E2E fleet sessions run at a 200K context window despite the harness's 1M setup #5274

Fix target and integration order — mandatory routing

Repository: TalonT-Org/AutoSkillit.
Fix PR base branch: develop, currently 381afb4e59cc23acf4ca9ace7a7644b75e6e7b14.
Scope: how E2E fleet sessions get their context window. The harness settings, the README text and the Claude Code version pin involved are identical on develop and on the #5255 branch, so this does not need the unmerged implementation E2E harness.

Open the remediation PR against develop. Once merged, the owner of parent PR #5255 integrates the develop commit.

Defect

The E2E harness is set up for a 1M-token context window, but fleet sessions run with a 200K window. The root therefore auto-compacted below 200k tokens in both retained implementation runs.

Evidence

Run 38431354139, attempt 1 at 93457f16ecc30249799c02084c20f41d9cd88455. Artifact e2e-implementation-38431354139-1.

What the harness configures (scripts/e2e/e2e_harness.py:26-27 and :128-132 on develop):

• ANTHROPIC_MODEL and ANTHROPIC_DEFAULT_{SONNET,OPUS,HAIKU}_MODEL are MiniMax-M3[1m].
• CLAUDE_CODE_AUTO_COMPACT_WINDOW is 1000000.
• scripts/e2e/README.md:74-75 says the [1m] suffix "makes Claude Code assume a 1M-token window for the unrecognized model id; the compact window is capped at it".

What fleet sessions actually run as.

• AutoSkillit passes an explicit model id, not the alias: CLAUDE_MODEL_ALIASES maps sonnet to claude-sonnet-5 (src/autoskillit/core/types/execution/_type_backend.py:348-349). The README states this too at :81 ("the explicit claude-* model ids AutoSkillit passes with --model").
• logs/sessions.jsonl: all nine sessions have configured_model: "sonnet", provider_used: "anthropic", profile_name: "".
• Claude Code's own output for two worker sessions (logs/sessions/fcaa0a53-…/raw_stdout.jsonl, logs/sessions/3a50360d-…/raw_stdout.jsonl): the init record has model: "claude-sonnet-5", and the result record has modelUsage: {"claude-sonnet-5": {"contextWindow": 200000, "maxOutputTokens": 64000}}. The assistant messages report MiniMax-M3.
• So the [1m] suffix, which sits on the custom id MiniMax-M3[1m], only affects sessions started without --model. Fleet sessions are sized as claude-sonnet-5.

Consequence.

• The root's raw output is not retained, but it was launched the same way and auto-compacted at preTokens 194567 (root transcript record 382). The root in run 38170521660 compacted at 166958.
• The root's 64,000-token output at transcript record 376 matches the maxOutputTokens above.

What the Claude Code documentation says (https://code.claude.com/docs/en/model-config and /env-vars, retrieved 2026-10-13):

• CLAUDE_CODE_AUTO_COMPACT_WINDOW: "The effective window is also capped at the model's context window."
• "Fable 5.1, Fable 5, Sonnet 5 and later, Haiku 5.5, and Opus 4.7 and later run with the 1M window by default, with no [1m] suffix needed."
• Behind a custom ANTHROPIC_BASE_URL, "Claude Code gives each model it recognizes the same context window the model has on the Anthropic API."
• "To keep sessions at a 200K window, set CLAUDE_CODE_DISABLE_1M_CONTEXT=1".

By that documentation claude-sonnet-5 should have had a 1M window here. Nothing in this repository sets CLAUDE_CODE_DISABLE_1M_CONTEXT, CLAUDE_CODE_MAX_CONTEXT_TOKENS or autoCompactWindow. The image pins Claude Code 2.1.280 (scripts/docker/Dockerfile:15).

Not established

• Why Claude Code 2.1.280 reported a 200K window for claude-sonnet-5 in this environment.
• Whether AutoSkillit sessions on the first-party API are sized the same way. If they are, this is wider than E2E.
• What context limit the MiniMax endpoint really enforces.

Required behaviour

• E2E fleet sessions run with the window the harness documents. Assert it from modelUsage.*.contextWindow in a session result, not from configuration.
• Correct the README so it says which sessions each setting covers.
• First step: in the E2E image, run one claude -p --model claude-sonnet-5 --output-format json call with and without the MiniMax base URL and read modelUsage.*.contextWindow. That separates a Claude Code behaviour from a harness configuration gap.
• Run pre-commit run --all-files and the configured task test-local-gate.

Coordination

Related to #5272, which requires that a fleet root does not auto-compact. With compaction blocked, a 200K window turns a compaction into a hard stop near 200k tokens, so the window still matters. Independent of #5270 and #5271; can be worked in parallel.