Roadmap

TypeScript coding harness for web projects: packages/core for the loop and gate, tsforge.dev for docs.

Shipped through 0.37

Strictness & the gate

• Stack detection → 22 ESLint rule packs (including Three.js) + a 31-rule meta-rule engine (config, CI, supply chain, container, testing, structure). See Rule packs · Meta-rules.
• 4 profiles (recommended → strict / security / opinionated); a type-aware ESLint overlay; progress-based loop termination (no blunt turn cap). See When the gate fails.
• Tests by default: test-sibling-required is an error on changed logic files. Opt-in gate oracles (coverage, boot smoke, property tests). See Tests by default · the gate.

Adapt to your repo

• tsforge setup: a conventions wizard that scans the repo and writes tsforge.config.json (interface naming, enums, test layout, component folders) without weakening the safety floor. See Set up a repo.
• tsforge.config.json overrides + external plugin packs. See config.

The agent surface

• Functional review (tsforge review, gate-aware) · workspace map (tsforge map) · pre-edit scout · declarative recipes (tsforge run) · plan mode · cross-session memory.
• Permissions & policy: 6 modes, deny-first config rules, and a run ledger summarized by tsforge trace. See Permissions & policy · Trace a run.
• Keyless web research: opt-in package metadata/docs, search, fetch, and local browser reads with no required API key. See Web access.
• Greenfield scaffolding: stands up the full BoringStack stack and builds it resource-by-resource against BoringStack’s own gate. See Greenfield scaffolding.

Models & observability

• Provider-agnostic ~/.tsforge/models.json (any OpenAI-compatible API; reasoning expressed as a declarative field-path profile, with presets for the common cases). See models.json.
• MCP servers as read-only context/tools. See MCP servers.
• Live tok/s + /metrics + a --log JSONL analyzer. See Token metrics.

Reliability internals: tool-call repair ladder (L0–L3), hashline edit_lines with snapshot recovery, TTSR stream-interrupting rules, instant per-write type diagnostics, statistical A/B reports, graded self-harness progress, real wall-time records, failure classification, and prompt-cache visibility.

Architecture & independent review: a generated, drift-checked map of all 30 harness subsystems (including dependency witnesses, entry points, and adapter seams); read-only model-driven subagents; and an independent reviewer panel with identity de-duplication, quorum enforcement, and explicit dropout causes. See Subsystems · Delegation.

Highlights since 0.27

• 0.28–0.31: the multiline TUI and /config hub, incremental gate caching, recipe-level profiles/procedure cards, image capabilities, and read-only parallel subagents.
• 0.32–0.35: BoringStack gate-parity and browser-acceptance hardening, live OpenAPI/DB boundary checks, the core↔adapter seam, modern layout intent, strict greenfield re-detection, and strictness-preserving resume.
• 0.36–0.37: the core audit fixes (fail-loud packs/flags/profile floors and normalized writable paths), declarative reasoning profiles, graded self-harness evaluation, reviewer independence/diagnostics, prompt-cache metrics, and the generated architecture map.
• 0.51: LSP navigation tools for the model (go_to_definition, impact, symbol_context) and an @ mention picker that surfaces workspace symbols alongside files. See TypeScript language server · Input editor.
• 0.52: Gate rail TUI — cycle Tasks/Gate in the right rail, structured gate errors with rule docs, Enter-to-steer, configurable pane keybindings, ? keymap overlay, OSC 8 file:line links. See Interactive CLI.
• Three.js pack: canonical three / three/addons/... imports, GPU dispose() contracts, scene-graph APIs (add/remove not .children), loader error paths, instanced-buffer needsUpdate. Detected from three or @react-three/fiber. See Rule packs.

Sequencing rule: Measured wins first. No feature defaults change and no 1.0 tag until A/B numbers against the reference model exist.

Road to 1.0

• Run sweeps: TSFORGE_FEATURE_VARIANTS=git,script across benchmark seeds (TTSR, hashline, and write feedback already graduated to always-on from earlier sweeps)
• Publish numbers (pass rate, edit success, tokens saved)
• Tune the remaining tool-availability defaults from data (git_context, script, web research)
• Freeze config and tool surface

Candidate work

• Rule packs: Hono, Prisma
• TTSR rules derived from eval failure patterns
• Meta-rule severity learning from repair-loop telemetry

Completed (v0.1 track)

Commit	Phase
2198361	stack-detection module for rule pack selection
9c0964e–2936a6f	vendored stack-agnostic ESLint rule packs + behavioral fixtures
8b7e9f3–c42b00f	Drizzle + BullMQ packs
9da5e82–ab64600	Elysia, structured-logging, frontend, auth packs
8c19dfd	wire stack-aware packs into gate and prompt
1d67899–3591e1b	meta-rule engine + gate integration + unified rule docs
6b88b54	tsforge.config.json escape hatch
a803d1a	cost-ordered tool-call repair ladder
f6a21db	hashline edit tool with snapshot recovery
9871ace–37f8795	TTSR stream-interrupting rules
a2126f0	instant per-file type diagnostics on write
4b33bb4	A/B eval support for TTSR, hashline, write-feedback