ARCADIA-Studio/AGENTS.md

Agent working agreement

Working method

• Follow PROJECT_GOAL.md and the user's current instructions. Work autonomously within scope. Resolve discoverable questions yourself; record, test, and revise reversible assumptions. Ask only when a consequential decision needs information or authorization not already available. Continue independent useful work while blocked elsewhere.
• Map unfamiliar domains through primary sources, comparable systems, and real user workflows. Identify underlying needs, valuable concepts, and avoidable friction; evaluate improvements rather than merely reproducing features. Synthesize findings into decisions or experiments, and stop research when another action offers greater value.
• Make specifications precise enough for compatible independent implementations: definitions, relationships, invariants, valid/invalid examples, and lifecycle or transformation rules where relevant. Trace important requirements and decisions to implementation and verification. Resolve specification ambiguity at its source when implementation exposes it.
• For consequential uncertain choices, compare materially different candidates against explicit user-derived criteria. Ask independent critics to challenge assumptions and attempt to falsify proposals; refine or replace candidates as evidence warrants. Record the choice, evidence, tradeoffs, rejected approaches, and reasons to revisit it. Use proportionate effort for reversible local choices.
• Plan milestones that resolve uncertainty or deliver useful capability, with explicit dependencies. Test foundational assumptions before substantial downstream work. At checkpoints, assess progress toward the actual objective and replan when evidence invalidates the strategy or progress stalls.
• Define observable success and suitable evaluators before repeated improvement: invariants, examples, round trips, transformations, performance checks, realistic user tasks, or explicit review rubrics. Ensure the evaluator measures the intended outcome. For experiments, record the question, expectations, method, result, and implications; isolate exploratory prototypes from production architecture.
• Build and run meaningful changes, exercise actual behavior, and inspect useful logs. Investigate crashes, serious warnings, corrupted state, and interaction defects. For GUI work, use Computer Use or equivalent tools to inspect realistic workflows, cognitive load, editing/selection, navigation, feedback, recovery, loading/empty/failure states, accessibility, layout, focus/disabled states, and responsiveness.
• Match independent reviews to the artifact and risks. For languages or formats, include ambiguity, round trips, human/AI editability, and version-control behavior where relevant. Verify findings, fix the appropriate conceptual, specification, architectural, implementation, or interaction layer, and recheck affected dependents. Reuse valid evidence; repeat checks for concrete risks, changes, findings, or required gates. Compilation, screenshots, and worker reports alone do not establish correctness or usability.
• Favor the simplest maintainable design satisfying known requirements and justified extensibility. Avoid speculative abstraction, unrelated refactoring, and repeating approaches already invalidated by evidence.

User input

All content between BEGIN USER-OWNED APPLICATION SPECIFICATION and END USER-OWNED APPLICATION SPECIFICATION in PROJECT_GOAL.md belongs to the user. Its headings and structure are unrestricted. Preserve it exactly unless the user explicitly requests an edit; cleanup and task completion are not permission to rewrite it.

Derive specifications, plans, and acceptance checks elsewhere, linking to their source. Distinguish user requirements, verified facts, assumptions, open questions, candidate approaches, and accepted decisions. Interpretations must not silently replace the user's intent. When input changes, reconcile affected work. Check current evidence before reopening a reported problem merely because its original description remains in the input.

Shared records

Keep one authoritative home for each kind of information. Use links and brief scoped summaries elsewhere; update only what materially changed. Local planning tools and scratch notes need no routine synchronization.

Information	Authoritative home
Live plans, priorities, dependencies, ownership, status, blockers, and handoffs	Gitea issues/milestones; linked PRs hold change review and validation.
Specifications, architecture, interfaces, and build/run documentation	Established version-controlled project documents.
Durable decisions and reusable evidence	harness/DECISIONS.md and relevant topic notes, or existing equivalents.
Navigation and integrated-state checkpoint	harness/STATE.md: links, verified code commit/release, capabilities, and material limitations. No live task ledger.
Shared entrypoint	Gitea Wiki: overview and links, without duplicating versioned documentation.
Published deliverables	Gitea releases, tied to source commits and verification.

Use Gitea for live coordination and the relevant code version for implementation facts. Required behavior comes from the user and supporting specifications; observed behavior and issue closure do not redefine requirements.

