docs: HARNESS, INVARIANTS, BOUNDARIES, DEVELOPMENT, EVAL_HARNESS, SECURITY, ARCHITECTURE + README, CONTRIBUTING, CLAUDE.md, CHANGELOG, TASKS.md (#25, #26)

Closes #25 + #26 in one PR (the README references docs/* paths and TASKS.md
references everything else; landing them separately would mean a half-broken
docs tree in between).

docs/* (all written for the template, not Teller-flavoured):
- HARNESS.md     — umbrella table mapping every layer to its config file
                   and to the meta-gate(s) that catch drift in it.
- INVARIANTS.md  — five portable rules with numbered slots 6+ for project
                   additions.
- BOUNDARIES.md  — ASCII layer diagram + the import-linter contract spec
                   + how to add a layer cleanly.
- DEVELOPMENT.md — prereqs, first-time setup, dev stack, justfile recipes
                   table, branching diagram, commit-prefix table, CI
                   workflow inventory, agent-hook setup, branch-protection
                   token setup.
- EVAL_HARNESS.md — runner architecture, three tolerance modes, wiring
                    your agent / LLM client, adding a case, opt-in for
                    nightly schedule.
- SECURITY.md    — threat model table + defence-in-depth ASCII map +
                    container hardening notes + explicit out-of-scope list
                    (auth, WAF, rate-limit, secret manager).
- ARCHITECTURE.md — scaffold component diagram, request lifecycle,
                    frontend lifecycle, slots that fill in as the project
                    grows.

Top-level docs:
- README.md       — what ships / quickstart / why-a-harness / docs index /
                    versions table / license.
- CONTRIBUTING.md — branching diagram, commit-prefix table, PR template
                    callouts, "adding a check" recipe.
- CLAUDE.md       — agent project instructions: read-first list, workflow,
                    code conventions, what-not-to-do, skills inventory.
- CHANGELOG.md    — release-drafter seed; first Unreleased entry summarises
                    the harness extraction.
- docs/TASKS.md   — full ticket table with phase grouping + status emoji,
                    matches the GitHub Project board.

Closes #25
Closes #26

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Author: constk

CHANGELOG.md

@@ -0,0 +1,21 @@
+# Changelog
+
+All notable changes to this project will be documented in this file.
+
+The format is loosely based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+Released versions are drafted automatically by [release-drafter](https://github.com/release-drafter/release-drafter); see `.github/release-drafter.yml` and `.github/workflows/release-drafter.yml`. Each entry on the GitHub Releases page corresponds to a tag of the form `vX.Y.Z`.
+
+## Unreleased
+
+### Added
+
+- Initial harness scaffold (Python 3.14 + FastAPI + Pydantic v2 + OpenTelemetry; React 19.2 + Vite + TypeScript strict).
+- 15 required CI status checks (lint, typecheck, tests, coverage ≥ 75 %, import-linter, pre-commit, frontend build/quality, security suite, two meta-gates, PR-title lint).
+- Release pipeline: tag-triggered build, push to GHCR, CycloneDX SBOM, GitHub Release publish.
+- Eval harness scaffold (provider-agnostic runner + LLM-judge Protocol + 1 example golden case + workflow_dispatch nightly).
+- `.claude/` agent integration (3 hooks, 6 auto-activating skills, settings example).
+
+### Notes
+
+- This template was extracted from a financial-agent take-home (Teller) and generalised. The harness is the product; the scaffold exists so every gate has something to operate on.

