CLAUDE_CODE_MAX_CONTEXT_TOKENS is real but conditionally honored

[jellologic]

Resolves the open question in #2

CLAUDE_CODE_MAX_CONTEXT_TOKENS is real — but it is conditionally honored, which is why it never appeared in the official env-vars reference.

Verified against the shipping binary, ~/.local/share/claude/versions/2.1.216 — not against docs, and not against the leaked v2.1.88 source (see #17 for why that distinction matters).

The two read sites

function a9c(){                                  // path A
  if(Z.DISABLE_COMPACT){                         // ← only when DISABLE_COMPACT is set
    let e=Z.CLAUDE_CODE_MAX_CONTEXT_TOKENS;
    if(e!==void 0&&e>0)return e
  }
  return
}

function l9c(e,t){                               // path B — default resolver
  if(Tb(e))return 1e6;                           // model id matches /\[1m\]/i
  if(t?.includes(Ege.header)&&RG(e))return 1e6;  // context-1m-2025-08-07 beta
  if(eM(e))return 1e6;                           // native_1m, first-party gated
  let r=b7n(e); if(r!==null)return r;
  let n=Z.CLAUDE_CODE_MAX_CONTEXT_TOKENS;
  if(n!==void 0&&n>0&&!po(mi(e)).startsWith("claude-"))return n;  // ← non-"claude-*" ONLY
  return EYt                                     // EYt = 200000
}

Constants: EYt = 200000, qtg = 1e6 (cap).

What this means for us

Model ID	DISABLE_COMPACT	Effect
glm-5.2, qwen-…, openrouter/…	unset	Honored — doesn't start with claude-
claude-opus-4.8	unset	Silent no-op — exactly the failure mode #2 warned about
anything	set	Honored via path A

Since cuckoocode's third-party presets use non-claude-* model IDs, this variable does work for our primary use case. It is not a no-op for us.

Note Tb(e) — a [1m] suffix returns 1e6 directly, confirming the mechanism in #14 independently of this variable.

Acceptance criteria

• Set CLAUDE_CODE_MAX_CONTEXT_TOKENS only when the configured model ID does not start with claude-
• Never set it for the first-party Anthropic provider
• Value drawn from catalog context_length, capped at 1e6
• Applied in addition to the [1m] suffix (BUG: z.ai preset silently runs at 200K context instead of 1M #14), never instead
• A code comment records the condition, so nobody later "simplifies" it into a no-op
• Test covering both branches: a claude-* id (must not be set) and a glm-* id (must be set)

Conflict note

#2 currently instructs implementers to omit this variable unless confirmed. That guidance was correct when written and is now superseded. Anything already built under the old guidance is conservative rather than wrong — it just leaves the fix incomplete.

[jellologic]

Closing — the conclusion is realised in the code. swisscode uses the per-variable [1m] model-id suffix (issue #14, closed) rather than CLAUDE_CODE_MAX_CONTEXT_TOKENS, which this investigation showed is only conditionally honored (the DISABLE_COMPACT path, or non-claude-* models). These binary-derived findings (verified against 2.1.216) informed #2/#14 and stay recorded here. Reopen if a future case needs the conditional variable.