Skip to content

docs: draft design spec for adversarial-tdd out-of-the-box workflow - #114

Open
DaveGerson wants to merge 19 commits into
masterfrom
claude/agent-baton-workflow-nryy6a
Open

docs: draft design spec for adversarial-tdd out-of-the-box workflow#114
DaveGerson wants to merge 19 commits into
masterfrom
claude/agent-baton-workflow-nryy6a

Conversation

@DaveGerson

@DaveGerson DaveGerson commented Aug 10, 2026

Copy link
Copy Markdown
Owner

Staged, model-tiered, adversarially verified TDD delivery pipeline as a
named workflow preset (baton plan --workflow adversarial-tdd). Draft —
under architecture review; will be amended before implementation.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo

claude added 19 commits August 9, 2026 07:48
@claude
docs: draft design spec for adversarial-tdd out-of-the-box workflow
2c7b75d 
Staged, model-tiered, adversarially verified TDD delivery pipeline as a
named workflow preset (baton plan --workflow adversarial-tdd). Draft —
under architecture review; will be amended before implementation.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
docs: amend adversarial-tdd workflow spec after architecture review (v2)
82a0a5a 
Incorporates two independent architecture reviews: audit-phase carryover
(compliance blocker), team-member re-tiering, explicit harvest predicate
without the IMPLEMENTATION name fallback, inverted field-preservation
rule, manager-mode mutual exclusion in v1, config sibling-key handling,
pure-applier gate injection, CLI interaction matrix, and scope cuts
(no validator/fable change, no baton workflows command, lossy DB fields
documented).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
feat(agents): add test-adequacy-reviewer agent and adversarial-tdd pl…
1a6e4c2 
…aybook

New opus reviewer agent that adversarially verifies authored tests
against a spec (coverage gaps, vacuous/tautological/over-mocked tests)
without reading implementation code — stage 4 of the adversarial-tdd
workflow. Bundled-agents mirror re-synced. Playbook added for the PMO
gallery.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
test: red-state TDD suite for adversarial-tdd workflow preset (198 te…
3b571fa 
…sts)

Pins the spec'd behaviors of the workflow preset registry, the
WorkflowApplier reshape semantics (harvesting, re-tiering incl. team
members, gates/approvals, audit carryover, fan-out, external verifier,
idempotency, graph invariants), WorkflowSettings loading, the two new
plan-model fields, and the plan CLI --workflow flag contract. All fail
by design: implementation follows after adversarial test verification.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
docs: spec rulings from adversarial test verification (v2.1)
d8b9f6d 
Rulings: last non-carryover phase invariant, member_id preservation,
config-default manager mode suppressed under --workflow, effective-tier
reporting in WorkflowDecisions, and normative test addenda (harvest
normalization, PREPARATION/REMEDIATION coverage, settings/gate CLI
wiring, idempotency-guard differentiator, stage scoping pins).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
test: close adversarial-verification gaps in workflow TDD suite (235 …
20149bc 
…tests)

Pins harvest name normalization (colon/last-word keying incl.
'Implement: API Layer', 'Backend Implementation'), PREPARATION and
REMEDIATION archetypes, baton.yaml->plan settings wiring (acceptance
criterion 4), CLI fallback-gate wiring on gateless plans, a real
idempotency-guard differentiator, per-reviewer briefing path scoping,
stage-1/stage-4 spec-path scoping, build-gate acceptance, harvested-phase
auditor negative, cross-phase dep re-keying, and config-default
manager-mode suppression under --workflow.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
feat(workflow): adversarial-tdd workflow preset, applier, settings, CLI
9f84649 
Named delivery-workflow presets reshaping the assembled MachinePlan
post-planning (manager-mode post-processor precedent):

- core/workflow/presets.py: WorkflowStage/WorkflowPreset registry with
  the built-in adversarial-tdd preset (spec->architecture->test
  authoring->test verification->implementation->implementation
  verification->final review, tiered fable/fable/opus/opus/sonnet/
  opus/fable).
- core/workflow/applier.py: pure, idempotent WorkflowApplier — explicit
  harvest predicate (last-word/colon keying, IMPLEMENTATION/PREPARATION/
  REMEDIATION only), inverted field preservation with recursive team
  re-tiering, gate/approval preservation, audit-phase carryover,
  final-review fan-out, optional external-vendor automation verifier.
- core/config/workflow.py: WorkflowSettings loaded from baton.yaml
  (stage agent/model overrides, external_command, fan-out knobs);
  manager config recognizes the sibling key.
- models/execution.py: MachinePlan.workflow + PlanStep.workflow_stage,
  conditionally serialized.
- plan_cmd: --workflow flag, mutual exclusions, application before all
  consumer branches, CLI-computed fallback gate, config-default manager
  suppression, --explain workflow section; goal_cmd passthrough.

