Unsupported `--reasoning-effort` silently resolves to `off` instead of erroring or clamping

[Roderick-Clemente]

Summary

Passing a --reasoning-effort value the target model does not support resolves silently to off (the weakest setting) instead of erroring or clamping to the nearest supported value. Exit code is 0 and nothing is printed. A caller requesting maximum reasoning gets none, silently, indefinitely.

Environment

• droid 0.186.0, macOS (darwin 24.6.0), droid exec mode

Reproduction

claude-haiku-4-5-20251001 advertises reasoning efforts [off, low, medium, high]. xhigh is not in its list.

# requested xhigh -> resolves to off, exit 0, silent
droid exec --model claude-haiku-4-5-20251001 -r xhigh -o json '...'
# resolved: claude-haiku-4-5-20251001 / off

# control: one flag value different, resolves exactly as asked
droid exec --model claude-haiku-4-5-20251001 -r high -o json '...'
# resolved: claude-haiku-4-5-20251001 / high

Resolved effort read from message.reasoningEffort in the session transcript (~/.factory/sessions/), since the result envelope does not carry it.

Expected

An unsupported --reasoning-effort either errors (like an invalid --model does — that fails closed at exit 1 with the valid list) or clamps to the nearest supported value (high).

Actual

Resolves to off, the weakest value — inverting the intent of the flag — at exit 0 with no output. A validator pinned at -r xhigh against a model lacking it runs with reasoning disabled and no signal.

[Roderick-Clemente]

Adding cross-version evidence for issue #4.
The recorded-field behaviour regressed between droid 0.180.0
and 0.186.0. At 0.180.0 the requested --reasoning-effort
value is preserved in the session transcript verbatim
(xhigh recorded as xhigh); at 0.186.0 Phase 0 captured the
same shape silently resolving to off. Both versions return
success / is_error:false / exit 0, so on the user-visible
surface they look identical — the regression is in whether the
platform announces the silent degrade, not whether the
degrade happens. Headline verdict below: "DOES-NOT reproduce
the recorded-field degrade Phase 0 cited at 0.186.0" — but
read with the regression framing in mind. The field's
observable semantics drift across the version line, and any
code that consults reasoningEffort by literal name cannot
treat 0.180 and 0.186 as the same surface.

---

Probe harness

• Three runs from a throwaway cwd /tmp/.../bug-04-reasoning-effort/,
  identical prompt: "Reply with the literal string OK and nothing else."
• Model pinned: --model claude-opus-5
• Autonomy: --auto low
• Format: --output-format json

---

The observable carrier is the session transcript, not the envelope

At 0.180.0 the exec envelope does not carry reasoningEffort.
Phase 0 Probe 8 documented that the field is recorded per
assistant message in the session JSONL under
~/.factory/sessions/<cwd-slug>/<session-id>.jsonl. The
envelope's session_id is the lookup key.

---

Three runs, side-by-side

droid --reasoning-effort	envelope exit / is_error / num_turns	transcript message.reasoningEffort
xhigh (Phase 0 cite)	0 / false / 1	xhigh (passed through)
high (control)	0 / false / 1	high
low (sanity)	0 / false / 1	low

---

What this means for the filed defect

Phase 0 captured xhigh resolving to off in the transcript at
0.186.0. At 0.180.0 the requested value is preserved verbatim.
Two interpretations both explain the differential:

1. A clamp-to-off step was introduced in the runtime between
   0.180 and 0.186 — the bug as filed did not exist at 0.180.0
   and was introduced later.
2. The runtime passes the value through unchanged but the model
   itself reports reasoningEffort=off after the call at 0.186 —
   the recorded value is the model's reply, not the runtime's.

Either way the field's observable semantics differ between
0.180 and 0.186. Code that reads reasoningEffort by literal
name cannot assume the two versions are the same surface.

---

Caveat — model-internal behaviour is opaque

We cannot verify from outside the CLI whether the model
genuinely executes with xhigh effort at 0.180.0, or quietly
ignores the unsupported value. A future probe that attaches a
side-channel signal (latency under load, per-token character
density, internal thinking traces if exposed) could close
that gap; not in this round.

---

What this does not change

The Phase 0 design recommendation — validate
--reasoning-effort against the model's advertised list rather
than trusting silent clamping — is unaffected by this
differential. If anything, this evidence makes the recommendation
stronger: silently-clamping and passing-through-unsupported are
both behaviours worth defending against, and the surface that
exposes them differs across versions.

---

Scope of this evidence

• Three runs at a single CLI version (0.180.0).
• Single model (claude-opus-5). The bug as filed is
  documented in the Phase 0 record as claude-opus-5 as well;
  cross-model behaviour is not asserted here.
• The verdict is "DOES-NOT reproduce the recorded-field degrade";
  if the filed issue is about a different observable, this
  comment is supplementary rather than contradictory.