@junghanacs/entwurf 0.14.1 → 0.15.0

package/pi-extensions/lib/meta-session.ts

@@ -12,8 +12,8 @@
  *   1. RECORD functions + types (mint / serialize / parse / certifyActiveStore /
  *      decideUpsert), the backend-agnostic authority. Pure beyond an injected
  *      `now`; backend capability (wakeMode/deliveryLevel) comes from the packaged
- *      registry via a cached fs read (loadMetaCapabilityRegistry) — see that seam
- *      below.
+ *      registry, re-read on every load (loadMetaCapabilityRegistry holds NO cache, so a
+ *      grade moved on disk is visible to a long-lived process) — see that seam below.
  *   2. The thin FS-BOUND STORE (step 3): `upsertMetaSession` wraps the pure core
  *      (readdir → `certifyActiveStore` → `decideUpsert` → atomic write) with the real
  *      filesystem. It lives in this module (not a sibling `*-store.ts`) on purpose:
@@ -74,12 +74,14 @@ export class MetaRecordError extends Error {
 // ---------------------------------------------------------------------------
 
 /**
- * The three native meta-bridge backends, declared from the start so the
- * per-backend adapter seam is forced (a different native layout each — that
- * difference is the whole reason for a thin adapter). Discriminator field on
- * every record.
+ * Native meta-bridge backends. Discriminator on every record. Adding a member
+ * is a sibling-reader redeploy (`./run.sh install-meta-bridge` and the matching
+ * sibling install), not a store archive: writers certify the whole active store
+ * against THIS set, so a deployed plugin still carrying the old set refuses
+ * every write, including its own. `./run.sh doctor-meta-bridge` already flags a
+ * stale deployed writer — run it after the change.
  */
-export const META_BACKENDS = ["claude-code", "antigravity", "codex"] as const;
+export const META_BACKENDS = ["claude-code", "antigravity", "codex", "copilot"] as const;
 export type MetaBackend = (typeof META_BACKENDS)[number];
 
 /**
@@ -127,6 +129,47 @@ export const META_BACKEND_DESCRIPTORS: Record<MetaBackend, MetaBackendDescriptor
 		deliveryLevel: "D6",
 		nativeIdLabel: "threadId",
 	},
+	// SELF-FETCH citizen since #82 RAIL 5. Copilot CLI 1.0.80 runs our plugin hook (birth,
+	// garden address, who-sent) AND forks our first-party extension, whose `joinSession()`
+	// holds an `fs.watch` on the garden mailbox signal and can `session.send()` a doorbell
+	// into an idle session. That is the same shape Claude's mailbox has, reached through a
+	// different vendor surface. Hence:
+	//   wakeMode  self-fetch — the doorbell announces, the model drains its own inbox with
+	//             `entwurf_inbox_read`, and THAT read is the receipt. The extension never
+	//             injects the body, so this label promises exactly what happens. It was
+	//             `direct-inject` while no doorbell existed (measured 2026-08-20: no
+	//             `FileChanged`, `asyncRewake` or `watchPaths` in the bundle) — that
+	//             absence was never a claim about the vendor's other surfaces, and the
+	//             extension rail is the one it missed. Copilot is still NOT a
+	//             `nativePushSupported` backend and has no native-push adapter.
+	//   D6        the PRODUCT grade, earned by the managed LIVE acceptance of 2026-08-23
+	//             (garden 20260823T181316-d9f6ba, native 20fe30c8-b2bc-4600-91a0-8a409131be51,
+	//             CLI 1.0.80): the receive log joins that native id and arms that garden id,
+	//             the mailbox state stamps lastEnqueuedAt 09:23:41.235Z and lastReadAt
+	//             09:23:56.480Z, and the model answered on the SAME record/native/gid chain.
+	//             This is still the PRODUCT grade and not the raw probe's: the earlier
+	//             transport probe was evidence about the mechanism, and it did not move this
+	//             number — a managed round trip did.
+	//             D7 is PARTIAL, deliberately: the reply and the read receipt were observed,
+	//             but the completion taxonomy as a whole and any long-haul operation were
+	//             not, and the reply envelope itself reaches this comment as an INHERITED
+	//             fact (it was not re-read from a transcript when this was written). D8 is
+	//             unproven and D3 (managed second-session isolation) is pending — its
+	//             decisive log was lost to a scratch cleanup before it could be preserved.
+	//             The route stays fail-closed on both ends: with no armed receiver marker
+	//             every dispatch is still refused `mailbox-undeliverable`, so replyability
+	//             is a fact about a live marker and never a constant of this backend.
+	//   sessionId the native join key, measured to be ONE id across all three surfaces:
+	//             the hook envelope, `record.nativeSessionId`, and the SDK's
+	//             `session.sessionId` (record 20260823T112003-9d069a ==
+	//             `ARMED sessionId=4fc16d8d-473d-4258-a1fd-f99d3cb375e9`, CLI 1.0.80).
+	//             That agreement is what lets the extension bind its arm to the record.
+	copilot: {
+		backend: "copilot",
+		wakeMode: "self-fetch",
+		deliveryLevel: "D6",
+		nativeIdLabel: "sessionId",
+	},
 };
 
