docs: simplify agent context and resolve conflicting guidance (#2287)

* docs: clarify agent task scope and validation

* docs: remove redundant and conflicting agent guidance

Author: thymikee

AGENTS.md

@@ -3,8 +3,9 @@
 `agent-device` is a CLI and daemon for automating Apple, Android, HarmonyOS, Vega, Linux, and web
 targets. A long-lived daemon owns sessions; registry-derived commands route to platform runtimes.
 
-This file contains repository-specific traps and invariants that are expensive to rediscover. Load
-task procedures only when needed:
+## Task routing
+
+Load only the procedures relevant to the task:
 
 | When the task involves | Read |
 | --- | --- |
@@ -14,35 +15,28 @@ task procedures only when needed:
 | Selector capture, polling, or interaction fast paths | `docs/agents/selector-capture.md` |
 | Adding or changing a CLI flag | `docs/agents/cli-flags.md` |
 | Opening or reviewing a PR | `docs/agents/pull-requests.md` |
-| Running against a real device | `docs/agents/device-verification.md` |
+| Apple runner changes or manual device verification | `docs/agents/device-verification.md` |
 | Writing issues or PRDs, and triage labels | `docs/agents/issue-tracker.md`, `docs/agents/triage-labels.md` |
 | Web backend setup or diagnostics | `docs/agents/web-backend.md` |
 
-Versioned CLI help is the source of truth for command behavior. Start with `agent-device help
-workflow`, then the relevant topic help.
+Versioned CLI help owns command semantics. Read the relevant topic when behavior is unclear;
+normal app-driving startup follows the skill's routing card.
 
 ## Incident-derived principles
 
-- Guarantees erode at path boundaries. Classify every interaction dispatch path in
-  `packages/contracts/src/interaction-guarantees.ts`; a registry claim is not proof that the native
-  implementation satisfies the guarantee's definition.
+- Classify interaction dispatch paths in `packages/contracts/src/interaction-guarantees.ts` and
+  verify each claim against the native implementation.
 - Delegation on error proves no success-path parity. A fast path may succeed on a candidate the
   shared rules would refuse.
-- Prove that a code path can fire before measuring it. An A/B test with an unreachable arm is two
-  green runs without evidence.
-- A green regression test counts only after it has been observed red against the pre-fix code. Plant
-  a violation for new structural gates and verify the gate names the invariant.
-- Repair recurring failures at their owning interface with types, a registry, or one construction
-  path. A custom guard that reconstructs another source of truth needs redesign, not another
-  exception.
-- Treat explanatory implementation comments as a failed design review. Do not narrate control flow,
-  preserve review history, or justify a workaround in code. Express the invariant through names,
-  types, module boundaries, and tests; put history in the PR or an ADR. Allow only public API docs,
-  tool directives, and a brief citation to an external constraint that cannot be encoded.
-- Key behavior on typed reasons and details, never error text. Existing message sniffs are owned debt
-  and must not be copied.
-- Snapshot output is the token budget: do not add per-node metadata when response-level metadata can
-  be emitted once. Append warnings through the shared response builder; never replace prior warnings.
+- Prove both paths are reachable before measuring a fast path against its baseline.
+- Repair recurring failures at the owning type, registry, or construction path; do not add guards
+  that reconstruct another source of truth.
+- Keep control-flow narration and review history out of implementation comments. Encode invariants
+  in names, types, boundaries, and tests; reserve comments for public API docs, tool directives,
+  and non-obvious constraints that cannot be encoded.
+- Key behavior on typed reasons and details, never error text; do not copy existing message sniffs.
+- Keep metadata at response level when it applies to the whole snapshot. Append warnings through
+  the shared response builder; never replace prior warnings.
 - Before preserving a compatibility shape, run `git tag --contains <commit>`. Unreleased API has no
   external compatibility obligation.
 
@@ -66,10 +60,10 @@ and progress in `@agent-device/capture-kit` (`request-cancel`, `request-progress
 contracts in `@agent-device/contracts`; CLI flags in `src/commands/cli-grammar`; cross-surface schema
 composition in `src/cli-schema`.
 
