fix(kiro): GPT-5.6 context windows follow Kiro's 1M listing

devlog/_plan/260926_kiro_lb_parity2/081_head_to_head_result.md

@@ -14,7 +14,7 @@ now survives restart and is fenced to the login that measured it (kiro-lb keys b
 fences a re-login), refresh stays leased where kiro-lb falls back to an unleased refresh, credentials are
 never reloaded from an unrelated source, region values are validated, and polling is on demand; load
 spreading is now at parity (deterministic, opt-in least-loaded against kiro-lb's weighted race or
-deterministic most-credits mode). Kiro-lb still leads on eight axes listed
+deterministic most-credits mode). Kiro-lb still leads on six axes listed
 below. Each is a deliberate choice or needs live evidence we do not have, so **the goal criterion "ahead on
 every compared axis" is not met as written**; "ahead or at parity on every adopted axis" is.
 
@@ -36,7 +36,7 @@ every compared axis" is not met as written**; "ahead or at parity on every adopt
 | P9 | Served success clears a verdict | `kiro/account_manager.py:1411` | `src/providers/kiro-usage.ts:344`; `src/server/responses/adapter-delivery.ts:125` | parity | `tests/providers/kiro/kiro-refusal-failover.test.ts` "kiro later served completion clears older exhaustion but not a newer refusal or different identity" |
 | E5 | Provider egress | `kiro/proxy_chain.py:176-198` | `src/adapters/kiro-retry.ts:169-196` | parity | `tests/providers/kiro/kiro-transport-parity.test.ts` "initial, reset, gateway alternate and 429 recovery use the supplied executor" |
 | E2/H | 502/503/504 rotation | `kiro/http_client.py:603-653` | `src/adapters/kiro-retry.ts:290-296` | parity | `tests/providers/kiro/kiro-transport-parity.test.ts` "canonical HTTP %i rotates once and final failure is fixed text", "500, 521 and custom URLs do not rotate" |
-| E6 | Timeout handling | `kiro/network_errors.py:65-145` | `src/adapters/kiro-retry.ts:200` | parity on header deadline → 504; ahead-lb on the connect/read split | `tests/providers/kiro/kiro-transport-parity.test.ts` "header deadline becomes 504 without rotating; caller abort preserves its reason" |
+| E6 | Timeout handling | `kiro/network_errors.py:65-145` | header deadline `src/adapters/kiro-retry.ts:191`; body inactivity `src/server/responses/adapter-delivery.ts:79-99`, `src/server/responses/adapter-continuation.ts:666-687` | parity (equivalent bound: our header deadline covers connect plus first byte, kiro-lb's covers connect only; read inactivity bounded on both sides) | `tests/providers/kiro/kiro-transport-parity.test.ts` "header deadline becomes 504 without rotating; caller abort preserves its reason"; `tests/server/terminal-guard-server.test.ts` "a stalled initial body fails with a 504 upstream error instead of a proxy error", "a stalled continuation body reports 504 even when cancelling it aborts the client signal" |
 | E7 | Fixed public 5xx text | `kiro/exceptions.py:27-36` | `src/adapters/kiro-errors.ts:215` | parity | `tests/providers/kiro/kiro-transport-parity.test.ts` "upstream 5xx marker reaches neither client nor debug ring or stderr"; `tests/server/server-kiro-refusal-e2e.test.ts` "unrotated Kiro 5xx retains fixed public text after raw-error handoff" |
 | E1 | Extra endpoint dialects | `kiro/endpoints.py:25-82` | runtime host plus `q.*` | ahead-lb | Reject holds: kiro-lb itself verifies only the runtime host (`kiro/endpoints.py:9-12`). |
 | E4 | Endpoint probe | `kiro/endpoint_probe.py:2-8` | absent | ahead-lb | Reject holds: the probe spends real credits. |
@@ -51,7 +51,7 @@ every compared axis" is not met as written**; "ahead or at parity on every adopt
 | U1 | Multiplier estimates | `kiro/model_costs.py:8-22` | none | ahead-lb (advisory) | Reject holds: a second, weaker number beside measured credits. |
 | C1/N | Per-account catalogue | `kiro/model_catalog.py:38-78` | `src/providers/kiro-model-catalog.ts:12,120` | parity | `tests/providers/kiro/kiro-model-catalog.test.ts` "management ListAvailableModels pairs each bearer with its own profile and region", "unrecognised or empty management replies preserve the last good list" |
 | C2 | Membership as evidence | `kiro/model_resolver.py:309-355` | `src/oauth/generic-account-failover.ts:237` | parity | `tests/providers/kiro/kiro-model-preference.test.ts` "reactive Kiro rotation prefers model evidence after room filtering", "no catalogue evidence never moves a healthy active account" |
-| C3 | Context limits | `kiro/model_costs.py:64-74` | `src/providers/kiro-model-catalog.ts:129`; `src/adapters/kiro/usage.ts:223`; `src/server/request-log.ts:1633` | parity when catalogue evidence exists; ahead-lb on the static fallback (272k GPT-5.6 window at `src/providers/kiro-models.ts:34` vs kiro-lb's 1M) | `tests/providers/kiro/kiro-model-catalog.test.ts` "a mixed known/unknown roster never reports more than the smallest known window" |
+| C3 | Context limits | `kiro/model_costs.py:64-74` | static `src/providers/kiro-models.ts:31-37` (GPT-5.6 1M per kiro.dev/docs/models, updated in wp10); observed minimum `src/providers/kiro-model-catalog.ts:129`; consumers `src/adapters/kiro/usage.ts:223`, `src/server/request-log.ts:1633` | parity (static window matches Kiro's page and kiro-lb; observed catalogue evidence preferred when present) | `tests/providers/kiro/kiro-adapter.test.ts` "1M-context models map to 1_000_000"; `tests/providers/kiro/kiro-model-catalog.test.ts` "a mixed known/unknown roster never reports more than the smallest known window" |
 | W1/W2 | IDE fingerprint | `kiro/utils.py:20-31,74-85` | `src/adapters/kiro/wire.ts:8` (`KIRO_IDE_VERSION = "1.0.0"`) | ahead-lb | Reject holds: needs our own capture. |
 | S2 | MCP web search | `kiro/mcp_tools.py:131-150` | none | n/a | Reject holds: a new feature, not auth or usage. |
 | K | Quota metrics | `kiro/metrics.py:262` | `src/providers/kiro-quota-metrics.ts:16`; `src/server/request-metrics.ts:192-196` | parity | `tests/providers/kiro/kiro-quota-metrics.test.ts` "fresh cached credits yield four opaque gauges without probing", "future, stale, reset-passed, and identity-mismatched rows are omitted" |
@@ -60,6 +60,9 @@ every compared axis" is not met as written**; "ahead or at parity on every adopt
 
