git:20260713.b41a73b to git:20260721.bd1e67a

- # AGENTS.md

- <!-- Generated by AIWG. Edit AGENTS.override.md for operator additions. -->

-

- ## Framework Context

-

- See [AIWG.md](./AIWG.md) for the full AIWG framework context

- (active frameworks, addons, agents, behaviors, rules).

- Tracker and delivery source of truth: [.aiwg/aiwg.config](./.aiwg/aiwg.config).

-

- Deployed artifacts live under your provider's native directory

- (for example `.codex/agents/`, `.warp/agents/`, `.github/agents/`).

- Use `aiwg discover "<intent>"` and `aiwg show <type> <name>` to browse

- skills, agents, rules, and commands across the installation.

-

- ## Tier 1 / Tier 2 / Tier 3 Loading Model

-

- - Tier 1: this bridge plus the finalization block; always-loaded orientation only.

- - Tier 2: compact quickref routing summaries; enough to choose a next lookup.

- - Tier 3: full skill, rule, agent, behavior, docs, and examples; load only through `aiwg discover` / `aiwg show` or an explicit skill invocation.

-

- ## Tier 2 Capability Map

-

- This is a quickref-style routing layer. Keep it in default context; load Tier 3 detail only through `aiwg discover` and `aiwg show`, not by traversing provider directories.

-

- Schema per entry: purpose, when to use, when not to use, curated discovery phrases, deep-load target, verification cue.

-

- ### Agents

-

- - Purpose: Route specialized work to deployed agent personas without loading their full bodies.

- - When to use: The task names a role, asks for domain review, or needs focused delegation.

- - When not to use: A direct CLI command or loaded skill already handles the work.

- - Curated discovery phrases: `aiwg discover "agent for <task>"`; `aiwg discover "<domain> specialist agent"`; `aiwg discover "review <artifact> agent"`

- - Deep-load target: `aiwg show agent <name>` after discovery selects the exact item.

- - Deployed summary: 195 recorded; examples: Accessibility Checker, Accessibility Specialist, Acquisition Manager, AgentSmith, AI/ML Engineer, +190 more.

- - Verification cue: Selected agent name and scope match the user request before delegation.

-

- ### Rules

-

- - Purpose: Expose policy and behavioral constraints as compact routing anchors.

- - When to use: The task involves safety, provider behavior, workflow discipline, or repo policy.

- - When not to use: The user asks for broad framework discovery rather than a specific guardrail.

- - Curated discovery phrases: `aiwg discover "rule for <constraint>"`; `aiwg discover "<provider> routing rule"`; `aiwg discover "context budget rule"`

- - Deep-load target: `aiwg show rule <name>` after discovery selects the exact item.

- - Deployed summary: 43 recorded; examples: RULES-INDEX, RULES-ONDEMAND, agent-deployment, api-abi-stability, auto-compact-continue, +38 more.

- - Verification cue: The applied rule is named and the action taken conforms to it.

-

- ### Skills

-

- - Purpose: Route operator intents to executable AIWG workflows and procedural guidance.

- - When to use: The user names an AIWG capability, workflow, issue action, or framework task.

- - When not to use: The task is ordinary code editing with no AIWG-specific workflow needed.

- - Curated discovery phrases: `aiwg discover "skill for <workflow>"`; `aiwg discover "address issues"`; `aiwg discover "regenerate context"`

- - Deep-load target: `aiwg show skill <name>` after discovery selects the exact item.

- - Deployed summary: 0 recorded; examples: none deployed.

- - Verification cue: The selected skill was read before invoking its workflow.

-

- ### Behaviors

-

- - Purpose: Summarize reactive behavior packs without carrying long interaction detail.

- - When to use: The task depends on session behavior, daemon interaction, or event-driven agent conduct.

- - When not to use: A static rule or skill directly covers the requested action.

- - Curated discovery phrases: `aiwg discover "behavior for <interaction>"`; `aiwg discover "daemon behavior"`; `aiwg discover "session behavior"`

- - Deep-load target: `aiwg show behavior <name>` after discovery selects the exact item.

- - Deployed summary: 43 recorded; examples: RULES-INDEX, RULES-ONDEMAND, agent-deployment, api-abi-stability, auto-compact-continue, +38 more.

- - Verification cue: The behavior trigger and expected effect are explicit before relying on it.

