#1719 [Feature]: a native agent "Harness" tab in the Strata web app (RFC)

Beschreibung

# RFC: a native agent "Harness" tab in the Strata web app

Draft for the upstream feature-request issue (Niko1221/Strata). Written in the
repo's docs style: plain words, measured numbers with what they were measured
on, no claims without a measurement.

---

## What do you want to do?

Strata today serves the model to *other* agents — Claude Code, Codex, Hermes,
MCP clients all point at `/v1`. But the Strata app itself has no agent: to run
a tool-using, multi-turn job on your own PC you still need a third-party
harness installed next to it.

I would like Strata to offer an optional fourth tab — **Harness** — a native
agent loop that runs against the local server in-process:

- multi-turn ReAct loop with tool calls (read/write/edit files, search, shell),
  streamed live to the tab;
- sessions that survive browser reloads: append-only JSONL event log, resume,
  fork, export (markdown/JSONL);
- approvals: every dangerous tool call is gated fail-closed (allow/deny/ask)
  with a durable policy event — the tab never runs a shell command without a
  decision;
- context management sized for local models: token meter driven by the
  monitor's own numbers, compaction that shadows old spans instead of
  rewriting them, spill files for oversized tool output;
- plan mode, goal mode, mid-turn steering, and subagents (child sessions with
  depth/budget caps);
- monitor awareness: the context-fill gauge and slot state come from
  `Service.metrics()` — the harness compacts before the engine 400s.

Goal, not implementation: "run an agent on your own PC without installing
another tool". The workaround I use today is pointing an external harness at
`127.0.0.1:8080/v1`, which works but keeps sessions, approvals and context
management outside Strata's view.

## Why this fits the project

- It answers demand that already shows in issues: agent workloads on Strata
  (#980 field report, #1431 tool-call recovery in agent workloads, #1615
  harnesses recognizing context-overflow, #728/#694 agentic-loop pain).
  A first-party harness can handle overflow and retry policy where the server
  actually knows the context state.
- It is additive and opt-in: `--harness` flag / `"harness": true` in the
  config, off by default. No changes to the engine, the C++ layer, the
  template layer, or any existing route. The fork boundary is
  `serve/harness/` (Python, stdlib only) + `serve/web/harness.js|css` +
  small hooks at the tail of the routes table and the tab list.
- It matches the UI philosophy: vanilla ES2020 JS, no build step, reuses
  tokens.css/components.css/st-* classes and the sprite icons.
- Security inherits the existing model: loopback by default, `_own_page`
  gates on control routes, approvals default to 'ask', keys read from env or
  run-config, never argv. Nothing is exposed beyond 127.0.0.1 without
  `--api-key`.

## Status: it exists and is tested

I built this as a fork ("strata-harness") over the last days, on upstream
0.1.40.4:

- ~4,300 lines added across 40 files (12 commits on the `harness` branch);
- 125 harness tests + the full upstream suite (262 serve tests) green under
  `python -m unittest`, plus ruff, pyright (strict on `serve/harness/`) and
  eslint gates;
- MockEngine-based tests — no GPU needed, same style as `test_server.py`;
- ADRs (9) and UX design/review docs for each UI milestone;
- a security audit of the fork (tracked files, full git history, key
  handling) before any push.

## Proposed PR series (small, reviewable)

1. **Scaffold**: `serve/harness/` package skeleton, `--harness` flag,
   routes table tail + hidden tab button, ADRs, lint/test configs,
   `tools/test_harness.sh`. No behaviour when the flag is off.
2. **Sessions + loop + tools**: events/session/loop/tools modules with tests,
   CLI driver only (no UI).
3. **The tab**: harness.js/css, SSE stream, approvals UI, monitor gauge —
   itself split into transcript/stream, then approvals, then gauge.
4. **Context management**: meter, compaction, spill, plan/goal modes,
   steering.
5. **Subagents + presets + export + keyboard nav.**

Each slice is independently testable and keeps the tab hidden until the
previous one lands. Happy to re-cut the series however review prefers, and to
follow whatever contribution process you want for a feature this size — I
could not find one documented, so this issue is my RFC.

## Extra notes

- MIT-licensed fork; I would contribute the whole thing under Strata's MIT
  license.
- The name "harness" is used here in the agent sense (the thing around the
  model), which collides with #1615's "harnesses" (external clients). Easy to
  rename the tab if that reads badly.