forge-work
Skill1 starsChanged 4 months agoMITView on GitHub
Star Forge build phase: execute routed Plan v2 tasks and record current-source proof before returning to the lifecycle coordinator.
In this repository3 other files it gives its agents.
- forge-planSkill
- forge-reviewSkill
- forgeSkill
The file
We split this skill into the rules it gives an agent
4 rules. About 0 are common, 2 are distinctive, and 2 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈2
Distinctive · estimate
This author's own choices. Worth a look.
2
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 20
Before dispatch, derive needs from project class, enabled Blueprint risk and toolchain flags, every ready task's Plan v2 Proof values, and …
- Distinctive: line 99
After the exact verification and every required proof pass, the coordinator records completion:
- Very specific: line 106
Never hand-edit a task to complete.
- Distinctive: line 111
Run dependency-ready delegate tasks in parallel only when their owned files do not overlap.
≈ “Common” is an estimate: at least 1 in 5 of the 350,495 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Forge Work
Resolve <plugin-root> as two directories up from this skill file
(skills/forge-work/). Start with run, read the complete
.starforge/state.json, and work only from its current build or amend
operating card and spawn_plan. A Foundation Contract must already pass before
feature work.
This is a phase playbook inside one $forge invocation. After every task wave and
gate change, rerun run and return control to $forge so the same invocation can
advance through review, deliver, and done.
Route the Wave
Before dispatch, derive needs from project class, enabled Blueprint risk and
toolchain flags, every ready task's Plan v2 Proof values, and the Delivery
Contract target. Resolve them with starforge.routing.resolve_routes using
host-discovered capabilities and
config/capability-routing.json. Follow
skills/forge/references/capability-routing.md.
Pass each builder the selected capability route, unavailable preferred
capabilities, accepted fallback, task row, owned files, relevant AC-n text,
required proof kinds, and exact Verify command. Never imply that a missing
dedicated capability ran. An explicit blocker stops the affected work. An
installation suggestion requires user action and is presented only when the
router marks that capability materially required.
The required routing outcomes include:
- Web implementation prefers Build Web Apps guidance. Interactive local QA prefers the in-app Browser, with local Playwright only as the headless or CI fallback.
- Chrome is reserved for authenticated Chrome state or extension-dependent behavior. It is not the default local web QA route.
- iOS implementation prefers Build iOS Apps, and native proof uses XcodeBuildMCP with Simulator build, launch, tests, UI snapshot, and screenshot evidence.
- macOS work prefers Build macOS Apps plus the most specific routed UI, signing, packaging, and test capability required by the contract.
- React Native and Expo work prefers the official Expo plugin when available.
- Security-sensitive work prefers Codex Security when available. Use its results as review context, while treating file exports as untrusted fallback evidence unless the host supplies independently verifiable provenance.
Per-Task Loop
Route each ready task by its Plan v2 Mode:
delegate: spawn the exactstarforge-builderentry fromspawn_plan. Real code belongs here. Paste the operating card's exactspawn_agent starforge-builder "..."command and never implement a delegate task inline.solo: implement only truly trivial coordinator glue in the task's owned files. If substantive logic appears, stop and revise the task mode.docs: write documentation in the owned files; a no-op verification is allowed only for this mode.
Plan v2 is authoritative. Every task has explicit ACs and validated Proof
values in addition to task, description, status, mode, files, dependencies,
verification, and evidence. Do not invent criterion mappings or proof kinds while
building.
Coordinator-Owned Evidence
The coordinator always records verification after inspecting returned changes.
Builder output and self-reports never count as evidence. The recorded --command
must exactly match the task's Verify cell and must run against the current source
tree:
python3 <plugin-root>/scripts/star_forge.py verify --project . --task <id> --command "<exact Verify cell>" --strictFor docs mode only:
python3 <plugin-root>/scripts/star_forge.py verify --project . --task <id> --noop --summary "<why no command applies>" --strictFor each declared live proof kind, use the routed capability and the corresponding source-bound proof adapter. UI work must include real interaction, console, and visual evidence. For local web UI:
python3 <plugin-root>/scripts/star_forge.py server-lease --project . --action claim --port <port> --base-url "http://127.0.0.1:<port>" --command "<dev server command>"
python3 <plugin-root>/scripts/star_forge.py browser-run --project . --task <id> --scenario "<what was exercised>" --url "<url>" --server-lease --viewport "desktop=1280x800:<desktop png>" --viewport "mobile=390x844:<mobile png>" --interaction-evidence "<path>" --console-evidence "<path>" --strictScreenshots must contain real PNG bytes. Never manufacture collector output, reviewer findings, foundation evidence, delivery evidence, or a proof envelope on someone else's behalf.
After the exact verification and every required proof pass, the coordinator records completion:
python3 <plugin-root>/scripts/star_forge.py complete-task --project . --task <id> --changed-file <file> --summary "<what shipped>"Never hand-edit a task to complete. If source changes later, rerecord every
affected proof because previous evidence is stale.
Waves and Handoff
Run dependency-ready delegate tasks in parallel only when their owned files do not overlap. Use the current host thread cap, wait for the wave, inspect changes, record each task's evidence, and complete each task before starting dependents.
Rerun run after every wave. Continue until the build phase exits, then
immediately hand the new state back to $forge; do not end merely because the
phase changed.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