CLAUDE.md

@@ -0,0 +1,63 @@
+# CLAUDE.md — agent project instructions
+
+You are working in `harness-python-react`, a template repo whose harness IS the product. Code quality here is enforced mechanically — every gate fails CI, not just tests. Keep that bar as you work.
+
+## What this repo is
+
+A production-quality LLM-driven coding harness over a minimal FastAPI + React scaffold. The point isn't the features (one `/health`, one `/echo`, one hello page); the point is that every layer of the pipeline — lint, types, architecture, security, eval, agent hooks — catches a different failure class without anyone remembering to run it.
+
+## Read first
+
+- [`docs/HARNESS.md`](docs/HARNESS.md) — umbrella; the controls and where they live.
+- [`docs/INVARIANTS.md`](docs/INVARIANTS.md) — the load-bearing rules. Every PR is checked against them.
+- [`docs/BOUNDARIES.md`](docs/BOUNDARIES.md) — layered import-linter contract; reverse imports fail CI.
+- [`docs/DEVELOPMENT.md`](docs/DEVELOPMENT.md) — branching, commit format, justfile, CI overview.
+
+## Workflow
+
+- One issue per change. Branch name: `feat|fix|chore|docs|test|refactor/<issue-number>-<kebab-title>`.
+- One PR per branch, base `develop`. PR title = the conventional-commit subject.
+- `develop → main` happens via a `release:` PR.
+- The pre-push gate is `just check` (lint + typecheck + architecture + tests). Run it before pushing.
+- For frontend changes, also run `just frontend-check`.
+
+## Code conventions
+
+- **Python:** 3.14, `uv run --frozen` everywhere, mypy `--strict`, ruff with the wide select set (`E W F I N UP B SIM TCH S RUF`).
+- **Type hints:** every public function. `from __future__ import annotations` at module top.
+- **Models:** anything crossing a module / process seam inherits from `StrictModel` (`src/models/_base.py`). `extra="forbid"`. Add `strict=True` to the class when you want strict type coercion (rejecting `"3.14"` → float).
+- **API:** every route under `/api/v1/`. Typed Pydantic responses, not raw dicts.
+- **Layer flow:** one-way. Reverse imports are a CI failure. See `docs/BOUNDARIES.md`.
+- **Observability:** OTel `agent_span(...)` for any operation in the request path; semconv-defined attribute keys only (constants at the top of `src/observability/spans.py`).
+- **Frontend:** React 19 + TS strict; functional components + hooks; never `dangerouslySetInnerHTML` on backend output; SSE consumers use the typed primitive at `frontend/src/lib/api/client.ts`.
+
+## What NOT to do
+
+- Don't bypass gates. `--no-verify` / `--no-hooks` / `--no-gpg-sign` are blocked by `pretooluse_bash.py` for a reason. If a hook is wrong, fix the hook.
+- Don't introduce a new commit-type prefix without updating both `pyproject.toml`'s commitizen schema AND `pr-title.yml` (the `Commit-type sync` meta-gate will fail otherwise).
+- Don't add a CI job without listing it in `.github/branch-protection/{develop,main}.json` (the `Branch-protection contexts sync` meta-gate will fail).
+- Don't skip the architecture contract by accident — `lint-imports` runs in CI and locally via `just architecture`.
+- Don't write code without tests. Coverage gate is 75% on `src/`.
+- Don't hand-roll secrets into config. Use env / `.env` (gitignored) → `Settings` from `src/models/config.py`.
+- Don't create files unless they're necessary. The scaffold has no dead modules.
+
+## Use the skills
+
+The agent-side skills in `.claude/skills/` auto-activate based on context:
+
+- `architect` — when designing module boundaries, API contracts, layer-flow decisions.
+- `code-reviewer` — after writing/editing code; runs the 10-point review checklist.
+- `devops` — when touching Docker, CI, pyproject.toml, observability config.
+- `frontend` — when working in `frontend/` (React 19 + TS + Vite).
+- `qa-engineer` — when writing tests or extending the eval harness.
+- `technical-writer` — when updating docs / READMEs.
+
+Trust their guidance — they encode this project's conventions.
+
+## When in doubt
+
+- If the change touches a gate, update the meta-gate inputs (`branch-protection/*.json`, `pr-title.yml`, `check_required_contexts.py`'s exemption list).
+- If the change touches an invariant, decide whether the invariant is wrong (update `docs/INVARIANTS.md` in the same PR) or the change is wrong (rework).
+- If a CI job is failing for a reason that doesn't match the change, dig — don't reroll. Recent fix patterns: tag-vs-commit SHA in pinned action references, `if: hashFiles(...)` startup failures (see project-memory), pytest exit-5 on empty test suites.
+
+The harness exists to make sloppy work hard. Lean into it — when a gate trips, it's protecting the next person reading this codebase.

