fix(llm): populate usage for the Claude Code local-agent path

[N3thunt3r69]

Fixes #139.

LocalAgentAdapter.chat() never set usage on its LLMResponse for the Claude Code local-agent path, so AgentLoop's per-task token budget check always compared against zero and never fired. localAgentChat() now requests --output-format json for the claude agent id only; LocalAgentAdapter parses the real envelope into usage.promptTokens/completionTokens, with a character-based estimate as a fallback if parsing fails. Codex and Hermes are untouched.

Contribution Receipt

• Change: Populate per-call usage for the Claude Code local-agent path (--output-format json) so AgentLoop's token budget check can fire.
• Scope class: local_lab
• Target authority: not_applicable (LLM backbone plumbing only; verification used a throwaway loopback fixture, not an authorized external target)
• Network use: loopback (verification target was a local throwaway HTTP server; the Claude Code CLI itself makes outbound calls to Anthropic under the operator's own account, inherent to exercising this adapter, not target network use)
• Run mode labels: local_agent, tool_backed
• Model/harness labels: model=claude-sonnet-5, provider=Anthropic (via Claude Code CLI, no API key), agent_runtime=Claude Code, harness=manual-review (ad hoc verification script, not a committed bench harness), tool_access=local_only
• Commands run:
  • npm run typecheck -> pass
  • npm test -> pass (675/675 vitest, 11/11 ops-preflight, 19/19 model-matrix)
  • npm run doctor -> pass (30/33, 3 warnings: semgrep/promptfoo missing, API health offline with the server not running — no blockers)
  • npm run verify-claims -> pass (27/27, no headline numbers changed)
• Artifacts: none committed. Verification used an ad hoc script (deleted after use) against a throwaway loopback HTTP fixture, plus two direct claude -p --output-format json calls confirming the envelope shape.
• Redaction: not_applicable (no secrets or evidence involved)
• Claims changed: none
• Abstentions/refusals: none
• Residual risk: Codex and Hermes local-agent paths are untouched and unverified (this fix targets claude only). The character-based fallback estimate (used only if the JSON envelope fails to parse) is a rough approximation, not exact. agentFailureOutput's plain-text failure detection in local-agents.ts is not updated for JSON-wrapped error envelopes (pre-existing behavior, unchanged here).

[jmagly]

Maintainer courtesy cleanup pushed as 11e3d4c.

I kept the change narrow:

• preserve the parsed Claude result when the envelope has missing, empty, zero, malformed, or negative usage;
• accept real usage only when token fields are finite, non-negative, and non-zero in aggregate;
• otherwise invoke the existing character estimate instead of returning zero usage;
• add behavioral coverage for real cache-token aggregation, every fallback shape, non-JSON output, and tool-call parsing from the envelope result.

Verification on a synthetic merge into current main (9020e5f): typecheck passed; focused suite 33/33; full suite 748/748 Vitest plus 11/11 ops preflight and 19/19 model matrix; lint 0 errors (145 existing warnings on the merged tree). I’m waiting for hosted CI on the new exact head before merge.

[jmagly]

Reviewed at 11e3d4c884e05931e8728a427a6333888da86513.

No blocking findings after maintainer-courtesy cleanup. Real Claude JSON usage is aggregated correctly; absent or invalid usage retains the parsed result and fails over to a non-zero estimate; non-Claude paths remain unchanged. Behavioral regression coverage exercises the reported budget-accounting boundary and JSON-wrapped tool calls.

Verification: hosted CI passed all gates; synthetic merge into current main passed typecheck, focused 33/33, full 748/748 Vitest plus ops/model checks, and lint with zero errors.