[hz89Dev]

Summary

Oh My Pi, Qoder, Kiro, ZCode, Trae, OpenClaw, Hermes and DeepSeek Harness set no session variable teamai reads. So a skill-driven workflow there could not resolve a step's model, and #1386's session judge never ran. --tools was refused with --driver skill, so the session had no way to say which tool it is. Now it names its tool with --tools default=<tool>, describe tells it when it has to, and the run keeps that tool to the end.

 teamai workflow start --describe <wf> --driver skill
+  says: "This coding session names no tool ...: describe the workflow again with --tools default=<tool>"
 teamai workflow start|resume ... --driver skill
-  --tools                    refused: "--tools is for --driver local"
+  --tools default=<tool>     the tool this session works in
 skill.json orchestrator
+  tool                       kept with the owner; runners resolve child models for it, done.score leases judging to it

• Every command teamai prints for the session keeps --tools: the starts describe offers, the describe a refused start prints, and the resume a refused resume prints.
• The tool moves with ownership: a takeover, or a session that comes to own an unowned run, works in its own tool. A resume with --tools repairs a run started without one.
• Models resolve only for tools teamai installs agents for. Judging works for every tool.
• A session whose host starts no child agent (h4) no longer gets the judge work. A judging-off event says why (no_child_agent).
• features.sessionJudge in TOOLS records, per host, whether it starts a fresh subagent, with evidence. The README scorers cell shows ✓³ from it for Oh My Pi, Pi, OpenClaw, Hermes, DeepSeek Harness, Qoder, Qoder CN, Kiro, Trae, Trae CN and ZCode, in every README.

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature causing existing behavior to change)
• Documentation only
• Refactor / internal cleanup

Test Plan

• npx tsc --noEmit passes
• npm run lint passes
• npx vitest run passes: 17633 passed.
• Added/updated tests for the change. Each new driver test was seen failing first.
• npm run test:e2e -- src/__tests__/e2e/workflow-skill-claim.test.ts src/__tests__/e2e/score-external-driver.test.ts src/__tests__/e2e/workflow-session-1086.test.ts: 6 passed.

Real CLI, after npm run build, with no agent session variable set (Oh My Pi 18.12.6 on PATH). It runs a workflow whose one prompt step pins model alias strong and is scored by review-quality, from describe through done to the stored label:

# Before (base commit)
## $ teamai workflow start --describe review --driver skill
{"kind":"elsewhere","reason":"The coding session's tool has no confined judge teamai runs."}
## $ teamai workflow start --describe review --driver skill --tools default=omp
{"error":{"code":"invalid_arguments","message":"--tools is for --driver local","next":null}}

# After (this branch)
## $ teamai workflow start --describe review --driver skill
{"judging":{"kind":"elsewhere","reason":"This coding session names no tool."},"says":"- This coding session names no tool for its prompt steps and judging: describe the workflow again with --tools default=<tool>, the tool it works in, and start it from that reply.\n- review is not judged on this machine. This coding session names no tool."}
## $ teamai workflow start --describe review --driver skill --tools default=omp
{"judging":{"kind":"session","tool":"omp","reason":"The coding session's tool, omp, has no confined judge teamai runs."},"starts":["teamai workflow start review --driver skill --tools default=omp","teamai workflow start review --driver skill --tools default=omp --no-judge"]}
## dispatch of review, prepared by its detached runner
{"step":"review","model":{"by":"step","name":"anthropic/claude-opus-4-5","alias":"strong"}}
## skill.json orchestrator
{"tool":"omp","lastCallAt":"2026-10-27T20:36:27.318Z"}
## $ teamai workflow next <run> --driver skill   (after the review step reported)
{"status":"done","actions":[{"kind":"done","outcome":"succeeded","prs":[],"score":"teamai score --next --json '{\"run\":\"r2580524fd85fbbe2\",\"driver\":{\"kind\":\"skill\",\"machine\":\"<machine>\",\"workspace\":\"<root>/app\",\"can\":{\"judge\":{\"isolation\":\"best-effort\",\"access\":\"read\",\"tools\":[\"omp\"]}}},\"running\":[]}'"}]}
## $ <done.score>
{"subject":{"runId":"r2580524fd85fbbe2","step":"review","pass":1},"judge":{"tool":"omp","by":"tool_default","isolation":"best-effort","access":"read"},"submit":"teamai score --lease s55e5b013922bc/ab5f49714 --submit --stdin"}
## $ teamai score --lease s55e5b013922bc/ab5f49714 --submit --stdin   (from the lease cwd)
thorough
## stored label
{"scorer":"review-quality","label":"thorough","source":"workflow","subject":{"runId":"r2580524fd85fbbe2","step":"review","pass":1},"judge":{"kind":"model","tool":"omp","model":{"reported":"anthropic/claude-opus-4-5"},"isolation":"best-effort","access":"read"}}

Not run here: the other providers (gitlab, github) and agents (Codex, CodeBuddy, OpenCode). This change touches only the skill driver's session tool, so I leave those to CI. The ✓³ cells come from each host's docs or source, as each sessionJudge note says. None was run in that host.

