Skip to main content

Copilot Workspace Instructions

Use this file as the GitHub Copilot entry point for workspace-wide guidance. AGENTS.md and CLAUDE.md remain sibling mirrors for other tooling; they are not the only workflow definition for docs-only Copilot waves.

  • For docs-only Copilot waves, GPT-5.5 high is the default committed model policy for current wave-critical Copilot agents (explorer, implementer, verifier, and resolver). Do not inherit the active Copilot model, and do not assign any current wave-critical Copilot agent to GPT-5.5 medium.
  • For Copilot implementation or validation work outside a named-chain carve-out, use the most specific available sub-agent for the requested subsystem or role and run it on GPT-5.5 high. Do not substitute a generic custom agent just because it exists, and do not use non-GPT-5.5 or GPT-5.5 medium for wave-critical Copilot agent work.
  • If the required specialist sub-agent is not available and the prompt does not explicitly require an exact named chain with a blocked-on-missing rule, dispatch the general-purpose agent on GPT-5.5 high with the full user prompt, repo tenets, forbidden-survivor list, validation requirements, and live-proof requirements. The fallback general agent must implement the prompt literally; it must not reinterpret a rewrite as a gap-fix, audit, sync, or preservation pass.
  • The orchestrator dispatches waves. Workers do not run multi-wave work inline, do not spawn other agents, and do not use scheduling/background tools.
  • If a prompt assigns the main agent as orchestrator or requires the exact named chain explorer -> implementer -> verifier, the orchestrator must perform real dispatches to those exact agents and must not simulate them with prose, todos, status text, plans, or inline work. If those dispatches are unavailable, stop and report exactly: Blocked: this environment cannot perform the required explorer -> implementer -> verifier named-agent orchestration.
  • In docs-only named-chain hard mode, every explorer, implementer, and verifier dispatch prompt must begin with the full hardening prefix in COPILOT_ORCHESTRATION_HARDENING_PROMPT.txt, placed before the wave brief, task body, or failure list. Missing or altered prefix text is a process FAIL and verifier must report it from dispatch prompt evidence.
  • Named-chain hard mode is stop-on-drift. If the orchestrator notices inline exploration, broad repo inspection, or non-agent analysis after hard mode activates, stop and report PROCESS FAIL: named-chain hard-mode drift rather than continuing inline.
  • The zero-exception git ban in orchestration hard mode covers mutating/destructive git or file-recovery commands, not read-only git evidence gathering already allowed by workspace agent guidance.
  • Resolver is for concrete blockers that would otherwise force a user question or halt. Do not use it for ordinary planning, coverage-matrix choices, or routine artifact selection.
  • Copilot workspace agents have the same repository file access as the orchestrator. Inspect committed docs and guidance in the workspace before asking the user for files already present.
  • Source of truth comes from committed repository docs first, then the current approved prompt. Chat history, Repomix bundles, and uploaded summaries are seed evidence only — always re-read the live committed file before acting on any prior-session memory of its contents. Mark proposals as proposals, keep missing evidence explicit, and do not invent files, facts, APIs, canon, lore, mechanics, diagnostics, or provenance links.
  • The current approved prompt outranks agent interpretation. If the prompt says rewrite, destructive rewrite, no backwards compatibility, no preservation, or names forbidden survivor surfaces, agents must treat existing code as evidence only and must fail rather than downscope the task to patching, polishing, wiring, tests, counters, or an "already mostly done" claim.
  • Question-only user turns are not permission to edit. If the user asks what changed, why something changed, what they asked for, or what should change, answer first without editing unless the user explicitly asks for a file change.
  • After a mistaken or disputed edit, do not auto-revert, restore, or rewrite files unless the user explicitly instructs you to do so. Answer the question first; wait for a direct edit instruction before touching files again.
  • When updating an existing prompt, instruction, or guidance file, preserve the document and patch it in place. Do not replace the whole file with a rewritten version unless the user explicitly asks for a rewrite or replacement.
  • For Valenar docs and guidance work, follow .github/instructions/valenar-docs-wave.instructions.md. That scoped instruction defines the allowed write surface, the explorer -> implementer -> verifier workflow, and the Valenar canon guard, including activity as the canonical behavior / planning noun and retained on_action metadata / extension-point vocabulary.
  • The exact Copilot docs-only custom agent names in this workspace are explorer, implementer, and verifier.
  • In docs-only waves, Verify is the validation step name, but the actual verification agent dispatch is verifier. Verify FAIL returns directly to implementer.
  • In docs-only named-chain hard mode, after verifier FAIL, only re-dispatch implementer with the exact failure list, then re-dispatch verifier. No inline repair, prompt patching, or orchestrator synthesis in between.
  • For mapv10 waves, use the exact Copilot custom agent chain mapv10-explore -> mapv10-implement -> mapv10-validate -> mapv10-tenets-check. GPT-5.5 high is the default committed model policy for the mapv10 chain. Do not inherit the active Copilot model, do not collapse Validate and Tenets, loop FAIL verdicts back to mapv10-implement until Step 4 returns PASS, and do not assign any current wave-critical Copilot agent to GPT-5.5 medium.
  • Every mapv10 agent dispatch prompt must start with the mandatory TENETS plus forbidden-pattern prefix from examples/map/mapv10/wave-protocol.md Part 0. Put that prefix before the Wave Brief or any failure list. mapv10-validate and mapv10-tenets-check must fail the wave on process grounds if dispatch prompt evidence shows the prefix was absent, altered, or not first.

