Tdmcp Roadmap Campaign

tdmcp Roadmap-to-1.0 Campaign

Drives _workspace/campaign_roadmap_v1/ROADMAP_1.0_BACKLOG.md to completion, gated by _workspace/campaign_roadmap_v1/ledger.json (idempotent, resumable, merge-safe). This is orchestration only — it never writes product code itself; it routes each ledger item to the existing specialist sub-harness and records the outcome back into the ledger.

Why this exists (vs tdmcp-backlog-campaign)

tdmcp-backlog-campaign drives a tool-shaped backlog through one pipeline (design→build→integrate→QA→release). The road to 1.0 is not uniform: it mixes new tools (M4), a policy-bounded mixer slice (M5), a coverage CI gate (G2), a bridge exec→REST sweep (G4), docs/governance (G1/G5/G6 prep), and live-validated recipes (G3). Each needs a different specialist. This skill is the thin router that sequences them and keeps one ledger of truth.

Confirmed campaign policy (2026-06-21)

• Release: commit + push, NO tag/bump. Everything lands in CHANGELOG

[Unreleased]. Cutting the version is a manual user step — never tag autonomously ([[no-premature-release-tag]]).

• Blocked items (GPU/CUDA, live-TD, external service): **build offline +

quarantine**. Implement, pass the offline gates, set live to the right UNVERIFIED-pending-* marker, keep going. Quarantine never blocks the campaign.

• Checkpoint: after Wave 1 only, then autonomous to the end. On repeat

failure: retry once → quarantine-and-continue (note it in the wave report).

Routing table (ledger route → sub-harness)

| route | Sub-harness | Agents | |---------|-------------|--------| | docs | tdmcp-docs-roadmap-update | roadmap-docs-editor, docs-cookbook-sync, docs-roadmap-qa | | tools | tdmcp-feature-lead | tdmcp-tool-builder ×N + single-writer integration | | mixer | tdmcp-pipeline | mixer-scene-contract-architect, soundcraft-ui24r-adapter-architect, mixer-policy-safety-qa | | coverage | tdmcp-test-coverage | tdmcp-coverage-writer, tdmcp-coverage-qa | | bridge | tdmcp-bridge-endpoint | tdmcp-bridge-engineer (sequential — bridge slices share files) | | recipes | tdmcp-pipeline | td-architect, td-builder, td-qa (needs live TD) | | submission | tdmcp-submission | submission-architect, submission-qa | | planning | — | tdmcp-backlog-planner (wave computation + ledger transitions) |

Workflow

Phase 0 — Context check (resume vs fresh)

1. Read ledger.json. If absent, the harness isn't initialized — (re)generate it

from ROADMAP_1.0_BACKLOG.md and docs/ROADMAP.md, then stop and report.

1. Reconcile against the tree first — a prior session or a merged PR may have

already shipped a ledger item. For each pending item, cheaply verify it's genuinely undone (grep for the tool/file/gate) before assigning it. Flip already-done items to shipped with an event: reconciled history row.

1. Determine the next ready wave: lowest wave number with pending,

non-blocked items whose depends_on are all shipped. Honor the checkpoint_after_wave gate.

Phase 1 — Plan the wave (tdmcp-backlog-planner)

Spawn tdmcp-backlog-planner (opus) with the ledger + backlog. It returns: the exact item set for this wave, their routes, dependency-safe ordering, and the proposed ledger status transitions. Write its report to _workspace/campaign_roadmap_v1/wave__planner_report.md.

Phase 2 — Execute the wave (route to sub-harnesses)

For each item, invoke its route sub-harness with a precise spec. Within a wave:

• Parallelize independent items across routes (docs ∥ coverage ∥ tools).
• Serialize same-file routes — all bridge items run sequentially (they

share td-client/td/ files); within tools, builders are parallel but the integrator is the single writer of shared registries.

• Blocked/needs-hardware items: build to the offline boundary, then mark

live quarantined per policy — do not wait on hardware.

Every sub-harness already runs its own gates; this skill does not re-implement them. Record per-item QA (typecheck/build/biome/vitest/recipes/bridge) into the ledger qa block, exactly like the project_rag campaign.

Phase 3 — Gate, commit, record

1. Run the four PR gates + recipe + bridge tests at the wave boundary

(npm run typecheck && npm run build && npm run lint && npm test, npm run validate:recipes, npm run test:bridge).

1. Commit + push the wave on its own branch (feature/roadmap-).

No tag, no version bump — CHANGELOG [Unreleased] only.

1. Update the ledger: flip items pending → shipped (or quarantined), stamp

qa, commit, and a history row. Write wave__report.md.

Phase 4 — Loop or checkpoint

• If the just-finished wave is the checkpoint_after_wave → stop and report;

wait for the user before the next wave.

• Otherwise advance to the next ready wave (Phase 0 → 3) until no pending

non-blocked item remains.

Phase 5 — Hand-back

When only quarantine items remain, report campaign complete (offline scope) and enumerate what the user must do to actually tag 1.0: provide GPU/TOX components or a reachable TD for the UNVERIFIED-pending-* live validations, run the bench/hardware M5 spikes, submit to the Connectors Directory, and — only on the user's word — cut the tagged 1.0 minor (closing G1).

Error handling

• A sub-harness that fails a gate: retry once with the failure fed back. On

second failure, mark the item quarantined with the reason, leave the rest of the wave intact, and note it in the wave report — never delete partial work.

• Conflicting tree state vs ledger: trust the tree, reconcile the ledger, note

the drift. Never silently overwrite shipped work.

• The ledger is the single source of truth and must stay valid JSON after every

transition; write it atomically (full rewrite), never partial.

Test scenarios

• Happy path: fresh ledger → Phase 0 picks Wave 1 → planner assigns 4 docs

items → docs sub-harness ships all four → gates green → commit no-tag → checkpoint stop with a Wave-1 report.

• Resume after merge: a follow-up run finds g2_coverage_ci_gate already

landed on main → reconcile to shipped (event: reconciled) → skip it → plan the remainder of Wave 2.

• Blocked item: Wave 4 m4_drive_streamdiffusion builds offline, gates green,

but no GPU → ship with live: UNVERIFIED-pending-gpu, campaign continues.

• Repeat failure: a tool fails its gate twice → quarantined with reason, wave

proceeds with the rest, report flags the gap.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: hybridlabor-api
• Source: hybridlabor-api/bdb-dev-optimized-agent-skills
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.