Orchestrate
You, the main session, own the plan and coordinate the work. This is a guide,
not a script: skip, reorder, repeat, or do a step yourself when that serves
the plan better, and say briefly why when you depart from it.
The plan follows the planning format; load it if a section or field is
unclear.
A typical run
- Read the plan and
git status --short. Confirm the plan exists, is
readable, and doesn't conflict with unrelated uncommitted work.
- Implement. Delegate to the
code-implementer agent with the plan path,
or implement a small plan yourself. The implementer edits code, writes
tests, and runs IMPLEMENTATION_CHECKS at most twice; the main session
owns final verification.
- Review. Delegate to the
code-reviewer agent with the plan path. It is
read-only and returns APPROVE, REQUEST_CHANGES, or ESCALATE.
- Settle disputed findings. If the review raises blockers or risk
candidates you or the user doubt, ask the
scientific-risk-reviewer agent
for a second opinion before acting on them.
- Verify. Run the plan's
IMPLEMENTATION_CHECKS yourself, following the
run-checks skill. Skip this when the implementer's last run passed and
the code is unchanged since, under the reuse rule in
Verification runs.
- Fix and re-review. For
REQUEST_CHANGES or a failing check, send the
specific findings or failure back to the implementer, then re-review with
the previous review supplied, so the reviewer checks only what changed.
- Finish. Once review approves and checks pass, run
FINAL_SUITE once
and report. Skip it only when the plan says REUSE_IMPLEMENTATION_CHECK
and that command passed on the final code.
If an agent isn't available on this platform, do implementation yourself or
load review-implementation or scientific-risk-review for those steps.
When to stop and ask the user
- The review returns
ESCALATE, or a finding is a plan gap: the plan needs
revising, which /create-plan does in place with the user's approval.
- The user would have to accept a blocker rather than fix it. Get a
scientific-risk-reviewer acceptance review first, then let the user decide.
- A check can't run within
run-checks' limits, such as an intensive
scientific analysis that has to be handed to the user.
- The same finding survives two fix rounds, or three rounds pass without
approval. Summarise what keeps failing rather than looping further.
- A scientific plan and its revisions reach four distinct blocking findings,
or a fix would need yet another validator, mapping, or precedence layer.
Propose simplifying instead, as
scientific-analysis describes.
Verification runs
- Batch related targeted checks into one invocation when they can run
sequentially.
- Defer the full suite until the code is stable after review.
- Run the full suite successfully at most once for the same code; don't rerun
a suite that already passed.
- Never rerun an unchanged failure; rerun only after a relevant code change.
- Reuse an earlier result only when it recorded the exact command, exit
status, and runner outcome for the same, unchanged code.
Ground rules
- Pass agents paths and findings, not your own reinterpretation of the plan.
- Never downgrade a scientific validity failure to get past a review.
- Commit or open a PR only when the user asks.
Report
Summarise what was built, the final review decision, the exact verification
commands and their results, anything unverified or accepted, open
questions, any proposed issues with their severity and likelihood ratings,
and any needs-human issues opened. If two or more items need the user, end
with a table that ranks them by how much work each one unblocks.
1---2name: orchestrate3description: A flexible guide for carrying out a plan from the main session by delegating implementation and review to agents and running verification yourself. Use when asked to implement and review a plan, carry a plan through to completion, or when the user invokes /orchestrate.4metadata:5 source: https://skillmd.com/skills/tobywise/orchestrate6---78# Orchestrate910You, the main session, own the plan and coordinate the work. This is a guide,11not a script: skip, reorder, repeat, or do a step yourself when that serves12the plan better, and say briefly why when you depart from it.1314The plan follows the `planning` format; load it if a section or field is15unclear.1617## A typical run18191. **Read the plan** and `git status --short`. Confirm the plan exists, is20 readable, and doesn't conflict with unrelated uncommitted work.212. **Implement.** Delegate to the `code-implementer` agent with the plan path,22 or implement a small plan yourself. The implementer edits code, writes23 tests, and runs `IMPLEMENTATION_CHECKS` at most twice; the main session24 owns final verification.253. **Review.** Delegate to the `code-reviewer` agent with the plan path. It is26 read-only and returns `APPROVE`, `REQUEST_CHANGES`, or `ESCALATE`.274. **Settle disputed findings.** If the review raises blockers or risk28 candidates you or the user doubt, ask the `scientific-risk-reviewer` agent29 for a second opinion before acting on them.305. **Verify.** Run the plan's `IMPLEMENTATION_CHECKS` yourself, following the31 `run-checks` skill. Skip this when the implementer's last run passed and32 the code is unchanged since, under the reuse rule in33 [Verification runs](#verification-runs).346. **Fix and re-review.** For `REQUEST_CHANGES` or a failing check, send the35 specific findings or failure back to the implementer, then re-review with36 the previous review supplied, so the reviewer checks only what changed.377. **Finish.** Once review approves and checks pass, run `FINAL_SUITE` once38 and report. Skip it only when the plan says `REUSE_IMPLEMENTATION_CHECK`39 and that command passed on the final code.4041If an agent isn't available on this platform, do implementation yourself or42load `review-implementation` or `scientific-risk-review` for those steps.4344## When to stop and ask the user4546- The review returns `ESCALATE`, or a finding is a plan gap: the plan needs47 revising, which `/create-plan` does in place with the user's approval.48- The user would have to accept a blocker rather than fix it. Get a49 `scientific-risk-reviewer` acceptance review first, then let the user decide.50- A check can't run within `run-checks`' limits, such as an intensive51 scientific analysis that has to be handed to the user.52- The same finding survives two fix rounds, or three rounds pass without53 approval. Summarise what keeps failing rather than looping further.54- A scientific plan and its revisions reach four distinct blocking findings,55 or a fix would need yet another validator, mapping, or precedence layer.56 Propose simplifying instead, as `scientific-analysis` describes.5758## Verification runs5960- Batch related targeted checks into one invocation when they can run61 sequentially.62- Defer the full suite until the code is stable after review.63- Run the full suite successfully at most once for the same code; don't rerun64 a suite that already passed.65- Never rerun an unchanged failure; rerun only after a relevant code change.66- Reuse an earlier result only when it recorded the exact command, exit67 status, and runner outcome for the same, unchanged code.6869## Ground rules7071- Pass agents paths and findings, not your own reinterpretation of the plan.72- Never downgrade a scientific validity failure to get past a review.73- Commit or open a PR only when the user asks.7475## Report7677Summarise what was built, the final review decision, the exact verification78commands and their results, anything unverified or accepted, open79questions, any proposed issues with their severity and likelihood ratings,80and any `needs-human` issues opened. If two or more items need the user, end81with a table that ranks them by how much work each one unblocks.