# Gait

> Offline-first durable agent runtime: dispatch long-running jobs with checkpoints, capture signed packs, gate tool calls and voice commitments with fail-closed policy, and turn incidents into deterministic CI regressions.

## Canonical
- https://github.com/davidahmann/gait
- https://davidahmann.github.io/gait/

## Core Commands
- gait demo
- gait demo --durable
- gait demo --policy
- gait tour
- gait verify <run_id|path>
- gait run replay <run_id>
- gait pack verify <pack.zip>
- gait pack diff <left.zip> <right.zip>
- gait pack export <pack.zip> --otel-out <otel.jsonl>
- gait regress bootstrap --from <run_id|path> --junit <path>
- gait regress run --json --junit <path>
- gait policy test <policy.yaml> <intent.json>
- gait gate eval --policy <policy.yaml> --intent <intent.json>
- gait job submit --id <job_id>
- gait job status --id <job_id>
- gait job checkpoint add --id <job_id> --type <type> --summary <text>
- gait job checkpoint list --id <job_id>
- gait job pause --id <job_id>
- gait job resume --id <job_id>
- gait job cancel --id <job_id>
- gait voice token mint --intent <intent.json> --policy <policy.yaml>
- gait voice token verify --token <token.json>
- gait voice pack build --from <call_record.json>
- gait voice pack verify <callpack.zip>
- gait voice pack diff <left.zip> <right.zip>
- gait doctor --json
- gait doctor --summary
- gait ui

## Tool Boundary
- Boundary location: the exact adapter call site before real tool side effects execute.
- Input: structured intent payload.
- Rule: `verdict != allow` means no tool execution.

## When To Use
- You need enforceable policy decisions before agent tool side effects.
- You need signed portable evidence artifacts and deterministic regressions.
- You need durable checkpointed agent jobs with inspectable state.

## When Not To Use
- No local Gait CLI/artifact path exists.
- No tool side effects or evidence requirements exist.

## Safety Model
- Fail-closed by default for ambiguous high-risk policy outcomes.
- Offline verification for packs, runpacks, and traces.
- Structured intent model for policy decisions (not free-form prompt filtering).
- SayToken capability tokens for voice agent commitment gating.
- Context evidence envelopes with fail-closed enforcement when evidence is missing.
- Signed traces and explicit reason codes for blocked actions.
- No hosted service dependency required for core operation.

## Canonical Docs
- /docs/adopt_in_one_pr/
- /docs/durable_jobs/
- /docs/integration_checklist/
- /docs/architecture/
- /docs/flows/
- /docs/threat_model/
- /docs/failure_taxonomy_exit_codes/
- /docs/contracts/packspec_v1/
- /docs/contracts/pack_producer_kit/
- /docs/contracts/compatibility_matrix/

## LLM Resources
- /llms-full.txt
- /llm/product.md
- /llm/quickstart.md
- /llm/security.md
- /llm/faq.md
- /llm/contracts.md