fix(proxy): forward content-frame API calls + errors to chrome shell - #2

…t lines (G3)

Adds a structured-parser layer over the G2 catch-all (internal/overlay):
recognized multi-line/noisy error blocks fold into a single
{kind, message, file:line} alert instead of surfacing as several `unparsed`
lines. Closes DoD #2 of the get_errors epic.

- structured_parsers.go: StructuredError + Compact(); parsePrisma (folds the
  `prisma:error` header + `Invalid \`prisma.x.y()\` invocation` banner + bare
  file:line call-site + cause into one line, extracting op + file:line + a
  remediation hint); parseDBAuth (raw pg/mysql `password authentication failed`,
  `database "x" does not exist`, `too many connections`). Registry +
  runStructuredParsers + isStructuralPrefix. Anchors informed by R1
  (docs/error-formats/prisma-orm-sql.md).
- alerts.go: ProcessLine runs structured parsers on unclassified lines BEFORE
  the catch-all; structural-prefix lines (block header/banner) are folded at
  the cause line rather than surfaced standalone; adds recentSnapshot().
- ECONNREFUSED intentionally left to the existing `connection-refused` pattern;
  stack frames are already non-surfaced (don't trip the catch-all).

Prisma DB-auth block now renders as one compact line carrying the cause +
prisma.user.findUnique() @ /app/src/db.ts:42 — measurably shorter than the raw
block (asserted). G1 catch-all contract preserved (block still surfaces, now
structured). testing-review passed.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…lking)

Replaced the hand-rolled map[string]interface{} walking in all 8
StatusFetcher.fetch* methods with typed DTOs (status_dto.go) decoded via
json.Unmarshal. Each per-field `m["x"].(string)` / `.(float64)` access is gone;
fetchers now decode into a json-tagged DTO and convert with a small toInfo().

This eliminates the brittle pattern that caused the earlier empty-event-log bug
(fetchStartupLog silently missing a field/key): adding or renaming a status
field is now a typed DTO change in one place instead of matching string keys by
hand across map walks. Behavior is unchanged — same daemon calls, same display
structs out; the overlay tests pass.

Note: this is the safe, type-safety half of the "#2" overview work. The
larger LOC collapse (8 fetches -> 1 batched OVERVIEW verb) was intentionally
NOT done: it would force the daemon to re-derive per-proxy log/page data
in-process (divergence risk) for no real latency gain now that the heavy
per-tick scans are cached. Net LOC here is ~flat; the win is type safety and
single-site field edits, not line count.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Five operator-reported rough edges surfaced from a live session:

- monitor firehose (#3): a single page load emitted per-node DOM mutations
  by the hundred, blowing the downstream consumer's rate limit. Add a
  mutationCoalescer that folds a same-target burst into one rolling
  "[mutation] N changes (M nodes)" summary (500ms / target-change / 50-cap
  buckets, tail-flush on disconnect). Compact stream only; JSON stays 1:1.

- raw gzip in compact log (#4): formatCompact/formatJSON dumped ResponseBody
  raw, but the traffic recorder captures on-the-wire bytes that stay
  compressed for any response not decompressed for HTML injection. New
  httpBodyPreview renders "<gzip-encoded, N bytes>" / "<binary, N bytes>"
  instead of garbage.

- misleading exec-timeout diagnostic (#5): the hub timeout message echoed the
  raw frame arg (usually ""), reading as "no frame" even when the exec was
  routed to the active content frame. New ProxyServer.DescribeExecTarget makes
  the message report the resolved destination plus a heavy-screenshot hint.

- access-URL typo (#6): listenAddr is a full host:port, so "http://localhost"
  + listenAddr produced "http://localhost127.0.0.1:47341". Extract testable
  proxyAccessURL; 0.0.0.0 now renders "http://<your-ip>:PORT".

- daemon start help (#6): the command blocks in the foreground but the help
  text implied auto-start/detach. Spell out that it holds the terminal until
  Ctrl+C and does not detach.

Also determinize TestPageTracker_SessionTimeout, which flaked in the
pre-commit race suite (unrelated to the above): its Active flag is recomputed
against wall-clock now on each read, so a single fixed 5ms sleep can race the
scheduler on an oversubscribed host. Poll for the post-timeout invariant with
generous headroom instead. Ref .claude/rules/testing-timing-assertion-flakes.md.

(Reported item #2 — watch emitting --target — was already fixed; the tool
emits --types. Verified against the built binary, no change needed.)

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Replaces the three LOCAL vendored process-lifecycle patches (agnt commits
77a98d2/613c74e9, b964cb2/e74fad05, 3804dda) with released upstream code
now that they merged and shipped as go-cli-server v0.5.7:

- PR #1 recycled-PID cleanup guard (carry scanner-captured identities through
  post-Wait cleanup instead of re-sampling a possibly-reused PID/PGID)
- PR #2 serialize spawn vs shutdown for StateStarting procs (startMu-ordered
  final-check/spawn/publish; ctx-bounded Shutdown wait)
- PR #3 negative-PID signal only a verified stable group leader

No go.mod replace directive. Vendored tree matches the released v0.5.7 tag
exactly. agnt's own internal/platform/* changes stay local (not upstreamed).

Verified: go vet, GOOS=windows cross-build, and full 'go test -p 1 ./...'
serial suite all green.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014Rk1F3dUJLaL6f9tqu5fRA

…gnals

The recorder's fixed +1200ms settle after __devtool appears is flaky under
CPU load (too short) and wasteful when idle (too long). Pin the replacement
to the ACTUAL signals a live bundle exposes — indicator/toast mounted AND the
websocket transport OPEN — so the gate keys on the connection being live, not
on a timer that fires before it (lessons-ssh-transport.md #2). Predicates are
not yet exported: RED.

worktrack-task: 01M08GG2XSD0PT8RRA8GC9K2RE
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>