-The enforced registries are self-declaring. A failing completeness, parity, coverage, timeout,
-layering, or construction gate means the new cell or path is unclassified; do not suppress or
-allowlist it. Interaction responses are built only through `buildInteractionResponseData`, and
-cross-language rules change through golden tables under `contracts/fixtures/`.
+Resolve registry completeness failures at the missing declaration. Diagnose other gate failures
+at their reported invariant; do not suppress them or add an allowlist to get a pass. Build interaction
+responses through `buildInteractionResponseData`; change cross-language rules through golden tables
+under `contracts/fixtures/`.
 
 ## Hard repository rules
 
@@ -78,9 +72,9 @@ cross-language rules change through golden tables under `contracts/fixtures/`.
 - Apple target changes keep the kernel device model, runtime-fact admission, dispatch resolution,
   Apple discovery, and xctestrun preparation in sync.
 - iOS simulator-set scoping must never hide the host macOS desktop target.
-- Skills may carry a minimal start/routing card; command semantics belong in versioned CLI help.
-- Do not add compatibility or fallback behavior without explicit approval. Complete migrations and
-  remove superseded paths.
+- Skills stay minimal routing cards.
+- Add compatibility or fallback behavior only when explicitly requested or approved. Otherwise,
+  complete the migration and remove superseded paths.
 - Keep changes within one command family or module group unless the task explicitly crosses a
   boundary. Platform-neutral work does not license inspecting every platform implementation.
 
@@ -93,7 +87,6 @@ cross-language rules change through golden tables under `contracts/fixtures/`.
 - Tests mirror source topology one-to-one. Split a source module and its test together; do not add to
   the legacy `interaction.test.ts` or platform `index.test.ts` aggregations. Pure moves carry their
   tests unchanged; rename-only hunks owe no new coverage.
-- Shared fixtures are named exports in a sibling fixture module, not repeated inline literals.
 - `src/daemon/handlers/session.ts` is already over budget; extract the relevant platform-specific
   concept before adding behavior.
 
@@ -103,30 +96,14 @@ cross-language rules change through golden tables under `contracts/fixtures/`.
   repository-wide, not path-scoped.
 - A fresh worktree requires `pnpm install --frozen-lockfile && pnpm build`. Until then package and
   optional-peer resolution may point at another checkout and produce false failures.
-- Source-checkout daemon state is worktree-scoped, but devices are not. Use `pnpm daemon:state-dir`
-  to inspect it and different devices for concurrent worktrees.
-- Run one full gate per host at a time. Subprocess-backed tests under concurrent worktrees produce
-  timeout-shaped contention failures.
-- Before pushing, run `pnpm check:affected --run && git push`. Use `pnpm check` for broad refactors.
-  GitHub remains authoritative for native, device, provider, and full-coverage lanes.
+- Parallel work needs disjoint edit ownership and distinct devices. Run one full gate per host;
+  concurrent subprocess-backed suites can produce contention timeouts.
 - The layering scan reads tracked files only. Stage a new module before trusting its result.
 - Fallow baselines are path-keyed. Move the matching baseline entry when renaming a file; never bulk
   regenerate baselines to accept unrelated findings.
-- The first Node process after a newly signed Apple runner launches may block during Gatekeeper
-  verification. Warm it with a throwaway `node -e 0` before measuring.
-- `DEVICE_IN_USE` has two flavors. "already in use by session X" is this daemon — follow its
-  `close --session` hint. "owned by session X in workspace Y" is another worktree's device
-  claim — non-retriable; run the error's `device status`/`device release --stale` recovery,
-  never PID hunting.
-- A changing timeout failure set that passes in isolation is host contention. Reproduce the same
-  test on `origin/main` under the same load before treating it as a regression.
 
