Claude Opus 5: effective context silently capped at ~166K (≈83% of 200K) despite 1M catalog/blog claims

[q8247990]

Summary

In a 24-hour OpenCode session routed through this plugin (kiro provider, model claude-opus-5, 806 step-finishes, zero 4xx errors), every single API call reported input_tokens ≤ 166,512 even though the local conversation grew to ~498K tokens by session end. Growth was asymptotic: ~100K at 39% of steps, ~130K at 65%, ~150K at 83%, ~160K at 93%, and a final plateau of 166,512 at step 805/806. The model only ever saw the most recent ~166K of the conversation — a silent sliding window, not a rejection.

This matters because Kiro's public blog advertises Claude Opus 5 with a full 1M context window (2.2x credit multiplier) and this plugin's model-registry mirrors that as limit.context: 1000000. OpenCode therefore believes it has ~834K headroom of free space and never auto-compacts. On long sessions the model is effectively amnesic beyond the recent ~166K tail while the UI context bar reads ~16%, which looks healthy.

Evidence

From one 24h / 806-step-finish OpenCode session on kiro / claude-opus-5, 2026-09-02 09:48 → 2026-09-03 10:08 UTC:

Metric	Value
API-reported input_tokens ceiling across all 806 requests	166,512 (no 4xx, no warnings)
cache_read_input_tokens on every request	0
Crossed 100K at	39% of steps (≈ step 314)
Crossed 130K at	65% of steps (≈ step 524)
Crossed 150K at	83% of steps (≈ step 669)
Crossed 160K at	93% of steps (≈ step 750)
Reached 166,512 at	step 805 / 806 (hard plateau)
Model switch to a different provider at 10:09: first request input_tokens	498,086 (= full local conversation history)

Delta after the switch (498,086 vs ≤166,512) confirms the local conversation had grown to ~498K while opus-5 was being served only the trailing ~166K window.

Where it comes from

Plugin arithmetic (v2.0.0, commit 08f06b4):

inputTokens = round(contextWindow × contextUsagePercentage / 100) − outputTokens
src/plugin/streaming/stream-transformer.ts:301-305

getContextWindowSize returns 200,000 unless the model name contains -1m (src/plugin/models.ts:11-13), and LONG_CONTEXT_MODELS is keyed off the -1m suffix (src/constants.ts:101). claude-opus-5 has no -1m variant in the plugin, and Kiro does not expose a separate long-context alias for Opus 5 (per upstream PR #116).

The reported ceiling fits 200K exactly: 166,512 ≈ 200,000 × 0.833, so the backend's contextUsagePercentage is consistent with a ~200K-class window, not 1M. The plugin's reported 166K is therefore a faithful fraction of 200K; the misleading part is the advertised limit.context: 1000000 that OpenCode reads from the model-registry, which prevents auto-compaction and makes the ~84% utilization look like "lots of headroom."

Why catalog discovery (PR #121) may not fix this

If Kiro's catalog reports 1_000_000 for claude-opus-5 (matching the public blog claim), then catalog-based discovery would confirm the 1M headroom while the CodeWhisperer/Q layer continues to serve a ~200K window. The silent truncation would continue undetected.

Suggested checks before trusting a catalog-derived 1M for opus-5:

1. Cross-check the catalog value against observed contextUsagePercentage behavior on a long session — if the plateau divides back to ~200K, the catalog claim is wrong or the long-context path is gated.
2. Until Kiro confirms 1M is actually served end-to-end, pin claude-opus-5 to contextWindow: 200000 so clients can compact appropriately (mirroring Bedrock's 1M beta requirement, see #16486).
3. If a long-context path does exist, surface it as a separate model id or beta flag so users opt in knowingly — same shape as AWS Bedrock's context-1m-2025-08-07 beta.

Comparable reports

• anomalyco/opencode#45168 — same failure pattern on another provider: input pinned at exactly 196,608 = 262,144 − 65,536, cache_read=0 on every subsequent request, silent cost blowup.
• anomalyco/opencode#12452 and #12507 — Bedrock Claude: 1M context requires the context-1m-2025-08-07 beta; catalog says 1M but API enforces 200K.
• anomalyco/opencode#16486 — 1M beta flag broke specifically on AWS Bedrock.

Request

1. Confirm whether the CodeWhisperer / Q layer currently serves claude-opus-5 on a ~200K-class window, or whether a 1M path exists behind an opt-in mechanism (separate model id / beta flag / header).
2. If 200K is the current reality, reflect that in the advertised limit.context so OpenCode auto-compacts on long sessions; or, if a 1M path genuinely exists, document how to unlock it (model alias, beta flag, upstream config) so users aren't silently running with a tail-only ~166K.

---

Evidence gathered from local session telemetry (OpenCode SQLite step-finish data) + plugin source inspection (v2.0.0, commit 08f06b4). Happy to share more detail on request.

[q8247990]

Duplicate — created this first with --assignee q8247990, which failed due to repo permissions; clean copy is #128. Closing.