Summary

contextTokens is a static config value used as a display denominator for "Context: 45k / 200k (22%)" in status output. The lookup functions that originally resolved it from model metadata are stubbed out (lookupContextTokens, DEFAULT_CONTEXT_TOKENS), making the denominator a guess (defaults to 200k). The actual token counting (input/output/cache) from CLI runtimes works correctly and should be preserved.

What to Remove

• contextTokens field from SessionEntry (src/config/sessions/types.ts)
• contextTokens field from AgentDefaultsConfig and AgentConfig (src/config/types.agent-defaults.ts)
• contextTokens from Zod schema (src/config/zod-schema.agent-defaults.ts)
• contextTokens parameter from deriveSessionTotalTokens() (src/agents/usage.ts) — it accepts it but never reads it
• lookupContextTokens stub and DEFAULT_CONTEXT_TOKENS stub in agent-runner.ts
• contextTokensUsed resolution chain in agent-runner.ts
• contextTokens passthrough in session-store.ts, session-usage.ts, get-reply-run.ts, get-reply-directives.ts, get-reply-directives-apply.ts
• contextTokens in cron runner (src/cron/isolated-agent/run.ts:424-428)
• Percentage display logic in formatTokens() and related formatting (src/auto-reply/status.ts, src/commands/sessions.ts, src/commands/status.summary.ts, src/commands/status.format.ts)
• contextTokens from gateway session utils (src/gateway/session-utils.ts)
• contextTokens from TUI types and session actions (src/tui/)
• contextTokens from UI types and presenter (ui/src/ui/)

What to Keep

• totalTokens, inputTokens, outputTokens, cacheRead, cacheWrite in session entries — these are real usage data from CLI runtimes
• deriveSessionTotalTokens() function (without the unused contextTokens param)
• hasNonzeroUsage() and normalizeUsage() — real usage utilities
• Token display in status, simplified to raw count: "Tokens: 45k" instead of "Context: 45k / 200k (22%)"

Acceptance Criteria

• Given the codebase
  When searching for contextTokens in source files
  Then no matches found (except changelog/git history)

• Given remoteclaw status output
  When a session has token usage
  Then raw token count is displayed (no percentage against a static denominator)

• Given agents.defaults.contextTokens in a config file
  When the config is loaded
  Then the field is silently ignored (Zod strips unknown fields) or explicitly warned

• pnpm check passes

• Existing tests pass (or updated to remove contextTokens assertions)

Rationale

The contextTokens denominator is misleading: it defaults to 200k regardless of actual runtime context window (Gemini has 1M, some Claude models have 1M). The stubs (lookupContextTokens → undefined, DEFAULT_CONTEXT_TOKENS → undefined) mean the resolution chain falls through to hardcoded 200k. Removing the misleading percentage is better than displaying an inaccurate one.

If runtime-reported context windows are wanted later (for proactive session health notifications), the token counting infrastructure survives — only the static config denominator is removed.

Depends On

None — independent of the provider ID unification chain (#2125-#2128). Can be executed in parallel.