Devin plan-quota awareness, zero-write watchdog, and advise task-shape routing (0.12.3)

[alexgreensh]

What

On a long Devin Pro session the shared daily plan quota can be exhausted, which blocks every plan-included model (glm/swe/kimi) at once. Outsourcerer modeled only the paid ACU pool, so it read that exhaustion as a "mis-gate, not a real limit" and pointed the user back at the same dead lane. Separately, advise recommended a pure-reasoning model for edit-run-verify work and had no glm-5.3-flash-high to prefer.

Changes

Devin plan quota

• Detect shared daily/weekly plan-quota exhaustion (distinct from a paid ACU refusal); mark the dv lane down until Devin's own stated reset; steer off Devin (OpenRouter via --provider cc, or a native lane) instead of another Devin plan model.
• Parse Devin's resets in 11h26m / resets at 09:00 UTC wording into the real lane-down TTL.
• Guard the dead devin usage call (the current CLI has no such subcommand); hints point at the usage dashboard.
• Per-session Devin-plan dispatch meter warns past OSRC_DEVIN_PLAN_JOBS_WARN that glm/swe/kimi/deepseek share one daily bucket.

Zero-write watchdog (lane-agnostic)

• A live-but-writeless run (a model spending its budget on reasoning without touching a file) is surfaced as no-progress-writes past OSRC_NOPROGRESS_SECS, mutating verbs only; read-only and text-only lanes exempt. Surface-only unless OSRC_NOPROGRESS_KILL_SECS is set.

advise

• glm-5.3 family incl. glm-5.3-flash-high in the model table, bench map, and Devin resolver.
• Edit-run-verify tasks classify as agentic; Kimi K3's near-frontier bump is reasoning-shaped only, so it no longer floats to Fix: cc-lane OpenRouter exhaustion (402) now self-heals to Devin for dual-lane models #1 on edit/agentic work (a hard pure-reasoning task still picks it).
• Recommendation prefers the cheaper capable variant within a family (flash-high over -high) unless --effort max; OSRC_ADVISE_VARIANT_PREF=0 disables.

Scope

Plan-limit detection + failover is Devin-specific in this release. Generalizing to the other subscription harnesses (Warp, Droid, Cursor, Cline, native) is tracked as follow-up. Proactive conservation (Claude/ChatGPT windows) and read-only transport-fallback are unchanged; the zero-write watchdog and advise changes are lane-agnostic.

Tests

Three new suites (test_devin_plan_quota, test_nowrite_watchdog, test_advise_task_shape) plus an updated test_model_selection_parity. Full conformance: 141 passed, 0 failed. Version parity green (0.12.3).