 // ---------------------------------------------------------------------------
@@ -140,9 +183,9 @@ export function requireNonEmptyString(value: unknown, field: string): string {
 	return value;
 }
 
-/** Validate the 3-backend NATIVE bridge axis (sender/receiver markers, capability
+/** Validate the native-bridge backend axis (sender/receiver markers, capability
  * drift guard). Not a record-schema validator: identity records take
- * `requireCitizenBackend` (which admits `pi`). Markers stay native-3 because the
+ * `requireCitizenBackend` (which admits `pi`). Markers stay on this set because the
  * pi adapter carries its record-established garden id into children via env rather
  * than using the native-hook pid marker rail. */
 export function requireBackend(value: unknown): MetaBackend {
@@ -210,7 +253,7 @@ function isoNow(now: Date): string {
 export const META_SCHEMA_VERSION_V3 = 3 as const;
 
 /** Every backend admitted by the one V3 record-citizen schema. */
-export const META_CITIZEN_BACKENDS = ["claude-code", "antigravity", "codex", "pi"] as const;
+export const META_CITIZEN_BACKENDS = ["claude-code", "antigravity", "codex", "copilot", "pi"] as const;
 export type MetaCitizenBackend = (typeof META_CITIZEN_BACKENDS)[number];
 
 /**
@@ -655,19 +698,30 @@ export function metaCapabilitiesFilePath(): string {
 // the rest of `delivery{}`, so today the registry is the sole home.
 // ---------------------------------------------------------------------------
 
-/** Memoized packaged registry; the file is immutable at runtime, so caching is honest (not stateful lying). */
-let cachedMetaCapabilities: MetaCapabilityRegistry | null = null;
-
 /**
- * Load + memoize the packaged capability registry — the live source of backend
- * honesty metadata as of 3D-3. A missing/corrupt file throws (the registry is a
- * packaged invariant; check-pack guarantees its presence).
+ * Load the packaged capability registry — the live source of backend honesty metadata
+ * as of 3D-3. A missing/corrupt file throws (the registry is a packaged invariant;
+ * check-pack guarantees its presence).
+ *
+ * NO CACHE, deliberately. This used to memoize into a process-lifetime singleton on the
+ * argument that "the file is immutable at runtime". That argument was false for the one
+ * process that matters: the entwurf-bridge MCP child lives as long as its harness
+ * session, so a grade or wakeMode moved by an install/upgrade was invisible to every
+ * already-running dispatcher until the operator restarted it — and nothing anywhere
+ * said so. The symptom is a citizen answering with last week's capability, which is the
+ * silent-wrong-answer class this repo refuses; "restart your session" is an instruction
+ * to work around a defect, not a fix for it (#82 RAIL 7).
+ *
+ * The subtraction is the whole repair, and it is deliberately not an invalidation
+ * scheme. Stat-based invalidation (mtime+size) cannot see an atomic same-size,
+ * same-timestamp replacement, so it would trade a certain staleness bug for an
+ * intermittent one; content hashing would have to read the file anyway. The registry is
+ * a few hundred bytes and both production callers — the v2 decider and production
+ * deliverability — reach it once per dispatch, not in a loop, so reading it there is
+ * cheaper than being wrong.
  */
 export function loadMetaCapabilityRegistry(): MetaCapabilityRegistry {
-	if (cachedMetaCapabilities === null) {
-		cachedMetaCapabilities = parseMetaCapabilityRegistry(fs.readFileSync(metaCapabilitiesFilePath(), "utf8"));
-	}
-	return cachedMetaCapabilities;
+	return parseMetaCapabilityRegistry(fs.readFileSync(metaCapabilitiesFilePath(), "utf8"));
 }
 
 /**
@@ -928,15 +982,18 @@ export function certifyActiveStoreDir(dir: string): ActiveStoreCertification & {
 
