feat(usage): track JSONL tokens + cost, gated by api_key auth, with subagent rollup

[manzil-infinity180]

Two related fixes shipped together so the api_key-gated cost tracking and the parent/subagent spend rollup can be tested in one POC.

What's new

• internal/usage — JSONL parser with (message.id, requestId) dedup and pricing table for opus/sonnet/haiku families (AFLOCK_PRICE_* overrides).
• internal/auth/claudeauth — auth-mode probe (env var → ANTHROPIC_API_KEY → ~/.claude/.credentials.json → macOS keychain).
• SessionState.AuthMode + richer SessionMetrics (cache 5m/1h, usageSource, costMeasured).
• Hooks refresh usage from JSONL at PostToolUse + SessionEnd before CheckLimits.
• Cost enforcement (maxSpendUSD) activates only under api_key; advisory under subscription/unknown via Evaluator.IsAdvisoryLimit. Token/turn limits enforce under both modes (transcript counts match Anthropic's per-turn accounting either way).
• Subagent rollup (closes fix(state): subagent spend not rolled up to parent until SubagentStop — parent fail-fast can be exceeded mid-flight (paper §5 invariant #2) #135): parent's PostToolUse enumerates child sessions via new state.Manager.ListChildSessions(parentID), sums any whose IDs aren't already in parent.ChildSessionIDs, and feeds effective metrics into fail-fast — closes the window where concurrent children could blow past the parent ceiling between SubagentStop merges. Race narrows to one child-tool-call delta (per paper §5 invariant #2 acceptance).

Test plan

• go test ./... clean
• api_key session over maxSpendUSD → block
• subscription session over maxSpendUSD → stderr advisory only
• maxTokensIn enforces under both modes
• Parent at $0.30 + child spending $0.30 with maxSpendUSD=\$0.50 → parent's next PostToolUse blocks, stderr shows Subagent rollup: including N in-flight child(ren)
• After SubagentStop, the next rollup correctly skips the merged child (no double-count)

[Copilot]

Pull request overview

This PR adds transcript-driven token/cache/cost accounting from Claude Code JSONL logs, gates cost-limit enforcement based on detected Claude authentication mode, and implements parent-side aggregation of in-flight subagent spend to narrow the fail-fast overspend window (issue #135).

Changes:

• Introduce internal/usage JSONL scanning + model-family pricing with env overrides, and plumb the resulting metrics into hook limit checks.
• Add internal/auth/claudeauth auth-mode detection and downgrade cost-based limits (e.g., maxSpendUSD) to advisory under non-api_key sessions across hooks/replay/verify.
• Add parent PostToolUse subagent rollup via Manager.ListChildSessions + rollupUnmergedChildren for near-real-time accumulation.

Reviewed changes

Copilot reviewed 18 out of 18 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
pkg/aflock/types.go	Extends SessionState and SessionMetrics to include auth mode + cache token breakdown + cost provenance flags.
internal/verify/verifier.go	Treats cost-based limit exceedance as advisory under non-api_key auth during verification.
internal/usage/testdata/sample.jsonl	Sample transcript fixture for JSONL parser tests.
internal/usage/pricing.go	Model-family pricing table + env overrides + cost computation.
internal/usage/pricing_test.go	Unit tests for pricing lookup, overrides, and unknown-model tracking.
internal/usage/jsonl.go	JSONL transcript scanner with (message.id, requestId) dedup and per-model rollups.
internal/usage/jsonl_test.go	Unit tests for JSONL dedup, filtering, and missing-file behavior.
internal/state/session.go	Adds Manager.ListChildSessions for parent-side rollup enumeration.
internal/state/session_test.go	Tests child-session enumeration correctness.
internal/replay/simulate.go	Aligns replay behavior with advisory cost limits under non-api_key auth.
internal/policy/evaluator.go	Adds Evaluator.IsAdvisoryLimit(limitName, authMode) helper.
internal/policy/evaluator_authmode_test.go	Tests advisory/enforced matrix for limits vs auth modes.
internal/mcp/server.go	Exposes new metrics fields + authMode in session API output.
internal/hooks/usage_refresh.go	Adds transcript refresh to populate token/cache/cost metrics in hook state.
internal/hooks/subagent_rollup_test.go	Tests rollup correctness, no double-count, and no parent mutation.
internal/hooks/handler.go	Wires auth-mode capture, transcript refresh, advisory gating, and in-flight subagent rollup into hooks.
internal/auth/claudeauth/detect.go	Implements auth-mode detection (env/api-key/credentials/keychain).
internal/auth/claudeauth/detect_test.go	Tests auth-mode detection precedence and file/env probes.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[manzil-infinity180]

test it with the anthropic api key method :)

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.

[Copilot]

Copilot encountered an error and was unable to review this pull request. You can try again by re-requesting a review.