 ## Runtime and diagnostics seams
 
-The OS-neutral Apple runner lives under `packages/platform-apple/src/runner/`. For connection errors,
-retry policy, or command typing, start at `runner-contract.ts`; transport stays below session/client
-behavior, and xctestrun build/cache logic stays outside request execution.
-
 Diagnostics use `@agent-device/capture-kit/diagnostics`. Request diagnostics belong in the session request log;
 session artifact paths come from `src/daemon/session-store.ts`. App/device logs remain in `app.log`;
 Apple runner and xcodebuild output remains in `runner.log`.
@@ -150,11 +127,8 @@ translate through the interaction root frame. Prefer selector or ref tests over
 
 ## Documentation ownership
 
-Before adding guidance, decide whether the command surface, CLI grammar/help, MCP projection, daemon
-runtime, ADR, or task procedure owns it. Link to executable registries instead of copying their
-contents. Keep a sentence in this file only when no gate, lint rule, versioned help, ADR, or
-decision-site comment can own it. `CONTEXT.md` is glossary-only: no implementation paths,
-architecture decisions, migration state, or workflows.
+Keep this file to repository traps and routing. Put task procedures in `docs/agents/`, decisions in
+ADRs, and command semantics in versioned help. Link to executable registries instead of duplicating
+their rules. `CONTEXT.md` is glossary-only: no implementation paths, decisions, or migration status.
 
-Behavior changes update their owning help/metadata and user docs. In the final summary,
-state whether docs or skills changed and why.
+Behavior changes update owning help/metadata and user docs. Report docs or skill changes when relevant.

docs/agents/cli-flags.md

@@ -1,16 +1,14 @@
 # Adding a CLI Flag
 
-A new flag touches only the layers that need to understand it. Stop at the layer where it stops
-mattering — threading it further is the common failure, not stopping too early.
+Thread a flag only through the layers that consume it:
 
 1. `packages/contracts/src/cli-flags.ts`: add to `CliFlags`; add the definition to the matching
    `src/commands/cli-grammar/flag-definitions-*.ts` owner and the relevant group in `flag-groups.ts`
    (for example `SNAPSHOT_FLAGS`). Then update the command family metadata/schema that exposes the
    flag; find the owner with
    `rg -n "<command>|supportedFlags|allowedFlags" src/commands src/cli-schema src/cli/parser`. For
-   schema-only CLI commands (`cdp`, `auth`, `connect`, `proxy`, `react-devtools`, `web`) the owner is
-   `SCHEMA_ONLY_CLI_COMMAND_SCHEMAS` in `src/cli-schema/command-overrides.ts`. New flags are
-   operator-only by default. Add a flag to `PROJECT_CONFIG_FLAG_KEYS` in
+   schema-only CLI commands, the owner is `SCHEMA_ONLY_CLI_COMMAND_SCHEMAS` in
+   `src/cli-schema/command-overrides.ts`. New flags are operator-only by default. Add a flag to `PROJECT_CONFIG_FLAG_KEYS` in
    `src/cli-schema/cli-config.ts` only when repository control is safe; this positive allowlist is
    the completeness gate.
 2. `src/commands/cli-grammar/*`: read the CLI flag into command input.

docs/agents/device-verification.md

@@ -1,20 +1,18 @@
 # Manual Device Verification
 
-Read this before running `agent-device` by hand against a simulator, emulator, or physical device.
+Read this for Apple runner changes or manual `agent-device` runs on simulators, emulators, or
+physical devices. Live verification steps apply when exercising a device-facing path.
 
-## Before the run: defeat staleness
-
-Dev-loop staleness has three layers, and each produces a convincing false negative.
+## Build freshness
 
 - After changing runtime code reached through `bin/agent-device.mjs` or the daemon: `pnpm build`,
   then `pnpm clean:daemon` — the daemon does not self-reload.
 - Before any Android verification from source: `pnpm build`, `pnpm build:android`, `pnpm clean:daemon`.
   `build:android` refreshes and verifies both bundled Android helper artifacts for the current
   package version.
-- `shutdown` deliberately HANDS OFF a healthy simulator runner. The adopted runner keeps serving the
-  old Swift binary until you kill its process or the source fingerprint changes, so "my change did
-  nothing" measured against an adopted runner is a classic false negative. If Swift runner code
-  changed, run `pnpm build:xcuitest`.
+- `shutdown` hands off a healthy simulator runner; a new daemon may adopt the old binary. After
+  Swift runner changes, run `pnpm build:xcuitest` before verification. Use the session cleanup
+  procedure below if ownership is stuck.
 