 ## Where kiro-lb still leads
 
+Updated in wp10 (`090_context_and_timeout_corrections.md`): the C3 static window and E6 timeout rows moved to parity.
+
+
 | Axis | Why we did not follow (yet) |
 |---|---|
 | W1/W2 IDE wire fingerprint | Changing the wire identity without our own capture risks every Kiro request. Needs a live capture. |
@@ -69,8 +72,6 @@ every compared axis" is not met as written**; "ahead or at parity on every adopt
 | U1 per-model credit estimates | Measured credits (`providerCredits`) are the source of truth; an estimate beside them is a weaker second number. |
 | Dashboard device-login UI | Native device login ships in the CLI and management API (060, option a); the dashboard still uses kiro-cli. A device-code dialog is the follow-up. |
 | Operations dashboard | Routable state and quota gauges are in the API, CLI and metrics export (070), not rendered in the GUI; the `health` field does not yet reflect Kiro suspension/exhaustion. |
-| C3 static context fallback | Without catalogue evidence our static GPT-5.6 window is 272k (`src/providers/kiro-models.ts:34`) where kiro-lb records 1M; the observed catalogue raises it when an account reports it. |
-| E6 connect vs read timeout split | A header deadline maps to 504 and never rotates on caller abort, but connect and read timeouts are not separated. |
 
