fix(claude): honor authoritative context windows in generated profiles

[park285]

Summary

• restrict generated Claude roster and self-profile [1m] markers to authoritative effective context windows of at least 1M
• align Claude Sonnet 4.6, Opus 4.6, and Opus 4.7 registry and generated metadata with their 1M source values
• preserve genuine routed [1m] model IDs while honoring provider caps and case-insensitive marker spelling
• preserve an existing [1m] selector when context metadata is missing or non-positive, stripping it only when a positive authoritative window proves sub-1M

Problem

Generated agent profiles reused the main-session auto-context predicate, so a 372K native route could be written with [1m]. The Anthropic registry and generated jawcode metadata also disagreed about Claude 4.6/4.7 windows, allowing valid 1M selectors to lose their marker during catalog-to-profile generation.

The marker is also overloaded: it can be Claude's context-accounting suffix or part of a real routed selector such as Kimi k3[1m]. Unconditional stripping loses the real selector, while unconditional preservation can ignore an effective provider context cap.

Missing metadata is not evidence of a sub-1M context. Catalog-derived selectors that already carry [1m] must therefore keep it unless a positive effective window explicitly proves otherwise.

Fix

• remove the Sonnet 4.6 generator-only 200K override and regenerate the owned metadata artifact
• add the missing 1M Anthropic registry windows and require every configured Anthropic model to have a positive window
• derive Claude 4.6/4.7 generated profiles through fetchProviderModels() and buildClaudeContextWindows() in regression coverage
• make exact selector metadata authoritative before falling back to the stripped base selector
• canonicalize only the trailing [1m] marker for exact lookup, preserving the rest of the model ID and honoring raw exact-key precedence
• distinguish absent or non-positive metadata from a positive sub-1M window, preserving existing markers in the former case
• cover incomplete catalog-derived Claude 4.6/4.7 metadata through both roster and self-clone generation paths, plus explicit 200K marker removal
• keep main-session auto-context behavior unchanged

This absorbs the shared source-of-truth reconciliation requested on #839.

Verification

• JAWCODE_MODELS_JSON=/tmp/opencodex-jawcode-854/opencodex-snapshot.json npx --yes bun@1.3.14 test --isolate tests/claude-agents-inject.test.ts tests/claude-context-windows.test.ts tests/codex-catalog.test.ts tests/provider-registry-parity.test.ts tests/jawcode-metadata-sync.test.ts — 178 passed, 0 failed
• npx --yes bun@1.3.14 x tsc --noEmit — passed
• JAWCODE_MODELS_JSON=/tmp/opencodex-jawcode-854/opencodex-snapshot.json npx --yes bun@1.3.14 scripts/test.ts — 6,963 passed, 9 skipped, 0 failed
• git diff --check origin/dev — passed for the PR-relative diff

Checklist

• Scope stays focused and avoids unrelated cleanup.
• Docs or release notes were updated when needed. (No user-facing configuration or documentation contract changed.)
• Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults. (No security boundary changed.)

Summary by CodeRabbit

• Bug Fixes
  • Improved context-window labeling for Claude models.
  • Ensured only models with authoritative 1M-token support display the [1m] marker.
  • Prevented incorrect markers from appearing on native roster and self-referential models while preserving valid routing labels.
  • Normalized marker formatting and respected configured context limits, including exact-capacity models.
  • Improved handling of existing labels when model context information is unavailable.

[coderabbitai]

📝 Walkthrough

Walkthrough

Claude model metadata now uses registry context windows. Claude roster and self-clone definitions normalize [1m] markers based on context metadata and auto-context settings. Tests cover routed selectors, caps, normalization, and incomplete metadata.

Changes

Subagent context marker handling

Layer / File(s)	Summary
Context-window metadata source scripts/generate-jawcode-metadata.ts	Generated model rows now use each model’s contextWindow value directly.
Context marker resolution src/claude/agents-inject.ts, tests/claude-agents-inject.test.ts	withSubagentContextMarker resolves marked, canonical, and bare selectors. Roster and self-clone models apply or remove [1m] according to context metadata. Tests cover catalog-derived markers, routed selectors, context caps, exact-cap precedence, case normalization, and incomplete metadata.

Estimated code review effort: 3 (Moderate) | ~25 minutes

Possibly related PRs

• lidge-jun/opencodex#839: Modifies Claude context-window metadata and [1m] marker behavior used by this change.

