provider agents: per-agent env injection for REPL sessions (small-fast model + CC traffic flags) — alias models are unreliable through the sol proxy

[olegbrok]

Summary

Provider-backed agents (#sol lane) can only inject ANTHROPIC_BASE_URL / ANTHROPIC_API_KEY / ANTHROPIC_AUTH_TOKEN into their REPL env (tmux_session._build_repl_env, mirrored by the SDK path). The proven sol recipe requires more: ANTHROPIC_SMALL_FAST_MODEL=gpt-5.6-luna[1m], CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC=1, CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK=1. Without them, CC's small-fast slot and any subagent with a model: alias resolve to claude-* names, which are only ~17-70% reliable through the codex backend (vs ~99% for gpt-5.6-sol).

Evidence (2026-07-12 night, solik on tmux via claude-code-proxy :18765)

Proxy log totals since install, /v1/messages only:

model	200	401
claude-sonnet-5	1993	9642
claude-opus-4-8	576	790
claude-haiku-4-5	330	144
gpt-5.6-sol	~all	a handful

Concrete incidents: solik's audit subagent fan-out (sonnet alias) died 08:38Z; a single opus-alias review subagent burned 1000+ failed calls 09:56→10:00Z before being killed. Interim mitigation: behavioral directive to solik — never pass a model override to subagents.

Ask

Per-agent provider env injection — either a provider_env JSON column on agents merged into _build_repl_env() (tmux) and the SDK session's provider_env, or first-class columns for small-fast model + the two disable flags. Registry-driven, so the flip survives respawns and applies to future provider agents (Luna fleet, Ryzhik/GLM). Sol phase-2 ("Luna small-fast") is blocked on exactly this seam.

Related: #868/#869/#870 (tonight's tmux-transport series).

🤖 Opened by Barsik

[olegbrok]

Third trap instance tonight (03:58): a loaded skill (/simplify) MANDATES four Agent calls with explicit model fields — agent killed them but lost the skill's value entirely. Behavioral directives can't cover skills whose instructions hardcode models. Strengthens the case for the alternative fix layer: model remap at the proxy (claude-sonnet-5/claude-opus-4-8/claude-haiku-* → gpt-5.6-sol / gpt-5.6-luna upstream) so every alias request from any source (small-fast slot, typed agents, skill fan-outs) lands on the reliable models — no per-agent env or behavioral rules needed. I maintain the local proxy patches; will scope the remap during the day.

🤖 Opened by Barsik

[olegbrok]

ROOT CAUSE FOUND AND FIXED — supersedes both earlier theories in this issue. It was never model flakiness: the proxy's codex token refresh had NO single-flight. Concurrent requests at the expiry margin each refreshed with the same single-use rotating token; losers got 401 and their error path cleared the winner's freshly saved tokens (clear_auth), producing minutes-long all-401 windows. The claude-alias correlation was pure sampling bias — alias models are what subagent fan-outs request, and fan-outs are what create concurrency at expiry time. Discriminating experiments: sequential AND 8-way-parallel probes on claude-sonnet-5 = 100% success outside windows; proxy already remaps aliases (sonnet→terra, opus→sol, haiku→luna), so the remap idea in earlier comments was moot.

Fixed in the fork (single-flight + rotated-token preservation, 2 unit tests; grok provider already had this exact guard — codex now at parity), deployed to the Mini 07:50, upstream PR raine/claude-code-proxy#42. Old binary kept as .bak-pre-singleflight for rollback.

Remaining value in THIS issue: per-agent env injection is still worth building (small-fast→luna saves sol tokens; CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC per the proven recipe) — but it's an optimization now, not an outage fix. Downgrading urgency. After a day of soak, solik's no-subagent-model-override directive can be relaxed (aliases are properly remapped and were never the problem).

🤖 Opened by Barsik

[olegbrok]

Final layer (2026-07-12 09:39-09:55 PT): the storm that recurred AFTER the single-flight fix was upstream revocation — OpenAI invalidated the proxy's OAuth grant entirely (Keychain auth cleared by the legitimate refresh-401 path; every model 401'd, sol included). Recovery: device re-auth (Brad approved 09:55), 24/24 probes green since. Two deployed mitigations remain in the fork: single-flight refresh (raine#42) and all-aliases→gpt-5.6-sol remap (the earlier terra/luna 'entitlement flap' reading is now uncertain — likely stages of the dying grant — but the remap is harmless and keeps skill fan-outs on the proven model). Watch-item: whether revocation recurs without refresh churn (if yes, it's ToS-flag territory per the original risk assessment, and Brad decides the lane's future). Codex CLI lane (Kuzya/Murzik, separate auth.json) untested since revocation — verify on their next use.

🤖 Opened by Barsik