 ## Narrowed by recorded decisions (not failures)
 

devlog/_plan/260926_kiro_lb_parity2/090_context_and_timeout_corrections.md

@@ -0,0 +1,55 @@
+# 090 — Context-window and timeout corrections after the head-to-head
+
+Work phase wp10. Branch `codex/kiro-lb2-081-corrections` from dev `5518653a9a`. This is the single
+executable plan (rewritten after audit; no later section overrides it).
+
+## Evidence
+
+- **C3 static window.** Kiro's model page (https://kiro.dev/docs/models/, fetched 2026-09-27; the page's
+  dateModified is 2026-09-25) lists GPT-5.6 Sol, Terra and Luna with a **1M** context window and states
+  two-tier pricing: requests up to 272K tokens bill at the listed rate, larger requests at double. Our
+  static table (`src/providers/kiro-models.ts:31-38`, cited as "page updated 2026-07-14") records 272K for
+  the three GPT-5.6 tiers. `gpt-6-sol`/`gpt-6-luna` are not on Kiro's page and are unchanged.
+- **E6 timeouts.** The Kiro header deadline is applied at `src/adapters/kiro-retry.ts:191` (020), and
+  body inactivity is bounded for every adapter stream — initial stream `src/server/responses/adapter-delivery.ts:79-99`,
+  continuation legs `src/server/responses/adapter-continuation.ts:666-687` — each mapping a stall to 504.
+  Verdict: **equivalent bound**. Difference recorded: our header deadline covers connect plus waiting for the
+  first byte; kiro-lb's connect timeout covers connect alone.
+
+## Consumers of `KIRO_MODEL_CONTEXT_WINDOWS` and what changes
+
+- Registry seed `src/providers/registry/entries-core.ts:610` → routed catalogue, Codex sync, Claude context windows.
+- `kiroObservedContextWindow` (`src/providers/kiro-model-catalog.ts:140`) → static floor for the **minimum
+  observed window across live accounts**; that shared minimum feeds catalogue display, the adapter's
+  context-percentage conversion and calibration, and request-log estimate capping.
+- `kiroUpstreamContextWindow` (`src/adapters/kiro/usage.ts:224`) → adapter `contextWindow`
+  (`adapter.ts:292`, `stream.ts:630`); `stream.ts:366-367` computes
+  `contextTotalTokens = window × contextUsagePercentage`, so reported context tokens for a GPT-5.6 turn
+  scale with the window.
+- `request-log.ts:1634` → estimated-token cap.
+
+## Change map
+
+| Path | Change |
+|---|---|
+| `src/providers/kiro-models.ts` | MODIFY: citation updated to the 2026-09-25 page; `gpt-5.6-sol`, `gpt-5.6-terra`, `gpt-5.6-luna` → `1_000_000`; one comment noting the 272K pricing threshold. |
+| `tests/providers/kiro/kiro-adapter.test.ts` (capped 2050) | MODIFY, net −3 lines: append the three GPT-5.6 ids to the 1M loop list (line ~1770) and delete the three `272_000` expectations (lines ~1778-1780). |
+| `devlog/_plan/260926_kiro_lb_parity2/081_head_to_head_result.md` | MODIFY: C3 → parity (static 1M now matches Kiro's page and kiro-lb; observed catalogue evidence still preferred); E6 → equivalent bound, anchors above, tests `tests/server/terminal-guard-server.test.ts` "a stalled initial body fails with a 504 upstream error instead of a proxy error" and "a stalled continuation body reports 504 even when cancelling it aborts the client signal", plus `tests/providers/kiro/kiro-transport-parity.test.ts` "header deadline becomes 504 without rotating; caller abort preserves its reason"; remove both from "Where kiro-lb still leads". |
+| docs-site | **No change**: no page states a Kiro GPT-5.6 window (the 272K mentions are the native Cursor guide). |
+
+## Risks
+
+- **Billing:** a client that plans against the larger window may send >272K-token Kiro GPT-5.6 requests,
+  which Kiro bills at double. The per-provider context cap only lowers the window clients plan against; it is
+  a planning aid, not a hard upstream limit or a billing guarantee. This is Kiro's published contract and
+  matches kiro-lb.
+- **Context accounting:** correct if Kiro's `contextUsagePercentage` is relative to its documented 1M window;
+  not verified live. The observed minimum across accounts still overrides the static value when present.
+
+## Verifiers
+
+`bun run typecheck`; `bun test tests/providers/kiro/` (includes `kiro-adapter`, `kiro-stream` context
+calibration, `kiro-model-catalog`); `tests/server/terminal-guard-server.test.ts`;
+`tests/usage/request-log-estimate-cap.test.ts`; `tests/ci-workflows/file-size-ratchet.test.ts`;
+`bun run privacy:scan`; `bun run structure:check`; the 081 mechanical check. Hosted CI on the PR head.
+

