docs: how to trim opencode under the Disengaged threshold, and where it has to happen

[chrischall]

Follow-up to #10 — you offered no preference, so here it is as the docs PR I said I'd
write. Happy to close it if you'd rather keep the issue and take none of this.

Promotes the conclusive parts into the reference docs per AGENTS.md
("Where findings graduate to"). Doc-only; no code touched.

Why this is worth a section

§9 says to keep the toolset lean and names opencode as the persistent-disengage case.
Both useful — but opencode gives a plugin no supported way to follow the advice.
Measured against 1.18.18:

Lever	What happens
config.tools from a plugin's config hook	Resolved into the config (opencode debug config shows it) and then ignored — a config disabling eight tools still produced a request offering all nine
experimental.chat.system.transform	Fires, accepts a replacement system array, request still carries the original prompt
agent.<name>.prompt from the config hook	Ignored the same way

So the trim has to land in whatever sits between opencode and Sydney. That is the bit I
think is worth recording: it explains why the existing advice wasn't actionable for that
harness, rather than opencode simply being too heavy.

Also records what actually fills the payload — <available_skills> was 34,263 of
53,676 chars, an inventory of every skill on the machine, none of which the trimmed
toolset can invoke — and the result: 1,853 chars, 9 tools down to 4.

Plus two traps for anyone else attempting it:

• opencode injects AGENTS.md/global rules as prose under an Instructions from:
  line, not in a structured block, so a "keep only the <tag> blocks" rule silently
  deletes the user's own project rules (I shipped that bug before catching it)
• 1.18.18 ships apply_patch where other versions ship edit/write

One number I did not change

opencode 1.18.18 offers 9 tools, not ~15 — below the ~12 borderline §9 states. So on
tool count alone, current opencode may no longer be the persistent case described. I have
not verified that, so I left your claim standing and flagged it in the blockquote
rather than editing it out. Your call.

Scope

Everything added is fenced in a blockquote as measured against opencode, not against
M365 — no live tenant was involved, and I make no claim the threshold moved. If/when I
run it against a real tenant I'll report back on #10.

Context: this came out of writing an independent opencode plugin from your docs
(chrischall/opencode-copilot-plugin,
MIT, no code copied, credited in its README). Your write-ups were detailed enough to
implement the protocol from scratch, which is a rare thing — thank you.