-

-

- <!-- aiwg-context-finalization:START -->

- ## Context Finalization

-

- This section is synthesized after template emission from the current workspace state. Preserve operator-authored content outside AIWG-managed blocks; rerun `aiwg regenerate` to refresh this section after provider, framework, or MCP wiring changes.

-

- ### Workspace Snapshot

-

- - Configured providers: claude, codex

- - Installed frameworks/addons: sdlc, media-marketing, all, security-engineering

- - Recorded deployments: claude, codex, copilot, cursor, factory, hermes, openclaw, opencode, openhuman, warp, windsurf

- - Normalized project context: `.aiwg/AIWG.md`

-

- ### Discover-First Protocol

-

- Classify every user turn FIRST: is it a **new directive** or a continuation? When a message names or references an AIWG command/capability — even as pasted content like an `address-issues` tracker table, an issue list, or a `flow-*` name — treat it as a new directive and ACT: run `aiwg discover "<the need>"`, fetch with `aiwg show <type> <name>`, and invoke it. Do NOT ask "what would you like me to do with these?" when the action is implied — a pasted `address-issues #1234` table means run the address-issues workflow on those issues.

-

- Also run `aiwg discover` before declining an AIWG request as out of scope or inventing a workflow from memory. The CLI ranks AIWG capabilities across the installed corpus and rebuilds the index from `$AIWG_ROOT` automatically, so a "no matches" for a command you know is deployed is a bug — not a signal it is absent. Commands AIWG deploys to your provider command directory (`.opencode/command/`, `.claude/commands/`, `~/.codex/prompts/`, …) ARE discoverable this way; fetch them with `aiwg show command <name>`. This prevents decline-without-search failures, ask-instead-of-act on new directives, and hallucinated skill or agent names. Full rule: `agentic/code/addons/aiwg-utils/rules/skill-discovery.md`.

-

- ### Engagement Verification

-

- When a user asks whether AIWG is active or engaged in this project, run or read `aiwg status --probe --json` and report the result plainly: engaged state, project root, deployed provider files, installed frameworks/addons, and the next action from the probe. Do not add AIWG attribution, signatures, generated-by text, or passive footers to user files, commits, PRs, comments, code headers, or docs.

-

- ### Tracker Authority Protocol

-

- - Source of truth: [.aiwg/aiwg.config](./.aiwg/aiwg.config)

- - Canonical tracker: `origin` (unknown; git@git.integrolabs.net:roctinam/aiwg.git)

- - Primary repo remote: `origin`; CI remote: `origin`

- - Secondary/mirror remotes: github (public-mirror)

- - Issue storage mode: not configured

-

- Tracker access order for issue, PR, release, and CI-sensitive tracker operations:

- 1. MCP/app tools for the configured tracker.

- 2. Tracker HTTP API with configured credentials.

- 3. Tracker CLI for the configured tracker, after confirming authentication.

- 4. Stop and report a blocker.

-

- - Project config decides tracker authority; installed/authenticated CLIs do not.

- - Git SSH remote access is repository sync, not issue-tracker API access.

- - Do not file on mirror or secondary remotes just because their CLI is authenticated.

- - Treat an unauthenticated tracker CLI as one failed access path, then continue probing MCP/app/API before blocking.

-

- ### Source Model

-

- - `.aiwg/AIWG.md` is the normalized project-local context entry point.

- - Root `AIWG.md` is the generated cross-provider companion loaded through `AGENTS.md` and provider twins.

- - `AGENTS.md`, `WARP.md`, `.hermes.md`, and `.github/copilot-instructions.md` are provider-facing bridges, not replacements for `.aiwg/AIWG.md`.

- <!-- aiwg-context-finalization:END -->

-

- <!-- AIWG-PARALLELISM-CAP:START -->

- ## Parallelism Cap

-

- This project caps parallel agent fan-out (#1359):

-

- - **max_parallel_subagents**: 4 (provider default for claude)

- - **max_parallel_ralph_loops**: 2 (provider default for claude)

- - **max_parallel_mc_missions**: 4 (provider default for claude)

- *Rationale*: Provider default for claude (migrated by aiwg refresh)

- When spawning parallel subagents, take the MIN of: this cap, `AIWG_CONTEXT_WINDOW` budget, the RLM 7-agent hard cap (RLM dispatches only), and the natural task decomposition. Bump via `aiwg config set --project parallelism.max_parallel_subagents N`.

- <!-- AIWG-PARALLELISM-CAP:END -->

- ---

- *See `AGENTS.override.md` for operator-authored additions.*