Related Issues

Closes #1479

Notes for Reviewers

• sessionJudge marks: Pi and OpenClaw are ◐. Bare Pi starts no subagent and answers h4, and an OpenClaw tool policy can remove sessions_spawn. DeepSeek Harness is ◐ because its minimal preset leaves the subagent tool out. JoyCode stays ?.
• New session variables (HERMES_SESSION_ID, DSH_SESSION_ID) are left to [feat] Read the sessions of more tools, so they can be scored #1418. They also change invoke evidence, recall and teamai score's current session.
• Rows R10, R11, R20, S1 and S4 of docs/dev/workflow-engine.md changed. The transition table test checks them cell by cell.

[github-actions]

• [P1 blocking] An explicit resume tool cannot replace the stored tool — src/workflow/drivers/skill/transitions.ts:448: Start in Oh My Pi with --tools default=omp, then resume from Qoder with --tools default=qoder, with neither host setting a recognized session variable. The existing tool wins over input.tool, so child models and done.score still target Oh My Pi. An explicitly declared resume tool needs to override the previous tool.

• [P1 blocking] Naming a missing tool on resume does not restore judging — src/workflow/drivers/skill/tick.ts:301: Start a scored prompt workflow without a detected tool, then resume with --tools default=omp before answering the host question. readiedRow discards the new judging decision because skill.json already exists; callRow fills only orchestrator.tool. The step can now run, but sessionJudge remains absent and completion prints no done.score, although judging was never opted out.

• [P2 non-blocking] Command-only workflows bypass the no-subagent check — src/workflow/drivers/skill/transitions.ts:585: A command-only workflow with a best-effort scorer and --tools default=pi gets sessionJudge, but its host is none, so this h4 check never runs. Bare Pi consequently receives done.score requiring a fresh subagent it cannot start. Check judging capability independently of whether the workflow has prompt steps.

The PR description documents sufficient testing, including a representative real-CLI run after build. I ran, built, and installed nothing from the PR.

[SaulMoro]

This was generated by AI during triage.

Approving at 1f5c467. Every acceptance item in #1479 works.

Built CLI, in a host that sets no session variable (base = main bd3adf9b)

• workflow start --describe --driver skill says to describe again with --tools default=<tool>. With it, the run dispatches with the step's model resolved for omp, done.score leases judging to omp, and the stored label has judge.tool: omp. Main refuses --tools with --driver skill.
• The three --tools refusals, the takeover case and the h4 host (no child agent, no judge work) behave as the body says.
• npx tsc --noEmit and npm run lint exit 0; the PR's 8 test files (237 tests) pass, also with a symlinked TMPDIR, and 15 of them fail on main. Every CI job passed at this head, including the three E2E shards.
• Covered by unit tests only: the --claim resume keeping --tools, the judging-off event text.

Non-blocking

• [P2 non-blocking] When neither session is known, a resume's --tools exits 0 but is ignored if the run already has a tool, so a run started with the wrong tool cannot be corrected. That follows #1479's "fills it when the run has none", but letting an explicit --tools replace the stored tool in callRow would allow the repair.
• [P2 non-blocking] A workflow whose only scored step is a workflow: step with command children never asks the host question, so it skips the h4 check; bare Pi then gets a done.score it cannot act on. The lease lapses and stays open, so nothing is lost.
• [P3 nit] A resume that names the tool does not turn session judging back on (describe had said so; main behaves the same).
• [P3 nit] The describe instruction does not list which ids <tool> accepts (dsh, qoder-cn); the refusal for an unknown tool does.

The CI review's three findings reproduce; we rate them as above (the issue's own rule for the first, main's behavior for the second).

[SaulMoro]

This was generated by AI during triage.

@hz89Dev this merged with three findings from the CI review still open. We reproduced all three; could you fix them in a follow-up PR?

1. An explicit resume tool does not replace the stored one (src/workflow/drivers/skill/transitions.ts:448). Start in Oh My Pi with --tools default=omp, resume from Qoder with --tools default=qoder, neither host setting a session variable: the resume exits 0, but child models and done.score still target Oh My Pi, so a run started with the wrong tool cannot be corrected. An explicit --tools should replace the stored tool in callRow.
2. Naming the tool on resume does not restore judging (src/workflow/drivers/skill/tick.ts:301). Start a scored prompt workflow with no detected tool, then resume with --tools default=omp before answering the host question: the step runs, but sessionJudge stays absent and completion prints no done.score, although judging was never turned off.
3. Command-only workflows skip the no-subagent check (src/workflow/drivers/skill/transitions.ts:585). A workflow whose only scored step is a workflow: step with command children never asks the host question, so the h4 check never runs and bare Pi gets a done.score it cannot act on. The judging check should not depend on the workflow having prompt steps.

[hz89Dev]

Fixed in #1481, each with a test that failed before the fix:

• A resume's --tools default=<tool> now replaces the run's tool, so child models and done.score follow it.
• Judging can no longer be lost to a missing tool. skill.json does not record --no-judge, so a later resume cannot tell an opt-out from "no tool at start". Instead, describe asks a session that names no tool for one (ask.kind: tool), and a start or resume that would drive a run without the tool its models or judging need is refused until it names one.
• A run its session judges asks the host question even without prompt steps, so h4 turns judging off. For such a run the answer decides only judging, so its commands still run as before.