 /**
  * The refusal an uncertifiable ACTIVE store earns, in the words both the runtime
- * writers and the install doctor use. One prescription, both invocation forms —
- * every defect kind (previous generation, corruption, drift, duplicate, symlink)
- * is answered by archiving the generation, so there is nothing to branch on.
+ * writers and the install doctor use. Rotten records (previous generation,
+ * corruption, drift, duplicate, symlink) are answered by archiving. An
+ * unknown-backend defect is a stale reader, not a rotten store — the listed
+ * defect says so; this headline tells the operator to distinguish. No branch:
+ * both sentences always print.
  */
 export function activeStoreRefusal(cert: ActiveStoreCertification & { dir: string }, shown = 3): string {
 	const headline =
 		`meta-record store ${cert.dir} holds ${cert.defects.length} entry/entries this generation cannot certify — ` +
 		`refusing to write (the active store is v3-only and carries no cross-generation continuity). ` +
-		`Archive the generation and open a fresh one with ${FRESH_CUT_PRESCRIPTION}.`;
+		`Archive the generation and open a fresh one with ${FRESH_CUT_PRESCRIPTION}. ` +
+		`If a listed defect says "backend" must be one of this reader's known set, the reader is stale — redeploy the sibling unit (the matching doctor names the install verb); do not archive those records.`;
 	// `shown = 0` is for a caller that ALREADY printed every cause per entry (the
 	// store-doctor): repeating them here doubles the wall on a large previous
 	// generation, which is exactly the aggregation lesson F8 taught. The count and
@@ -1608,12 +1665,28 @@ export function readMetaSenderMarker(opts: ReadMetaSenderMarkerOptions): MetaSen
 // ── meta-receiver presence marker (SE-2 active-receiver signal) ──────────────
 
 /**
- * The arm-capable hook events. Only these can emit watchPaths (and therefore arm
- * the idle-wake), so only these write a receiver presence marker. UserPromptSubmit
- * is deliberately absent: it can backfill the record but cannot re-arm the watch, so
- * it must NOT mint or refresh an "active receiver" claim it cannot back.
+ * The arm-capable events, across every backend that can arm an idle wake.
+ *
+ * CLAUDE (`session-start`, `cwd-changed`, `file-changed`): only these hook events can
+ * emit watchPaths (and therefore arm the idle-wake), so only these write a receiver
+ * presence marker. UserPromptSubmit is deliberately absent: it can backfill the record
+ * but cannot re-arm the watch, so it must NOT mint or refresh an "active receiver"
+ * claim it cannot back.
+ *
+ * COPILOT (`extension-join`): the same rule, one backend over. Copilot's arm is not a
+ * hook at all — the CLI forks a first-party extension and the extension's own
+ * `joinSession()` is what puts a live `fs.watch` on the garden mailbox signal and gives
+ * it a `session.send()` channel to ring. The membership test is unchanged and is about
+ * capability, not event vocabulary: this provenance is admitted because the process that
+ * writes it holds the watch and can start a turn on an idle session. A Copilot hook event
+ * has neither, which is why the birth hook still arms nothing (#82 RAIL 5).
  */
-export const META_RECEIVER_ARM_PROVENANCES = ["session-start", "cwd-changed", "file-changed"] as const;
+export const META_RECEIVER_ARM_PROVENANCES = [
+	"session-start",
+	"cwd-changed",
+	"file-changed",
+	"extension-join",
+] as const;
 export type MetaReceiverArmProvenance = (typeof META_RECEIVER_ARM_PROVENANCES)[number];
 
 function requireArmProvenance(value: unknown): MetaReceiverArmProvenance {

package/pi-extensions/lib/mux-fresh-call.ts

@@ -47,6 +47,11 @@
 
 import { randomBytes } from "node:crypto";
 import { classifyTmuxCwd, type TmuxCwdRejectReason } from "./classify-tmux-cwd.ts";
+import {
+	COPILOT_PREFLIGHT_HINT,
+	type CopilotPreflightRejectReason,
+	copilotFreshPreflight,
+} from "./copilot-fresh-preflight.ts";
 import {
 	assertLaunchTarget,
 	LaunchPreconditionError,
@@ -66,27 +71,47 @@ import {
 	type WindowHandle,
 } from "./mux-placement.ts";
 
