Triss mascottrisscoworker
GitHub ↗
triss coder

The part where she writes the code.

triss coder runs bounded implementation tasks through your selected coding engine. Inspect the resulting changes before accepting them.

Follow the implementation workflow — from a bounded task to an inspected, accepted change.

a whole task, delegated
$ triss coder init
engine? provider? key? → done once
$ triss coder run \
"add input validation to /signup" \
--isolate
→ 3 files changed in .triss/wt/signup
→ your working tree: untouched
engines

Pick one engine. Keep the same public model IDs.

Every engine panel is rendered below — with JavaScript they collapse to tabs, without it they all stay visible.

opencode

default

The V1 engine and the stable default. Its deny-first bash allowlist in opencode.json is enforced. Zen and Go runs keep native OpenCode provider behavior. Protected mode forwards only User-Agent plus session, request, and client identity; the project fingerprint stays local.

isolation
opt-in via --isolate
credential default
best_effort_raw (warned); --protect-credentials uses the parent-owned proxy
triss coder run --engine opencode "your task"

opencode2

beta

The V2 beta. OpenCode 2 0.0.0-beta-19059 or newer is accepted by default — the supported floor plus every newer parseable version, never one pinned build; Triss checks required option declarations, not help descriptions. It shares the same opencode.json config and provider catalogue.

isolation
opt-in via --isolate
credential default
best_effort_raw (warned); --protect-credentials uses the parent-owned proxy
triss coder run --engine opencode2 "your task"

crush

interim

Provider-neutral single-envelope engine: any canonical provider projects onto a run-scoped config with $ENV credential references. Isolation is ON by default: the run happens in a disposable worktree that separates changes from your checkout. That worktree is not an operating-system sandbox. By default the engine auto-approves its tools; --restrict opts into the CLI allowlist enforcement for this run.

isolation
ON by default
credential default
protected_proxy by default; --no-protect-credentials is the explicit raw choice
triss coder run --engine crush "your task"

omp

new

Native OMP headless adapter with structured events and named sessions. Run-private config prevents profile inheritance. It defaults to worktree isolation — and a worktree is not an OS sandbox.

isolation
ON by default
credential default
best_effort_raw (warned); --protect-credentials uses the parent-owned proxy
triss coder run --engine omp "your task"

harness bundle

DSH plugin, not --engine

Not an --engine value — opt-in DSH companion bundle. One profile dependency instead of hand-authoring provider mapping.

isolation
owned by Harness
credential default
owned by Harness
dsh plugin add triss-dsh-provider-bundle

From run to accepted change

The same path with or without JavaScript: run, locate the result, inspect every class of change, verify, then decide.

  1. Run a bounded tasktriss coder run --engine <engine> --isolate "the bounded task"The prompt describes the task; it does not create a sandbox policy. --isolate gives you a disposable Git worktree, not an operating-system sandbox.
  2. Locate the resulttriss coder result listOr list sessions per engine: triss coder session list --engine <engine>. The JSON envelope reports the worktree path for isolated runs; retention of a result artifact is not guaranteed for every run.
  3. Inspect the actual worktreegit -C "$WORKTREE" status --short && git -C "$WORKTREE" diff --cached && git -C "$WORKTREE" diffAlso check untracked files: git -C "$WORKTREE" ls-files --others --exclude-standard. Staged and untracked changes are part of the result — a diff alone is not the whole picture.
  4. Verify, then accept or rejectRun the project's focused checks in the worktree yourself; a zero exit code is not proof the task is correct. Accept through your normal Git process — review, commit, pull request. To discard a specific retained result: triss coder result clean <run-id>; a specific inactive session: triss coder session clean <slug> --engine <engine>.

Capability and isolation verdicts (effective_isolation, execution_capabilities) describe the current implementation per engine and route. Where a run reports best_effort or unavailable, treat it as a disclosed limitation of this version and decide accordingly — not as a guaranteed property of every route, and not as a reason an engine is permanently unusable. See the reliable delegation contract.

The envelope you actually review

One JSON envelope printed to stdout per run — honest about isolation, files changed, and capabilities. Every field note below is static content; no script is required to read it.

session_slugNever an implicit conversation

Either the slug you passed or a generated anonymous one. A run never silently continues a previous session.

triss coder session list
files_changedNull unless isolated

Only an isolated run can honestly count changed files. Otherwise this is null.

--isolate
worktreeA disposable checkout

The isolated run happens here, never in your working tree. Triss waits for residual processes.

.triss/wt/<slug>
effective_isolationEnforced, or honestly downgraded

Fails before spawn unless you opted into best-effort. Advisory only when downgraded.

--allow-best-effort-caller-worktree
execution_capabilitiesEight honest verdicts

Each capability reports enforced, best_effort, or unavailable — verified for the run.

docs/reliable-delegation-contract.md