage workflow shape
StagePrepareWorktree() -- still needed: AgentStage mounts tc.Dir into the
agent container: internal/domain/workflow/agent.go:74
passes tc.Dir straight through. No cheaper
primitive exists in this codebase for a
directory-less classification call.
AgentStage "classify" -- ReadOnly, ONE cheap agent turn, no exploration
tools required beyond what ReadOnly already
grants. Mirrors feasibility.go's decideCaptureTools
pattern: a structured capture tool, called
exactly once, MaxCalls: 1.
Mission: read the task's title/body/labels (already in taskPromptBlock)
and decide, via a captured decide tool call:
{"needs_code_change": bool, "workflow": "implement" | "tdd" | "feasibility", "reasons": "..."}
needs_code_change: false → close path: reuse closeNoChangesIssue
(steps.go) directly — issue closed if forge-backed, task marked
StatusMerged/"completed -- no changes required" otherwise. No baseline,
no plan, no build, no commit, no PR. This is the actual fix for the
production incident: a no-op request now costs one classification call
(low thousands of tokens) instead of 669,421.
needs_code_change: true → sets tc.Task.Workflow to the classifier's
chosen workflow (implement/tdd/feasibility, defaulting to
implement if the field is missing or unrecognized) and an Outcome that
requeues the task under it, the same requeue mechanism
chatTaskControllerAdapter.ApproveChatTask already uses
(main.go, requeue(ctx, taskID, fromStatus, workflow)). The task then
re-enters Route() on its next claim with t.Workflow now set, and takes
the normal path for that workflow from there — triage never runs its
classification twice for the same task.
Non-goals (v1)
- Full bug/feature intent classification quality — the classifier only
needs to be directionally right; a wrong
tddvsimplementcall still produces a working PR, just via a slightly less-tailored pipeline. Tightening this is a follow-up, not a blocker. - Any change to label-based routing,
Route()'s explicit-workflow branch, orBootstrap()/Feasibility()/TDD()themselves. - Triage for forge-sourced (real GitHub issue) tasks with no labels — in scope structurally (same fallback branch), but the dominant case this fixes is chat-spawned tasks, which essentially never carry labels.
Testing
workflow_test.go:Route()picksreg["triage"]overreg["implement"]for an unlabeled task when both are registered; falls back toreg["implement"]unchanged whentriageisn't registered (regression guard for every existingRoute()test).- New
triage_test.go: classify →needs_code_change: falsecloses without ever callingtc.Trees.Preparea second time or touchingtc.Agent.Runagain; classify →needs_code_change: truesetstc.Task.Workflowand produces a requeue Outcome, using a fakeagentRunnerFuncreturning a captured decision, mirroringfeasibility_test.go's existingdecideCaptureToolstest style. - Integration-shaped test reproducing the production scenario: a
chat-spawned task with no labels, title "this is just a test, close it
when you receive it" — asserts the task closes without ever reaching a
baseline/plan/buildstage.