Harnesses and roles
Every agent harness Gantry drives, the environment keys that configure each one, and which role runs which stage of the pipeline.
The harnesses
Agent harnesses — 4 rows exported from
src/engine/harness/mod.rs
Claude
picked with c
| field | environment key | if that is unset |
|---|---|---|
| bin | GANTRY_CLAUDE_BIN | GANTRY_CLAUDE |
| model | GANTRY_CLAUDE_MODEL | GANTRY_MODEL |
| small_model | GANTRY_CLAUDE_SMALL_MODEL | GANTRY_COMMIT_MODEL |
| args | GANTRY_CLAUDE_ARGS | the harness's own default |
OpenCode
picked with o
| field | environment key | if that is unset |
|---|---|---|
| bin | GANTRY_OPENCODE_BIN | the harness's own default |
| model | GANTRY_OPENCODE_MODEL | the harness's own default |
| small_model | GANTRY_OPENCODE_SMALL_MODEL | the harness's own default |
| args | GANTRY_OPENCODE_ARGS | the harness's own default |
Codex
picked with x
| field | environment key | if that is unset |
|---|---|---|
| bin | GANTRY_CODEX_BIN | the harness's own default |
| model | GANTRY_CODEX_MODEL | the harness's own default |
| small_model | GANTRY_CODEX_SMALL_MODEL | the harness's own default |
| args | GANTRY_CODEX_ARGS | the harness's own default |
Gemini
picked with g
| field | environment key | if that is unset |
|---|---|---|
| bin | GANTRY_GEMINI_BIN | the harness's own default |
| model | GANTRY_GEMINI_MODEL | the harness's own default |
| small_model | GANTRY_GEMINI_SMALL_MODEL | the harness's own default |
| args | GANTRY_GEMINI_ARGS | the harness's own default |
The roles
Harness roles — 4 rows exported from
src/domain/mod.rs
| role | stages it runs |
|---|---|
| plan | Brief, Plan, Test setup, Replan |
| build | Execute, Fix, Investigate, Troubleshoot, Resolve |
| review | Review |
| util | Test, Merge, Support, Done |
The stages
Pipeline stages — 14 rows exported from
src/domain/mod.rs
| stage | shown as | description | run by |
|---|---|---|---|
| environment-build | Brief | Writes the task's environment brief after planning and before gate build | plan |
| plan | Plan | Splits the plan into tasks or revises the remaining plan | plan |
| gate-build | Test setup | Writes the project gate on the clean baseline | plan |
| execute | Execute | Implements one task in a fresh context | build |
| gate | Test | Runs the project-supplied gate | util |
| fix | Fix | Makes one bounded repair after a red gate. Bundled `map` schedules it; other recovery paths use `troubleshoot` | build |
| investigate | Investigate | Diagnoses a failed task and writes a report without repair. No bundled pattern schedules it; it remains so older journals render | build |
| troubleshoot | Troubleshoot | Diagnoses the intact failed tree, repairs what a real check verifies, or declares a briefed retry or handover | build |
| review | Review | Assesses a task's result, applies safe improvements, and judges scope | review |
| replan | Replan | Plans the remaining gap in an `until` run after its acceptance check remains red | plan |
| merge | Merge | Merges the run branch into the run's pinned home branch | util |
| resolve | Resolve | Resolves a merge-back conflict in the disposable worktree | build |
| support | Support | Handles a blocker exchange or pattern-declared run-record bookkeeping | util |
| done | Done | Records a terminal summary card after the run finishes green and merges | util |