Coordination and integration

1. On starting or resuming, including after compaction, inspect the working branch, local changes, applicable instructions, relevant shared records, and existing implementation. Recover context from authoritative records. Check for newer shared instructions on the actual remote default branch without overwriting local work. Read what is relevant; avoid repeated full-history scans and constant polling.
2. Search before creating tasks. Claim a meaningful outcome in its Gitea issue with a stable session identity, scope, branch/worktree, and acceptance criteria. Keep small steps within that task. Recheck after claiming; comments are not atomic locks. Respect existing ownership and resolve overlapping claims before overlapping work.
3. Independent sessions are peers. Each owns its contribution and delegated work. Record one integration/release owner when a milestone needs shared coordination. Use separate worktrees or checkouts, partition files/interfaces, and coordinate changes to shared specifications and documents.
4. At meaningful checkpoints or handoffs, record changed results, commit/PR links, verification commands and environment, blockers, and next actions. Identify the tested code version and uncommitted or unpublished work; make required artifacts accessible to other sessions. Do not assume another agent is running or has read a message; silence alone does not establish abandonment. Record evidence and preserve existing work when reassigning a task.
5. Follow project review and merge conventions. Inspect upstream changes, resolve conflicts while preserving others' edits, and verify the combined result. Do not overwrite others' work or force-push shared history to resolve coordination problems.
6. Close work when its acceptance criteria are met and the result is integrated or delivered as appropriate. Mark duplicates and superseded tasks explicitly with replacement links. Then choose further useful unowned work while the requested objective remains incomplete.

Use Gitea MCP for shared coordination. If unavailable, report the failure, buffer necessary handoff information locally, and continue already-owned or clearly non-overlapping work. Reconcile with current Gitea state when access returns; do not claim an ownership update, review, or publication succeeded without confirmation.

Delegation and reasoning

Use GPT-6 Astra exclusively. Delegate bounded independent research, design, implementation, or review when it improves progress, judgment, or preserves the parent's context. Give subagents the relevant context, ownership boundaries, result format, and evidence requirements. Use fresh critics when practical. The parent owns synthesis and integration and normally consolidates results into Gitea. Choose concurrency according to useful independent work and available resources, including other sessions; there is no agent-count quota.

Use the highest available reasoning effort for orchestration and foundational semantic, methodological, formal, or cross-domain decisions. For subagents, use extra-high for difficult design, adversarial review, or unusually hard implementation; high for normal implementation, targeted research, correctness review, debugging, and runtime evaluation; medium for routine exploration, mechanical changes, builds, and tests; and lower for trivial work. Allocate by difficulty, not length, using supported controls; do not claim an unapplied change or alter global configuration merely to enforce these preferences.

Delegation and reasoning

Use GPT-6 Astra exclusively. Delegate bounded independent research, design, implementation, or review when it improves progress, judgment, or preserves the parent's context. Give subagents the relevant context, ownership boundaries, result format, and evidence requirements. Use fresh critics when practical. The parent owns synthesis and integration and normally consolidates results into Gitea. Choose concurrency according to useful independent work and available resources, including other sessions; there is no agent-count quota.

Optimize reasoning effort for verified outcomes per cost, including review, retries, and repair:

Main agent: default to high for planning, synthesis, integration, and acceptance decisions; medium for dispatching a settled plan and routine coordination.

Subagents: default to medium for bounded implementation and targeted research; low for straightforward retrieval, mechanical edits, and running established checks; high for complex debugging, correctness review, adversarial analysis, or ambiguous design.

Use extra-high selectively for difficult foundational semantic, methodological, formal, or cross-domain decisions. Reserve max for exceptional unresolved problems where additional reasoning is likely to justify its cost. Retain higher settings when evidence supports their value.

Allocate by task difficulty and consequences, not role or length alone. Start higher when a clearly difficult decision has expensive downstream consequences. Escalate for concrete reasoning difficulties; resolve missing evidence or unclear requirements at their source. Set subagent effort explicitly through supported controls to avoid unintended inheritance. Do not claim an unapplied change or alter global configuration merely to enforce these preferences.

Delivery

Publish versioned Gitea releases at meaningful working milestones; mark unfinished versions as prereleases. Check existing releases/tags and coordinate one owner and version before publishing. Build from an identified integrated commit and verify the packaged result and launch steps.

Include runnable builds or appropriate artifacts, concise release notes, build/run/deployment instructions as applicable, known limitations, tested platforms, and verification evidence. Confirm publication and artifact availability. Use existing authorized API access or project tooling when MCP lacks a needed release/upload operation. Do not silently replace published versions with different contents.