 ## Prove the path under test was actually active
 
@@ -25,19 +23,30 @@ Dev-loop staleness has three layers, and each produces a convincing false negati
   simulator, physical-device, Metro/dev-client, and app-surface steps. An already-installed
   `com.callstack.agentdevicelab` is not sufficient — the README's Metro/dev-build and `snapshot -i`
   checks must prove the expected app surface is running.
-- For Android RN/Expo/dev-client apps on any local Metro port, `adb reverse tcp:<port> tcp:<port>` is
-  harmless and should be run before opening the app or URL.
+- For Android RN/Expo/dev-client apps that use local Metro, configure
+  `adb reverse tcp:<port> tcp:<port>` for the app's Metro port before opening the app or URL.
 
-## Session hygiene
+## Worktree ownership and runner diagnostics
+
+- Source-checkout daemon state is worktree-scoped, but devices are not. Use `pnpm daemon:state-dir`
+  to inspect it and different devices for concurrent worktrees.
+- The first Node process after a newly signed Apple runner launches may block during Gatekeeper
+  verification. Warm it with a throwaway `node -e 0` before measuring.
+- `DEVICE_IN_USE` has two flavors. "already in use by session X" is this daemon — follow its
+  `close --session` hint. "owned by session X in workspace Y" is another worktree's device
+  claim — non-retriable; run the error's `device status`/`device release --stale` recovery,
+  never PID hunting.
 
-Every manually opened session is a resource that must be closed — including exploratory sessions and
-failed verification attempts.
+The OS-neutral Apple runner lives under `packages/platform-apple/src/runner/`. For connection errors,
+retry policy, or command typing, start at `runner-contract.ts`; transport stays below session/client
+behavior, and xctestrun build/cache logic stays outside request execution.
+
+## Session hygiene
 
-- Every `agent-device open` needs a matching `close` with the same `--session`, `--platform`,
-  `--udid`, and `--state-dir` before the agent finishes.
+- Close manually opened sessions, including failed verification attempts, using their original
+  `--session`, `--platform`, `--udid`, and `--state-dir` values.
 - Use a purpose-specific session name for experiments, and an isolated `--state-dir` under
   `/private/tmp` when you need cleanup isolation beyond the current worktree's default daemon.
-- Track opened sessions in working notes; close each one before the final response.
 - If `close` is blocked or ownership looks stuck, inspect it with
   `agent-device device status --stale` (daemonless), stop the owning daemon with
   `agent-device daemon stop --state-dir <dir>` (add `--clean` to remove retained runners), and
@@ -48,8 +57,7 @@ failed verification attempts.
 
 ## Sandboxed environments
 
-Start the daemon outside the sandbox with escalation. The daemon binds localhost, and sandboxed runs
-fail before any product code executes, with `listen EPERM: operation not permitted 127.0.0.1` or
-repeated `Failed to start daemon`/metadata cleanup messages. Those are not agent-device regressions —
-rerun with escalation. Unit tests, typecheck, lint, and build can stay sandboxed unless they need
-devices or listener access.
+The daemon binds localhost. If the sandbox rejects the listener with `listen EPERM`, rerun with
+host access when permitted. Generic `Failed to start daemon` or cleanup errors alone do not prove a
+sandbox cause; inspect the underlying failure. Run other checks in the sandbox unless their tools
+require host access.

docs/agents/domain.md

@@ -1,14 +1,7 @@
 # Domain Docs
 
-Single-context repo. Before architecture, diagnosis, TDD, triage, PRD, or roadmap work, read
-`CONTEXT.md` for domain vocabulary and the relevant ADRs in `docs/adr/`. Selector-capture work also
-reads `docs/agents/selector-capture.md`.
-
-Use `CONTEXT.md` vocabulary in issue titles, refactor proposals, test names, and architecture notes.
-If a proposed change contradicts an ADR, say so explicitly and explain why the decision should be
-reopened.
-
-`AGENTS.md` routes to the rest of this directory by task type.
+Use `CONTEXT.md` for domain vocabulary and `docs/adr/README.md` to locate decisions relevant to the
+change. A proposed ADR contradiction requires an explicit decision update.
 
 ## Test-harness vocabulary