Imported from agent-orc/agent-studio (
AGENTS.md). Install upstream withnpx skills add agent-orc/agent-studio. Copyright stays with the author.
AGENTS.md
Root navigation card for Codex CLI, Claude Code, GitHub Copilot coding agent, and future agent surfaces. Keep this file short. Mandatory guardrails live here; domain system-of-record maps live in
docs/.
Compatibility shims: CLAUDE.md, GEMINI.md, and
.github/copilot-instructions.md.
Frontend-scoped rules live in frontend/AGENTS.md and
apply only under frontend/.
Start Here
- Product overview: README.md.
- Future-only product direction: ROADMAP.md.
- Canonical contribution and style conventions for all agents: docs/start/contribution-and-style-guide.html.
- Documentation index: docs/start/README.md. Start there when you do not already know the right document.
- Operator setup and troubleshooting: docs/operations/setup/.
- Familiar runtime, CLI, permission, filesystem, runner, or state-machine failure: search the thematic common-problems libraries (docs/operations/common-problems/ and docs/system/common-problems/) before debugging from scratch.
- Current state of a designated topic (so we do not re-discover the same ground):
docs/concepts/designated-topics/.
The opt-in
post-agents-wiki-syncstep keeps these pointers consistent and the per-topic Current State pages fresh.
Non-Negotiable Rules
- UI work obeys the style-guide hard rules in docs/quality/design/style-guide-hard-rules.md. Most-cited: no coloured left accent line or bar on cards, panels, rows, banners, or pill groups (encode status via background tint, badge, or dot).
- New or touched admin UI also obeys the flat-surface contract and reference mockup in docs/operations/admin-design-guideline/index.html. Apply it incrementally; do not turn a scoped card into a broad restyling pass.
- Work only in the active dev checkout or assigned task worktree. Never edit
agent-taskboard-stable/; stable updates only by the parentupdate-stable.shafter a verified dev batch. - Never directly create, move, edit, delete, or rename anything under
agent-taskboard-workspace/projects/**oragent-taskboard-workspace/.metadata/**. Use the application API. If the API lacks the operation, queue a task instead of bypassing it. - For task creation, movement, reissue, archive, or triage, read the Task API skill first: .agents/skills/task-api/SKILL.md.
- Do not commit, push, amend, or mutate remotes unless this exact interactive task asks for it. Managed task runs leave git ownership to the platform; see docs/operations/git/commit-push-doctrine.md.
- Platform-owned commits are remote-durability boundaries: push them immediately by default in every managed repository. Keep network work off the run path, retry failures with bounded backoff, and surface final failures as typed operator-feed events rather than treating a local commit as durable.
- Follow the canonical contribution and style guide linked above. Its language policy applies to UI strings, backend errors shown to the UI, prompts, comments, docs, commits, and PR text. Do not introduce em dashes.
- When adding a document under
docs/, add one row to docs/start/README.md in the same change. - CLI crashes, run-outcome classification, retries, or orchestrator drive-to-conclusion: read docs/concepts/orchestrator-drive-to-conclusion.html before changing that logic, and maintain it after. Append each incident to its case log (date, slug, what crashed, which terminal it reached).
- A repeatedly busy orphan worktree is a bounded drive-to-conclusion failure:
after five preparation attempts, escalate it with a visible
worktree-blockedgate item containing the busy path, and pause retries. Repeatedpick-reverted-no-runnotices remain rate-limited per task. - Agent shell policy: default to bash/sh, prefer existing
.shscripts, and avoid PowerShell-specific file creation.
Domain Maps
| Area | Read first | Owns |
|---|---|---|
| Runner | docs/system/domains/runner.md | Pickup, CLI run loop, outcome policy, supervisor loops, and the standalone remote runner (runner/, runbook). |
| Pipeline | docs/system/domains/pipeline.md | Pre/core/post steps, pipeline history, step contracts, and the review/aspect evidence contract (branch diff + results/ inventory + card mode; when "deliverables missing" is legitimate). |
| Review | docs/system/domains/review.md | Remote Review material, merge-base diff budgets, semantic verdict citations, grading, and acceptance-rail behavior. |
| Model routing | docs/system/domains/model-routing-policy.md | Canonical model and thinking-level tiers, weighted selection, correctness floors, benchmark evidence, and quota handling. |
| Tasks | docs/system/domains/tasks.md | Job folders, lane states, API mutations, task access, Epic lifecycle and archive-inclusive history, and the project onboarding and central-store contract. |
| Frontend | docs/system/domains/frontend.md | Angular surfaces, design system, Playwright proof. |
| Style guides | docs/quality/README.md | Technology-aware Angular and .NET guidance selected into coding prompts; incorporates the hard design baseline. |
| Design rules | docs/quality/design/style-guide-hard-rules.md | Hard, non-negotiable design rules (no left accent bars, full-bleed views, aggregate = sum of visible children, acute-only signals, both themes). |
| CLI | docs/system/domains/cli.md | Claude, Codex, Copilot, Gemini drivers and quota probes. |
| ADRs | docs/system/architecture/decisions/adr-archive.md | Load-bearing decisions and deliberate non-goals. |
| Skills | .agents/skills/README.md | Portable specialist workflows. |
Product Boundaries
- Default task execution is sequential within one project.
- Parallelism across projects is in scope.
- Opt-in intra-project parallelism exists only through orchestrator-gated
maxParallelism, isolated worktrees, and pipeline-owned git steps (ADR-0052). - Do not add workflow engines, per-task workspaces, unbounded fan-out, or a design where the run agent manages git.
- The Companion App is reference code only. Do not extend it or advertise it as current product capability.
Runtime And Stable
- Dev backend port 5030 is offline by default. Agents may bring it up only via a Playwright spec from stable that uses frontend/e2e/fixtures/dev-backend.ts.
- Do not start dev backend from a supervisor session, auto-mode loop, background watcher, or routine shell command.
- Stable is the supervisor seat. Never update stable mid-run. After a stable restart, verify runner mode and retry the resume path until the intended mode is visible.
- Backend lifecycle scripts are shell scripts. Prefer repo scripts such as api.sh over ad-hoc process control.
Verification
- Regression reports need data first: reproduce, measure, find the test gap, write the failing regression test, fix, then repeat the original measurement.
- Prompt changes under
prompts/runtime/are CLI behavior changes. Run the matching live probe, such as the@billablehello-world spec, or state why it could not be run. - Visual or behavioral frontend changes require relevant Playwright coverage and
screenshots. Persist review screenshots under the task
results/folder when they must survivetest-results/cleanup. - Pure documentation-only changes do not need builds or Playwright. Verify the navigation, link targets, and diff.
Finish Criteria
- Keep changes scoped to the task.
- Update docs and ADRs only when the change affects product direction, public behavior, architecture, CLI contracts, filesystem contracts, or agent workflow.
- ADRs are for load-bearing decisions and deliberate non-goals, not changelog entries or bug-fix notes.
- Report changed files and verification. If no documentation update is needed for a code change, say so briefly.