src/providers/kiro-models.ts

@@ -28,12 +28,13 @@ export const KIRO_MODELS = [
 ];
 
 // Per-model context windows as documented on Kiro's official model catalog
-// (https://kiro.dev/docs/models/ — "Quick comparison", page updated 2026-07-14).
+// (https://kiro.dev/docs/models/ — "Quick comparison", page updated 2026-09-25).
 // "Auto" is a router with no fixed window on Kiro's table, so it is intentionally omitted.
+// GPT-5.6 accepts 1M tokens; requests above 272K bill at double the listed rate (two-tier pricing).
 export const KIRO_MODEL_CONTEXT_WINDOWS: Record<string, number> = {
-  "gpt-5.6-sol": 272_000,
-  "gpt-5.6-terra": 272_000,
-  "gpt-5.6-luna": 272_000,
+  "gpt-5.6-sol": 1_000_000,
+  "gpt-5.6-terra": 1_000_000,
+  "gpt-5.6-luna": 1_000_000,
   "gpt-6-sol": 272_000,
   "gpt-6-luna": 272_000,
   "claude-sonnet-5": 1_000_000,

tests/providers/kiro/kiro-adapter.test.ts

@@ -1768,16 +1768,13 @@ describe("kiro adapter — per-model context windows (kiro.dev/docs/models)", ()
   });
 
   test("1M-context models map to 1_000_000", () => {
-    for (const id of ["claude-sonnet-5", "claude-opus-5.5", "claude-opus-5", "claude-opus-4.8", "claude-opus-4.7", "claude-opus-4.6", "claude-sonnet-4.6"]) {
+    for (const id of ["gpt-5.6-sol", "gpt-5.6-terra", "gpt-5.6-luna", "claude-sonnet-5", "claude-opus-5.5", "claude-opus-5", "claude-opus-4.8", "claude-opus-4.7", "claude-opus-4.6", "claude-sonnet-4.6"]) {
       expect(kiro.models ?? []).toContain(id);
       expect(cw[id]).toBe(1_000_000);
     }
   });
 
   test("smaller-context models match Kiro's published limits", () => {
-    expect(cw["gpt-5.6-sol"]).toBe(272_000);
-    expect(cw["gpt-5.6-terra"]).toBe(272_000);
-    expect(cw["gpt-5.6-luna"]).toBe(272_000);
     expect(cw["claude-opus-4.5"]).toBe(200_000);
     expect(cw["claude-sonnet-4.5"]).toBe(200_000);
     expect(cw["claude-sonnet-4.0"]).toBe(200_000);