Suggested reviewers: ingwannu, lidge-jun, wibias

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 60.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes the main change: honoring authoritative context windows in generated Claude profiles.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}

[lidge-jun]

Pre-merge review result: not merge-ready yet. The narrow GPT fix direction is correct (gpt-5.6-sol is authoritatively 372k while the synthetic Cursor route is 1M), but the Anthropic side of the metadata contradicts upstream truth:

1. scripts/generate-jawcode-metadata.ts:30 forces both Sonnet 4.6 variants to 200_000, and tests/codex-catalog.test.ts:2216 locks that value — but Anthropic documents Sonnet 4.6 (and Opus 4.7) as 1M-context models. The source-of-truth values need to move together, not just the two literals in this PR. (This is the same reconciliation fix(providers): give Claude 4.6/4.7 their 1M context windows #839 needs; consider coordinating.)
2. The static Anthropic registry (src/providers/registry.ts:216) lists Sonnet 4.6 and Opus 4.7 but omits their context windows, and the PR's [1m] strip-then-consult flow in src/claude/agents-inject.ts:62-64 can regenerate a valid saved 1M selector without its marker when the map is incomplete.
3. The regression test pins the synthetic correction but never passes real catalog-derived Claude 4.6/4.7 metadata through buildClaudeContextWindows, so the inconsistency stays false-green. Please add an end-to-end generated-profile regression.
4. Full cross-platform CI never ran on this head and there is no maintainer approval yet — both are required by merge policy.

Happy to re-review once the metadata source of truth is reconciled and CI runs.

[park285]

Addressed the pre-merge findings in 2f5a20be:

1. Removed the Sonnet 4.6 200K generator override and regenerated both base and [1m] metadata at 1M.
2. Added 1M registry windows for Sonnet 4.6, Opus 4.6, and Opus 4.7, plus a completeness guard for the configured Anthropic roster.
3. Added catalog-derived generated-profile coverage through fetchProviderModels() → buildClaudeContextWindows() → buildClaudeAgentDefs().
4. Added exact-selector, provider-cap, genuine routed [1m], and case-insensitive marker boundary coverage for roster and self profiles.

The branch now includes the latest dev, and the PR-relative diff remains limited to seven files. Local verification on this head: 177 focused tests passed, typecheck passed, generated metadata byte-sync passed, and the full suite completed with 6,913 passed / 9 skipped / 0 failed.

This PR now absorbs the overlapping source-of-truth work requested on #839. Please re-review this head; hosted cross-platform CI and maintainer approval remain the external merge gates.

[coderabbitai]

Actionable comments posted: 1

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@src/claude/agents-inject.ts`:
- Around line 62-71: Update withSubagentContextMarker to preserve an existing
[1m] marker when no positive matching window metadata exists; only strip the
marker when an authoritative positive window indicates a sub-1M context, while
retaining current auto-marking behavior for eligible windows. Add an end-to-end
regression using incomplete catalog-derived Claude 4.6 and 4.7 metadata that
covers both roster and self-clone generation paths.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 01e943b1-2719-4c8c-831f-db41e09ad3e0

📥 Commits

Reviewing files that changed from the base of the PR and between 04459a6 and 2f5a20b.

📒 Files selected for processing (2)

• scripts/generate-jawcode-metadata.ts
• src/claude/agents-inject.ts

[park285]

Follow-up on the CodeRabbit finding:

• withSubagentContextMarker() now treats only a positive context window as authoritative. An existing [1m] marker is preserved when exact/canonical/base metadata is absent or non-positive, and is stripped only when a positive window proves sub-1M.
• The regression derives incomplete Claude 4.6/4.7 catalog rows through buildClaudeContextWindows() and verifies both roster and self-clone generation, plus explicit 200K marker removal.
• The branch now includes dev at f9b9440c and remains limited to the same seven PR-relative files.

Verification on head 463749e6:

• focused suite: 178 passed, 0 failed
• typecheck: passed
• generated metadata byte-sync: passed
• full suite: 6,963 passed, 9 skipped, 0 failed
• git diff --check origin/dev: passed

[lidge-jun]

Superseded by #889, which ports this PR's withSubagentContextMarker design and its full regression set (the modified 372k roster test plus the five additions: catalog-derived markers, routed [1m] preservation, 350K cap unmarking, marker-case precedence, incomplete-metadata preservation) together with #839's registry map, the generated metadata rows, and the codex-catalog/provider-registry-parity test updates. Consolidated as one fix crediting both PRs. Thanks — the port survived an adversarial audit unchanged.

[park285]

특붕이 잘쓰고있어요 유지보수 감사합니다.