-/** The two backends this rail can open. Fixed set, not a profile — a third one is a decision,
- * not a config entry. */
-export const FRESH_CALL_BACKENDS = ["pi", "claude-code"] as const;
+/** The backends this rail can open. Fixed set, not a profile — a further one is a decision,
+ * not a config entry. `copilot` was added by #82 RAIL 9 under the step 9 admission contract. */
+export const FRESH_CALL_BACKENDS = ["pi", "claude-code", "copilot"] as const;
 export type FreshCallBackend = (typeof FRESH_CALL_BACKENDS)[number];
 
-/** The fixed runtime each backend resolves on PATH. Same reason `mux-launch` uses PATH rather
+/**
+ * The fixed runtime each backend resolves on PATH. Same reason `mux-launch` uses PATH rather
  * than a compiled-in location: the official binary is whatever the operator's environment gives
- * when they type the name. */
+ * when they type the name.
+ *
+ * `copilot` resolves `entwurf`, NOT the vendor CLI, and that is the contract rather than a
+ * convenience. Step 9 clause 1 requires ONE fixed MANAGED runtime path, and a bare `copilot`
+ * is not one: it would start without the `COPILOT_CLI_ENABLED_FEATURE_FLAGS=EXTENSIONS` token
+ * whose absence skips the extension scan SILENTLY, so the sibling would look launched and could
+ * never be delivered to. `entwurf copilot` is the accepted managed invocation that owns that
+ * flag, its recursion fence and its receiver precondition; fresh call reaches Copilot only
+ * through it. The cost is named: a Copilot fresh call needs a current `entwurf` on PATH, the
+ * way a pi fresh call needs `pi`.
+ */
 export const FRESH_CALL_RUNTIME: Record<FreshCallBackend, string> = {
 	pi: "pi",
 	"claude-code": "claude",
+	copilot: "entwurf",
 };
 
 /**
  * The callback tool NAME differs per backend and that is not cosmetic: native pi exposes the
- * capability directly (`entwurf_v2`), while a Claude Code session reaches it through the MCP
- * bridge under its namespaced name. Naming the wrong one costs the whole first turn.
+ * capability directly (`entwurf_v2`), while an MCP-hosted session reaches it under whatever
+ * name that harness composes. Naming the wrong one costs the whole first turn.
+ *
+ * `[측정]` Copilot CLI 1.0.80 composes `<mcpServerName>-<mcpToolName>` — NOT Claude Code's
+ * `mcp__<server>__<tool>`. Read from two independent sessions' own event logs
+ * (`~/.copilot/session-state/<id>/events.jsonl`): `assistant.message.toolRequests[].name` and
+ * `tool.execution_start.toolName` both carry `entwurf-bridge-entwurf_v2`, with
+ * `mcpServerName`/`mcpToolName` beside them as the parts. Derive-and-measure, never copy a
+ * sibling's spelling (`docs/adding-a-harness.md` step 5).
  */
 export const FRESH_CALL_CALLBACK_TOOL: Record<FreshCallBackend, string> = {
 	pi: "entwurf_v2",
 	"claude-code": "mcp__entwurf-bridge__entwurf_v2",
+	copilot: "entwurf-bridge-entwurf_v2",
 };
 