CONTRIBUTING.md

@@ -0,0 +1,79 @@
+# Contributing
+
+Thanks for taking a look. This template's harness is the product, so the contribution flow is opinionated — every change goes through the same gates as a feature.
+
+## Branching
+
+```
+main  ◄── release PR ◄── develop ◄── feat/123-short-name
+                                  ◄── fix/124-bug-name
+                                  ◄── chore/125-config-change
+```
+
+- `main` is the release line. Protected: 15 required status checks, code-owner approval, no force pushes.
+- `develop` is the integration branch. Same gates, less strict (PRs don't need rebases).
+- Feature branches are short-lived and named `<type>/<issue-number>-<kebab-title>`. Open one issue per branch so the project board stays usable.
+
+## Commit messages
+
+Seven prefixes (enforced in three places — `[tool.commitizen]` in `pyproject.toml`, `pr-title.yml`, `check_commit_types.py`):
+
+| Prefix | When |
+|---|---|
+| `feat:` | New capability |
+| `fix:` | Bug fix |
+| `docs:` | Documentation only |
+| `test:` | Tests / eval harness |
+| `refactor:` | Internal change with no behaviour delta |
+| `chore:` | Tooling, deps, infra |
+| `release:` | `develop → main` release PRs only |
+
+The subject is **lowercase** after the colon. Title Case prose (`Add the thing`) is rejected; all-caps initialisms (`CI failure`, `SDK upgrade`) are fine.
+
+## Pull requests
+
+1. Open the issue first. Use a feature/bug template; fill every section.
+2. Branch off `develop` with the matching name.
+3. Land one logical change per PR. Stack PRs if the work is naturally split.
+4. The PR template asks five things — answer each (`None` is valid where applicable):
+   - **What & why** (1–3 lines)
+   - **Test plan** (checkbox list; CI covers most of it)
+   - **Invariants affected** — cite numbered rules from `docs/INVARIANTS.md`
+   - **New deps / actions / external surface** (anchor for supply-chain review)
+   - **Screenshots** (UI changes only)
+5. Wait for green CI + a code-owner review before merging.
+
+## Local pre-push gate
+
+```sh
+just check                     # ruff + mypy + import-linter + pytest
+cd frontend && npm run lint && npm run format:check && npm run check && npm run test && npm run build
+uv run pre-commit run --all-files
+```
+
+A green pre-push run is a high-confidence predictor of a green CI run. The `just check` gate is intentionally a subset of CI — fast feedback over coverage.
+
+## Adding a check
+
+When the harness grows a new gate:
+
+1. Add the workflow job in `.github/workflows/`.
+2. If it's a required gate, add the job's display name to the `contexts` arrays in `.github/branch-protection/{develop,main}.json`.
+3. If it's NOT required (scheduled / dispatch-only / push-to-main-only), add the workflow filename to `EXEMPT_WORKFLOWS` in `.github/scripts/check_required_contexts.py`.
+4. Update `docs/HARNESS.md` and `docs/SECURITY.md` (if security-relevant).
+5. Land in one PR — the meta-gate `Branch-protection contexts sync` will fail if you skip step 2 or 3.
+
+## Code of conduct
+
+Be kind. Disagree on substance, not on people. If review feedback gets sharp, take it offline and come back when both sides are ready.
+
+## Reporting security issues
+
+If you find a vulnerability that affects users of the template, **do not open a public issue**. Email the maintainer (see commit history for contact). Include:
+
+- Repro steps
+- Affected version / commit SHA
+- Severity assessment (informational / low / medium / high / critical)
+- Suggested fix if you have one
+
+We'll acknowledge within 72 hours.

README.md

@@ -1,5 +1,72 @@
 # harness-python-react
 
-Production-quality coding harness for Python (FastAPI) backends and Vite + React + TypeScript frontends. Designed for LLM-driven development: every gate — lint, types, architecture, security, eval — is enforced mechanically so code quality stays consistent across many human and AI contributors.
+> A production-quality coding harness for Python (FastAPI) + Vite/React/TypeScript projects. Designed for LLM-driven development: every gate — lint, types, architecture, security, eval — is enforced mechanically so code quality stays consistent across many human and AI contributors.
 
