feat(opencode): make local turns survive, and make failures visible - #48

Investigation of "OpenCode stops mid-task on a local model and I have to type
continue", grounded in OpenCode's own session database (58k assistant turns),
its log, the meter's telemetry, and a throwaway OpenAI-compatible endpoint that
recorded exactly what OpenCode puts on the wire.

The stalls were overwhelmingly the endpoint dying, not OpenCode losing its place.
Turns that never completed over 7 days: glm-meter 0.9%, local-orch 0.0%,
local-dev 38.1%. By hour, 2026-08-25 00:00-04:00 shows 40 local turns and 40
failures -- exactly the M2 wedge window. Historically, 5,167 of 5,430
local-review stream errors were Bad Gateway clustered on Aug 2-6, the M2 outage
that went unreported for four days. What made it look like an OpenCode fault is
that it fails silently: three retries, an AI_RetryError line in its log, then the
turn ends with nothing on screen.

Changes, each tied to something measured:

- limit.output 32000 (was 16384, set earlier today). local-orch has produced
  32,000-token turns and been truncated at exactly that ceiling three times; 2 of
  the 34 recorded "continue" nudges followed such a turn. 16384 was my error --
  it halved a ceiling the model actually uses. OpenCode clamps the wire value at
  32000 regardless, verified against the recorder.

- options.timeout / chunkTimeout per local provider. Worst observed
  time-to-first-token on these endpoints is 209s, and OpenCode does not retry a
  chunk timeout at all (anomalyco/opencode#20466), so a default in the tens of
  seconds would silently kill healthy turns.

- configs/opencode/plugins/local-stall-alert.mjs raises a desktop notification on
  session.error, with a one-minute cooldown so an outage does not fire forty
  times. It fixes nothing upstream; it makes a silent failure announce itself.

- Dropped the `instructions` entry. ~/.config/opencode/AGENTS.md is a symlink to
  the same file OpenCode auto-loads, so every house rule was being sent twice:
  12,555 duplicated characters in every request, 15% of the system prompt.
  Verified end to end -- the system prompt fell from 81,046 to 67,421 chars and
  the whole request from 121,076 to 106,121 bytes.

- Disabled local-ai-delegate_run_local_plan: never called once in 47,240 recorded
  tool calls. Three more never-called tools (list_mcp_resources,
  list_mcp_resource_templates, read_mcp_resource) CANNOT be disabled from config
  -- `mcp*`, `*resource*` and `*mcp_resource*` were each tested against the
  recorder and left all three in place. A test now rejects those dead patterns so
  the config never claims coverage it does not have.

- small_model moved off the paid cloud; 21,734 title generations were going there.

- The mlx launchers set --max-tokens (server default is 512, and applies to any
  request that omits it -- a trap for curl probes and new clients, not for
  OpenCode which sends its own) and gain optional --chat-template-args.
  DELIBERATELY unset by default: thinking off measured 89% -> 78% recall on the
  code-review eval, so it belongs on a bulk lane, never on the reviewer.
  Takes effect on the next server restart; nothing restarted here.

- rotate-logs.sh now covers OpenCode's log directory. Its log was one unrotated
  374 MB file, mostly 137,653 "duplicate skill name" warnings. Rotated in place.

Not done, deliberately: `steps` stays unset, because capping it forces a
text-only summary mid-task, which looks exactly like the stall being fixed. And
the duplicate skills themselves are spread across ~/.agents, ~/.claude,
~/.opencode and more -- shared with four other agents, so consolidating them is
the operator's call, not a change to make from inside this repo.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>