235-test TDD suite green; neighbor suites green (4 pre-existing golden
snapshot failures confirmed on baseline).

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
docs: document the adversarial-tdd workflow preset across the doc matrix
a523f07 
cli-reference (--workflow flag + workflow: config section), agent roster
(test-adequacy-reviewer, count 30->31), orchestrator-usage recipe 15,
package-layout (core/workflow + config/workflow), baton-engine reference
plan table, baton.yaml.example workflow section, installed-template
CLAUDE.md flag list, root CLAUDE.md/GEMINI.md architecture rows.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
docs: spec amendments from independent verification (v2.2); revert te…
701d6ec 
…mplates/CLAUDE.md

Rule 2(a) also fires when qualifying phases yield no harvested steps;
rule 5.8 partitions by dispatch unit with the reviewer formula exact
(never step-clamped); rule 5.1 no-op guard performs zero writes.
templates/CLAUDE.md restored — it is a distributable artifact the root
conventions mark do-not-modify.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
fix(workflow): close independent-verification findings
36238be 
- Research-support guidance appended to every applier-created stage
  briefing (spec §3 Notes), pinned by tests.
- Final-review fan-out partitions by dispatch unit: the reviewer formula
  is exact and never clamped by harvested step count; a team step
  contributes one unit per member (member description + step paths).
- Unknown --workflow names now fail before any planning work runs.
- WorkflowSettings numerics validated (ge=1); unknown stage ids in the
  stages override map rejected with an actionable error.
- Harvest fallback (a) fires when qualifying phases yield no harvested
  steps, not only when no phase qualifies.
- No-op guard performs zero writes (plan_diagnostics untouched).
- Unrecognized non-harvesting stage ids raise instead of silently
  receiving final-review briefings; workflow_stage rendered in plan.md.

Target suite 258 tests green (stable across 20 hash seeds); neighbor
sweep 800 green with only the 4 pre-existing golden failures.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
docs: final-review contract fixes — ADR-26, counts, anchors, spec rul…
e6da8e6 
…ings

ADR-26 + doc-audit entry; agent count 30->31 across README, roster
header, GEMINI, pillars, bundled mirror; real heading for the
workflow-presets anchor; baton goal --workflow row; 300s automation-cap
caveat; baton.yaml.example overrides commented out; spec v2.3 rulings:
carryover dep re-keying, carryover-excluded gate scan and approval OR,
inline-auditor exclusion ruling, settings/preset validation notes,
status -> Accepted.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
fix(workflow): re-key carryover deps; exclude carryover phases from g…
8aba6b9 
…ate/approval scan; preset invariants

Final-review fixes: carryover steps' depends_on are re-keyed intra-phase
and dropped otherwise (a verbatim old id silently pointed at a reshaped
stage phase, or failed plan-load validation at execute start); the
first-test/build-gate scan and the approval OR now skip
carryover-eligible phases (no gate aliasing/double-run, no doubled
sign-off from a single source); presets validate structural invariants
at registration (non-empty stages, exactly one harvesting stage, review
stage last) and the bare next() paths carry explicit errors; real
StackProfile in the CLI gate wiring. 268 target tests green; each
regression test verified to fail against the pre-fix code.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
wip: docs overhaul in progress (README + agent-facing discovery, mid-…
3642742 
…fleet checkpoint)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
docs(readme): engine-first install, adversarial-tdd quickstart, deliv…
df7d830 
…ery-workflows section

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
wip: docs journeys page + cross-links (mid-fleet checkpoint)
b95aff9 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
docs: finish delivery-workflows journeys page
3467580 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
fix(workflow): re-key team member ids on harvest+carryover, auditor-a…
abb66ac 
…ware carryover, phase-order preservation, reshape validation

- Re-key TeamMember.member_id (and intra-team depends_on) when harvested or
  carried-over team steps are renumbered, mirroring engine _renumber_phases;
  fixes DISPATCH Parent-Step/Record-With pointing at the wrong step and the
  resulting re-dispatch deadlock.
- Detect auditor across the flattened team tree so team-consolidated Audit
  phases carry over; compute carryover eligibility before harvesting so
  fallback harvest cannot steal steps from auditor-bearing phases.
- Record dropped base steps in WorkflowDecisions.dropped_steps with warnings
  when auditor work would be lost.
- Preserve base-phase ordering across the flattened Implementation phase and
  clear stale parallel_safe flags invalidated by the new edges.
- Validate the reshaped plan via MachinePlan round-trip; raise typed
  WorkflowReshapeError with actionable message.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
wip: engine/config patch wave checkpoint (worktree team-context sync,…
460b04b 
… fable tables, automation timeouts)

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
@claude
fix(runtime): plan.json rescue for automation timeout on reloaded steps
ed63777 
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FVZ9JqpfgfH4FesTs6uWCo
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@DaveGerson @claude