[linhdmn]

Summary

The PRD said Last updated: 2026-10-05 and nothing in it recorded that the 0→1 pipeline had ever run through the web UI. It has now, headless, and getting there took three fixes (#46, #47, #48). Per the repo's own rule — a change that lands without the PRD reflecting it is incomplete — this brings it current.

Docs only. No source, no tests.

What changed

Header stamp → 2026-10-07. Names the run (goal-met, 30 steps, $0.1648, five phases, an artifact that builds) and the three fixes in one line each, so a reader who stops at the top knows the pipeline has now run for real.

New §7.3 — the headless run, and the three bugs it took to get one. Each bug with its measurement and the test that fails when it is reverted:

Bug	The measurement	The test
The SSE ping never refreshed the watcher flag	approval/asked in the log, SSE connection ESTABLISHED, GET /api/state → pending: [] for 10+ minutes, every settle POST 409	test/dashboard.test.ts fails on the old code and on a 25s TTL
/product and /loop never started a turn	command/run + command/done, no turn/start	test/loop-command.test.ts fails with the handler reverted
Every subagent ran its own phase machine	four subagents, each stopped at exactly 24 steps, all stopReason: 'refusal', parent told "declined the task", run blocked at 2% of budget	test/subagent.test.ts fails with the policyFor branch reverted

It closes on the method, because that is the finding: all three were found by driving the real thing and reading what it actually did, and all three survived a suite of 800+ passing tests. §7.1 and §7.2 already made that point about wiring bugs found by reading; this is the same lesson from the other direction.

§8 — three limits this session established:

• "Is a front end watching?" is still a heartbeat window, not a connection count. Both halves are now TTLs checked against each other by a test — which is what caught the bug — but a tab that is open and wedged still reads as watching for up to one TTL.
• The fl-live verify.sh is weaker than a build. It asserts a page exists and references assistant-ui, and the first run passed it with a page that would not compile (layout.tsx imported ./globals.css; no CSS was written). Deployment configuration rather than plugin code, but the lesson generalises: a gate that checks shape rather than a successful build can pass work that does not run.
• A 10% research share of a 240-step budget is tight — both live runs spent 23–24 of their 24 steps. Recorded as configuration to tune from a measurement, not as a defect.

§9.4 — the plugin-wiring item is no longer "nothing is covered." Three wirings gained source-shape assertions; the item says plainly that this is a weaker instrument than the fake-context harness it still asks for, and keeps that harness as the highest-value remaining work.

Verification

Docs-only, so the check is that the document is internally true:

• every markdown link resolves to a real file (KNOWN-ISSUES.md, PLAN-close-open-issues.md)
• every §-reference names a heading that exists (1, 7, 7.1, 7.2, 7.3, 8, 9.4)
• every test and source path the file names is real (test/dashboard.test.ts, test/loop-command.test.ts, test/subagent.test.ts, src/subagent.ts, src/watcher-ttl.ts)

All three checked programmatically, not by eye. Full suite and CI are untouched by this change and were green at 808f398.