 /** Mirrors the `entwurf_v2` message bound. This is an INTERFACE cap for symmetry with the
@@ -112,12 +137,41 @@ export function isSafeFreshCallModel(model: string): boolean {
  *                 Flag-first submitted no message; Pi rejects the equals form for `--model`.
  *   claude-code — prompt, then `--allowedTools=` and `--model=` as ONE token each. The space form
  *                 for allowedTools is variadic and eats the prompt as an option value.
+ *   copilot     — the managed VERB first, then the prompt as the value of `-i/--interactive`,
+ *                 `--model`, value as two tokens, and the policy as the explicit `--yolo`
+ *                 token. Measured from `copilot --help` (1.0.80).
  *
- * Both failures looked identical from outside: window open, record and socket minted, no turn.
+ * Both pi/claude failures looked identical from outside: window open, record and socket minted,
+ * no turn.
  *
- * The equals form is NOT a permission guarantee — on the measured host the tool was already
- * permitted, so the option's effect was unobservable. What was observed is that it does no harm
- * to the argv. Permission stays a documented host precondition.
+ * For pi and claude-code the equals form is NOT a permission guarantee — on the measured host
+ * the tool was already permitted, so the option's effect was unobservable. What was observed is
+ * that it does no harm to the argv. Permission stays a documented host precondition there.
+ *
+ * Copilot is the backend where the policy IS carried explicitly (step 9 clause 2), and three of
+ * its argv facts are load-bearing:
+ *
+ *   - `copilot` is argv[0] of the RUNTIME `entwurf`, i.e. the managed verb — see
+ *     `FRESH_CALL_RUNTIME`. Everything after it is forwarded byte-identical by
+ *     `scripts/copilot-launch.sh`.
+ *   - the prompt rides `--interactive`, never `-p/--prompt`: `-p` runs the prompt and EXITS,
+ *     which would close the window on a sibling that is supposed to stay open and be delivered
+ *     to. `--interactive <prompt>` is non-variadic, so the space form is safe here.
+ *   - the policy token is `--yolo`, STATED HERE rather than left to the launcher: the launcher
+ *     injects `--yolo` only when the argv names no policy, and step 9 clause 2 requires the
+ *     fresh composition to state its model and permission policy explicitly, never to rely
+ *     invisibly on someone else's default.
+ *
+ * `--yolo` is a GLG operator decision, not a drifted default. The first cut passed a
+ * callback-only `--allow-tool=entwurf-bridge(entwurf_v2)` grant, and GLG's 2026-08-25 operator
+ * LIVE measured the consequence: the fresh sibling's footer showed no `YOLO`, and every tool
+ * its task needed stopped on a confirmation prompt, which made the sibling impractical to work
+ * with. GLG then set the policy explicitly: a fresh Copilot sibling carries the same managed
+ * `--yolo` profile a human-typed `entwurf copilot` gets. (Copilot 1.0.80 help: `--yolo` = all
+ * tools + all paths + all URLs.) The permission GRAMMAR lesson from that first cut — Copilot's
+ * `--allow-tool` takes `<mcp-server-name>(tool-name?)`, a different dialect from the
+ * model-facing tool name — stays recorded in `docs/adding-a-harness.md` step 9's worked
+ * example; it is a measured vendor fact even though this argv no longer uses it.
  */
 export function buildBackendArgs(backend: FreshCallBackend, prompt: string, model: string): string[] {
 	switch (backend) {
@@ -125,6 +179,8 @@ export function buildBackendArgs(backend: FreshCallBackend, prompt: string, mode
 			return [prompt, "--entwurf-control", "--model", model];
 		case "claude-code":
 			return [prompt, `--allowedTools=${FRESH_CALL_CALLBACK_TOOL["claude-code"]}`, `--model=${model}`];
+		case "copilot":
+			return ["copilot", "--interactive", prompt, "--model", model, "--yolo"];
 	}
 }
 
@@ -167,6 +223,7 @@ export type FreshCallRejectReason =
 	| PlacementRejectReason
 	| LaunchRejectReason
 	| TmuxCwdRejectReason
+	| CopilotPreflightRejectReason
 	| "caller-identity-unavailable"
 	| "model-empty"
 	| "model-invalid"
@@ -270,6 +327,17 @@ export function freshCall(
 		throw err;
 	}
 
+	// Backend capability, still PRE-MUTATION (step 9 clause 3). It runs AFTER the runtime is
+	// proven, because "entwurf is not on PATH" is the more fundamental answer — telling an
+	// operator to run `entwurf install-copilot-bridge` when they have no `entwurf` at all sends
+	// them to the wrong repair. It runs BEFORE placement for the reason this whole ordering
+	// exists: a refusal here cannot leave a window behind, while the launcher's own equivalent
+	// check (receiver only, manual `entwurf copilot`) necessarily runs after one is open.
+	if (params.backend === "copilot") {
+		const missing = copilotFreshPreflight(env);
+		if (missing) return { ok: false, reason: missing };
+	}
+
 	const inspected = inspectPlacement(env);
 	if (!inspected.ok) return { ok: false, reason: inspected.reason };
 	const placement = inspected.placement;
@@ -319,6 +387,9 @@ export function freshCall(
 /** Why each refusal happened, in the caller's terms. A reason a caller cannot act on is a reason
  * they will guess about. */
 const REJECT_HINT: Record<FreshCallRejectReason, string> = {
+	// The Copilot capability reasons keep their repair text on the leaf that decides them, so
+	// the sentence an operator reads cannot drift away from the predicate that produced it.
+	...COPILOT_PREFLIGHT_HINT,
 	"no-tmux-context": "this agent is not running inside tmux, so there is no session to open a sibling beside",
 	"anchor-malformed": "TMUX_PANE is not a native pane id",
 	"anchor-unresolved": "tmux resolved no pane for this agent's anchor",