Additional Fixes to Make Junie Work

[MatthewJones517]

Fix Junie CLI support

tl;dr

JetBrains docs for Junie CLI lie. This PR is the result of testing stuff against the actual harness.

Overview

Fixes Junie CLI support, which was silently broken. The original #45 entry
pointed Junie at a documented proxies entry in ~/.junie/config.json.
Verified against a real Junie CLI install (build 1543.24): that mechanism is
never consulted, headless or interactive, regardless of kind — Junie just
keeps talking to its normal authenticated backend, with no error and an empty
logs folder.

This PR switches Junie to the mechanism that actually works — a Custom LLM
model profile at ~/.junie/models/*.json, selected explicitly with
--model custom:<profile-name> — and fixes three more bugs found while
driving it end-to-end against a live proxy:

1. 5f06839 — Switch from the broken proxies/config.json route to a
   working custom-model-profile route (~/.junie/models/request-logger.json,
   --model custom:request-logger). Also fixes the model profile's
   baseUrl, which must be the full endpoint path (/v1/messages or
   /v1/chat/completions), not a bare origin — Junie posts to it verbatim
   with no path appended.

2. 4e33f22 — Fix blank [request-logger] upstream error: log lines.
   Node throws an AggregateError on a failed dual-stack connection (e.g. to
   localhost), and its .message is always empty — the real reasons are
   one level down in .errors[]. Added a describeError() helper that
   unpacks it, and wired connection failures into the proxy's existing
   burst-throttle so a stuck client retrying in a loop doesn't flood the
   console.

3. b87cce8 — Stop setting anthropic-version in the model profile's
   extraHeaders. Junie's Anthropic client already sends that header
   itself; an extraHeaders entry with the same name doesn't override it,
   it's appended as a second occurrence, and the two get joined with a comma
   on the wire (2023-06-01,2023-06-01) — which Anthropic rejects with a
   400.

4. 1c80009 — Send a real model ID in the profile's id field instead
   of the placeholder "request-logger". Junie sends id verbatim as the
   model field of every request, so a made-up name gets a 404 from the real
   backend. The wizard now asks for a real model ID up front (the same way it
   already does for Pi), discovered from /v1/models when available.

Every fix above was root-caused and confirmed against a real, installed
Junie CLI (build 1543.24) with a live proxy in front of it — not just against
JetBrains' published docs, which turned out to disagree with the real
client's behavior. request-logger/README.md documents this testing status
and explicitly warns against the documented-but-broken proxies route so
nobody rediscovers it the hard way.

Setup instructions for Junie users

1. Run the wizard:

   npm run request-logger

2. Pick Junie when asked which coding agent you use.

3. Enter your backend's base URL — e.g. https://api.anthropic.com for a
   real Anthropic key, or a local server's URL.

4. Pick the wire format — Anthropic-compatible or OpenAI-compatible,
   matching your backend.

5. Pick a model ID — the wizard tries to discover one from /v1/models;
   if that fails (common for Anthropic, which requires auth to list models),
   type one by hand (e.g. claude-sonnet-4-5). This must be a real model
   your backend serves — Junie sends it verbatim as the model field of
   every request.

6. The wizard prints a JSON block and tells you to save it to
   ~/.junie/models/request-logger.json. Do that, and replace the
   placeholder API key (YOUR_ANTHROPIC_API_KEY / Bearer YOUR_OPENAI_API_KEY) with your real key. This file holds a real
   credential in plaintext — don't commit it.

7. Run Junie with the printed command, which selects the profile
   explicitly:

   junie --model custom:request-logger

8. Every request now shows up as a readable Markdown file in
   request-logger/logs/.

If you change your backend or model later: re-run npm run request-logger -- --force to go through the wizard again, and update
~/.junie/models/request-logger.json by hand to match — the file is written
once at wizard time and isn't kept in sync automatically.

Do not use Junie's documented proxies entry in ~/.junie/config.json
for this — it's silently never consulted by the real client, and you'll get
an empty logs folder with no error.