condense ignores the agent provider: the reflector is wired to a global command with a hardcoded Claude model and Claude-only flags, so an all-OpenCode floor still burns a hidden Claude session #596

Summary

The memory reflector (MemoryReflector / condense) does not use the agent's provider. It is wired to a single global command with a hardcoded claude fallback, a hardcoded Claude model, and Claude-CLI-specific flags. On a floor where every agent runs on OpenCode — which is the whole point of deploying OpenCode — the condense pass is still a hidden Claude session that consumes the Claude subscription.

So the one subsystem that was supposed to be provider-agnostic is the one that still requires Claude.

Evidence (0.4.6, Linux x64, read from app.asar)

1. The reflector is constructed with a global command, not a per-agent provider:

const reflector = new MemoryReflector(
  () => readConfig().harnessHome,
  () => readConfig().defaultCommand ?? "claude",   // <- single global value
  () => memory.env(),
  reflectSettings,
  (event) => { ... }
);

2. The constructor's own doc comment says it is Claude:

 * @param getCommand   The base `claude` command (only its binary name is used).

3. The model is hardcoded, not configurable:

const CONDENSE_MODEL = "claude-haiku-4-5";

4. The hidden call builds Claude-CLI flags regardless of which binary was resolved:

const args = [
  "--model", opts.model,          // claude-haiku-4-5
  "--permission-mode", "bypassPermissions",
  "--disallowedTools", ...disallowed
];

--permission-mode / --disallowedTools are Claude Code flags. They are passed whether the resolved binary is claude or something else.

5. The fallback is baked into the runner too:

const binary = (opts.command || "claude").trim().split(/\s+/)[0] || "claude";
const exe = resolveCommand(binary);

6. There is no provider/model control in the reflect settings at all — reflectSettings() exposes only enabled, intervalMs, byteTriggerPct, sectionTrigger, recentKeep, minBytes. Grepping the bundle for condenseProvider, condenseCommand, reflectCommand, memoryCommand returns zero occurrences.

Our situation

Config:

"defaultCommand": "claude",
"godProvider":   "opencode",
"godModel":      "local/deepseek-v4.1-flash:cloud"

Every non-archived agent has provider: opencode. Live process check: 5–7 opencode --model … processes, zero claude processes. Yet condense runs a hidden Claude session on a ~30-minute timer, and has been failing hourly:

condense-abort  reason=summarize-failed
detail: Error: condense: response contained no parseable JSON

Why it fails (the mechanism)

summarize() requires strict JSON:

Produce STRICT JSON: {"condensed": "<text>", "hoist": ["<line>", ...]}.

When the Claude subscription hits its weekly limit, claude -p returns plain text:

You've hit your weekly limit · resets Sep 25, 6am (UTC)

That is not JSON, so parseSummary() returns nothing and the call throws. Verified directly — both claude-haiku-4-5 and claude-sonnet-5 return the limit line (subscription pro).

Note: the limit is the symptom. Even with quota available, this path still (a) burns Claude usage on a floor that was deliberately moved to OpenCode, and (b) spends a hidden invocation every cycle it cannot succeed.

Why this matters

1. It defeats the purpose of the OpenCode deployment. Operators move to OpenCode/self-hosted models specifically to get off the Claude subscription. Condensation is the one piece that never made the move, and it is the piece that runs unattended on a timer.
2. It fails silently. Nothing in fleet.json distinguishes "condensing is failing" from "condensing is idle", and the abort shares its log line with the unrelated JSON-escaping failure in Memory condensation aborts ~80% of the time: the payload has to survive JSON string escaping #459 — three distinct causes collapse into no parseable JSON.
3. It is not configurable away. With defaultCommand: "claude" the binary is Claude; pointing defaultCommand at opencode does not help either, because the model (claude-haiku-4-5) and the flags (--permission-mode, --disallowedTools) are Claude-specific. There is no supported way to say "condense with my configured provider".

Expected

• The reflector should resolve the provider of the agent being condensed (or at minimum a configurable reflectProvider / reflectModel / reflectCommand), and
• build the arguments from that provider's preset (providerPreset() already exists in the bundle and is used for spawning) instead of hardcoding Claude flags, and
• detect a limit/quota response and record it as its own reason (e.g. rate-limited with the reset time) rather than retrying blind.

Related

• Memory condensation aborts ~80% of the time: the payload has to survive JSON string escaping #459 — "Memory condensation aborts ~80% of the time". Same subsystem; our failure is a different mechanism (see my comment there).
• Spawn request with provider but no command launches the wrong binary with another provider's flags #462 — "Spawn request with provider but no command launches the wrong binary with another provider's flags" — same class of bug (provider ignored, flags assumed to be Claude's), but on the spawn path. The reflector is the unattended twin of this.
• First-class provider support for subscription CLIs: OpenCode Go and Command Code (GOAT) #545 — provider support for subscription CLIs; relevant if the reflector is to become provider-aware for more than Claude/OpenCode.

Happy to test a patch — we have a reproducible floor and log evidence.