-> **Status:** bootstrap. Full documentation, scaffolding, and the harness itself land across [issues #1–#28](https://github.com/constk/harness-python-react/issues). Track progress on the [project board](https://github.com/users/constk/projects/3).
+## What ships
+
+- **Backend:** Python 3.14, FastAPI, Pydantic v2 (`StrictModel` base), `uv` deps, OpenTelemetry SDK + OTLP exporter, structured JSON logs, generic tool-registry pattern.
+- **Frontend:** Node 24 LTS, React 19.2, Vite 8, TypeScript strict, ESLint 10 flat config, Prettier, Vitest + jsdom + Testing Library.
+- **Eval harness:** provider-agnostic runner + LLM-judge `Protocol`, three tolerance modes (exact / numeric / semantic), one example golden case, nightly workflow (disabled by default).
+- **CI:** 15 required status checks across `ci.yml` (lint/format, mypy strict, unit tests, coverage ≥75%, import-linter architecture, pre-commit, frontend build, frontend quality, branch-protection sync, commit-type sync) + `security.yml` (gitleaks, pip-audit, npm audit, trivy) + PR-title lint.
+- **Release:** tag-triggered workflow that builds the image, pushes to `ghcr.io`, generates a CycloneDX SBOM, and publishes the GitHub Release.
+- **Agent integration:** `.claude/hooks/` (forbidden-flag blocker, secret scan, formatter dispatch, SessionStart context) + six auto-activating skills (architect / code-reviewer / devops / frontend / qa-engineer / technical-writer).
+- **Docker:** multi-stage Dockerfile (non-root, healthcheck), `docker compose up` boots app + frontend + Jaeger.
+
+## Quickstart
+
+```sh
+git clone https://github.com/constk/harness-python-react.git
+cd harness-python-react
+
+uv sync --extra dev
+uv run pre-commit install --hook-type pre-commit --hook-type commit-msg
+(cd frontend && npm ci)
+
+docker compose up   # backend :8000, frontend :5173, Jaeger :16686
+```
+
+The pre-push gate is `just check` (= ruff + mypy + import-linter + pytest). For frontend changes add `just frontend-check`.
+
+## Why a harness
+
+The differentiator isn't the scaffold — it's that every layer of the pipeline catches a different failure class **without relying on the human or LLM coder remembering to run anything**. The same posture protects code regardless of who wrote it.
+
+See [`docs/HARNESS.md`](docs/HARNESS.md) for the full umbrella. Highlights:
+
+- **Pydantic `StrictModel` everywhere a contract crosses a seam** (rejects unknown keys at construction).
+- **`import-linter` enforces one-way layer flow** (`api | eval → agent → tools → data → observability → models`).
+- **Three independent secret scans** (PreToolUse hook → pre-commit gitleaks → CI gitleaks).
+- **Two meta-gates** that catch *drift in the gates themselves*: `Branch-protection contexts sync` (workflow jobs vs branch-protection JSON) and `Commit-type sync` (commitizen schema vs PR-title allowlist).
+- **CycloneDX SBOM attached to every release** for supply-chain attestation.
+
+## Documentation
+
+| File | Purpose |
+|---|---|
+| [`docs/HARNESS.md`](docs/HARNESS.md) | Umbrella: every control + where it lives |
+| [`docs/INVARIANTS.md`](docs/INVARIANTS.md) | The numbered load-bearing rules |
+| [`docs/BOUNDARIES.md`](docs/BOUNDARIES.md) | Module layering + the import-linter contracts |
+| [`docs/DEVELOPMENT.md`](docs/DEVELOPMENT.md) | Local setup, branching, justfile, CI |
+| [`docs/EVAL_HARNESS.md`](docs/EVAL_HARNESS.md) | Eval flywheel + opt-in for the nightly workflow |
+| [`docs/SECURITY.md`](docs/SECURITY.md) | Threat model + defence-in-depth map |
+| [`docs/ARCHITECTURE.md`](docs/ARCHITECTURE.md) | Scaffold-level component view |
+| [`CONTRIBUTING.md`](CONTRIBUTING.md) | Branching, commit format, PR flow |
+| [`CLAUDE.md`](CLAUDE.md) | Agent-facing project instructions |
+
+## Versions
+
+Verified April 2026 (`endoflife.date`):
+
+| Layer | Version | Sunset |
+|---|---|---|
+| Python | 3.14.4 | active feature release |
+| Node LTS | 24.15.0 | through 2028-04-30 |
+| React | 19.2.5 | current stable |
+| Vite | 8.x | current stable |
+| TypeScript | 6.x | current stable |
+
+Bump together (Python in `pyproject.toml`, Node in `frontend/package.json`, both in `Dockerfile` + the CI matrix). Document the bump in `docs/DEVELOPMENT.md`.
+
+## License
+
+[MIT](LICENSE).