Keywords: wave-protocol-auto-iteration, orchestrator-not-worker, copilot-workspace-agent, gpt-5.5-high, gpt-5.5-high-general-agent-fallback, specialist-sub-agent-first, prompt-outranks-agent-interpretation, rewrite-not-gap-fix, forbidden-survivor-fail, docs-only-gpt-5.5-high-default-policy, docs-only-explorer-gpt-5.5-high, docs-only-implementer-gpt-5.5-high, docs-only-verifier-gpt-5.5-high, docs-only-resolver-gpt-5.5-high, mapv10-wave, mapv10-gpt-5.5-high-default-policy, mapv10-explore, mapv10-implement, mapv10-validate, mapv10-tenets-check, mapv10-explore-gpt-5.5-high, mapv10-implement-gpt-5.5-high, mapv10-validate-gpt-5.5-high, mapv10-tenets-check-gpt-5.5-high, model-downgrade-blocked, gpt-5.5-medium-blocked-for-wave-critical-agents, agent-dispatch-required, explore-implement-verify-required, sub-agent-no-spawn-task, sub-agent-same-workspace-file-access, no-inline-wave-execution, parallel-agent-dispatch-when-safe, fail-validation-loops-to-implement, verify-subagent-prompts, verify-deviation-notes, fail-on-bad-agent-prompt, docs-only-hardening-prefix, named-chain-hard-mode-stop-on-drift, blocker-only-resolver, mutating-git-ban-not-readonly-git, verifier-itemized-pass-fail, source-of-truth-first, no-invented-project-facts, proposed-is-not-committed, missing-files-not-invented, docs-only-wave, no-code-edits, no-content-secs-edits, no-generated-edits, no-runtime-edits, no-client-edits, no-server-edits, no-test-edits, no-invented-game-canon, canon-vs-proposal-explicit, no-unapproved-lore-retcon, mechanics-change-requires-design-record, balance-numbers-require-source, worldbuilding-source-of-truth-first, valenar-no-legacy-docs, valenar-acts-not-stages, valenar-core-not-settlement, valenar-core-not-realm-birth, valenar-camp-core-settlement-separated, valenar-act-1-first-people, valenar-act-2-first-outpost, valenar-sites-first-class, valenar-dungeons-are-sites, valenar-characters-skills-not-fixed-jobs, valenar-pops-fill-ordinary-jobs, valenar-clues-missions-objectives, valenar-activity-canonical-behavior-planning, valenar-on-action-metadata-retained, valenar-reclamation-starts-early-full-reclamation-last, secs-source-hygiene