skills: add execution agent dispatch protocol

This commit is contained in:
2026-03-07 20:58:32 +10:00
parent dd1884e4c2
commit f0ed738453
4 changed files with 194 additions and 0 deletions
+1
View File
@@ -39,6 +39,7 @@
- Historical temporary preference once was: all subagents **primary** = `openai-codex/gpt-5.2`, **fallback** = `github-copilot/claude-opus-4.6`.
- Current important exception / newer rule: **codex_worker** should use **primary** = `openai-codex/gpt-5.4`, prefer high/extra/xhigh thinking when available, and should **not auto-fallback** to another model. If its primary model is unavailable, report to Tao and let Tao decide the replacement model.
- Additional long-lived execution agents initialized locally for repeat use: **docs_worker**, **qa_worker**, **ops_worker**.
- Execution-agent system still needs a standardized dispatch/handoff layer: task input template, result format, and blocker/escalation rules.
## 8) Watchdog / automation policy
- Tao preference: avoid watchdog-style auto-restart automation.
+36
View File
@@ -60,3 +60,39 @@
- New agent-splitting rule clarified: create multiple execution agents when work is parallelizable, responsibilities differ, contexts would otherwise get mixed, or separate validation/release tracks are needed; avoid splitting tiny or tightly coupled work.
- Long-lived execution agents initialized locally: `docs_worker`, `qa_worker`, `ops_worker` (with dedicated agent dirs + workspaces + basic role files). `codex_worker` remains the default general execution worker.
- This architecture is the current default, but Tao may revise it as needs change.
## OCM / workflow / agent architecture updates (late evening)
- OCM work completed today:
- Rewrote README opening around first-run value, support positioning, and product framing.
- Refreshed README hero screenshots using newer redacted Dashboard / Agents / CLI captures.
- Adjusted README screenshot layout to use one large Dashboard hero image plus a secondary row (Agents / CLI / Actions) because mixed aspect ratios looked awkward in a single row.
- Unified OCM built-in CLI completion text to English:
- success: `✅ Done (exit 0)`
- non-zero exit: `❌ Exit code N`
- OCM GitHub flow today confirmed again that the repo is effectively PR-only even when approvals are set to 0:
- PR #5: README positioning + screenshot refresh
- PR #6: version bump to `v0.9.1`
- PR #7: English CLI completion messages
- PR #8: README screenshot layout A
- Release published: `v0.9.1`
- Tao explicitly said that for status updates, accuracy matters more than token saving; concise is fine, but not at the cost of ambiguity.
- Tao approved a multi-agent dev architecture as the default pattern:
- main = PM / architect / reviewer / user-facing coordinator
- execution agents = hands-on implementers
- use `codex_worker` as the default coding execution agent
- create more execution agents when work is parallelizable, responsibilities differ, contexts would mix, or separate validation/release tracks are needed
- avoid splitting tiny or tightly coupled work
- Tao approved a stronger delegation rule:
- unless work is truly tiny and can be completed in one short pass, main should avoid personally implementing/editing and should delegate execution to `codex_worker` first
- Long-lived execution agents initialized locally for repeat use:
- `docs_worker`
- `qa_worker`
- `ops_worker`
- `codex_worker` remains the default general execution worker
- Two workflow/architecture skills were created and published to GitHub:
- `gh-pr-release-flow` → https://github.com/dtzp555-max/gh-pr-release-flow
- `execution-agent-planner` → https://github.com/dtzp555-max/execution-agent-planner
- Tao's preferred lifecycle rule for execution agents was clarified:
- keep long-lived role-based agents for reuse
- delete or archive project-specific temporary agents after the project ends to avoid an “agent graveyard”
- Next improvement approved: add a standardized execution-agent dispatch / handoff protocol so workers receive clearer task packets and report blockers/results consistently.
+53
View File
@@ -0,0 +1,53 @@
# execution-agent-dispatch
A lightweight skill for turning PM decisions into clear worker handoffs.
## Why this exists
It is not enough to have execution agents.
They also need a consistent way to receive work and report results.
Without that, the system drifts into familiar failure modes:
- vague delegation
- role boundary drift
- workers doing too much or too little
- silent blocking
- messy result reporting
This skill exists to standardize the handoff layer between **main** and execution agents.
## What it standardizes
- the task packet main sends to a worker
- the result format a worker returns
- when a worker must escalate instead of staying silent
- milestone / ETA discipline for delegated work
## Core idea
A good worker handoff should answer:
1. What exactly should be done?
2. What is out of bounds?
3. What does success look like?
4. When should the worker report back?
5. What should trigger escalation?
## Recommended task packet
- Task
- Scope
- Deliverable
- Constraints
- Milestone 1 + ETA
## Recommended result packet
- `status:` ok | blocked | failed
- `changed:` ...
- `notes:` ...
- `next:` ...
## Notes
This is a dispatch / coordination skill, not an implementation skill.
It works best after `execution-agent-planner` has already decided which workers should exist and what they should own.
+104
View File
@@ -0,0 +1,104 @@
---
name: execution-agent-dispatch
description: Standardize how main dispatches work to execution agents and how execution agents report back. Use when main has already decided which execution agent(s) should do the work and needs a clear handoff packet, result format, blocker/escalation rules, and milestone discipline so workers do not drift or go silent.
---
# Execution agent dispatch
Use this skill after execution architecture is decided.
## Goal
Turn vague delegation into a clear handoff.
Every dispatched task should make it obvious:
- what to do
- what not to do
- what counts as done
- when to escalate back to main
## Dispatch packet
When main assigns work to an execution agent, include these sections:
### 1) Task
One short statement of the goal.
### 2) Scope
Specify:
- repo / path / files if known
- what area is in bounds
- what is explicitly out of bounds
### 3) Deliverable
Define the expected output:
- code change
- docs update
- validation report
- PR / commit / test result
- recommendation only
### 4) Constraints
List non-negotiables, such as:
- do not change unrelated files
- do not switch models automatically
- do not message users directly
- ask before destructive actions
### 5) Milestone 1 + ETA
Tell the worker what first checkpoint matters and when main expects the first update.
## Worker response format
Execution agents should respond compactly with:
- `status:` ok | blocked | failed
- `changed:` files changed/created, commands run, or "none"
- `notes:` short summary
- `next:` next action or handoff need
## Escalate immediately when
A worker should report back to main instead of silently stalling when:
- permissions are missing
- the model/tooling is unavailable
- repo rules block the intended action
- requirements are contradictory or underspecified
- the task is crossing role boundaries
- the ETA has clearly slipped
- the requested change is riskier or broader than the original handoff suggested
## Silence rule
Workers should not chatter, but should also not disappear.
Default rule:
- acknowledge briefly
- do the work
- report at milestone
- report immediately on blockers
## Completion rule
A task is not complete just because files changed.
Completion should include:
- the requested deliverable exists
- basic verification happened when relevant
- blockers or caveats are disclosed
- main has enough information to review and report upward
## Main's responsibility
Main must not delegate sloppily.
Before dispatching, main should decide:
- why this worker is the right worker
- what the boundary is
- what review criteria will be used
- whether the task is small enough that delegation is unnecessary
## Good default tone
- short
- operational
- explicit
- non-dramatic
Prefer a clean task packet over a long motivational speech.