@try-works/dsh-recursive-mode 0.4.5 → 0.4.7

package/lib/recursive_ask.tool.d.ts

@@ -125,24 +125,101 @@ export declare function createRecursiveAskTool(recursive: RecursiveRuntime): imp
 /**
  * PHASE 0 — record the answer to the run-start gate, and start the run only if it says so.
  *
- * ⚠ THIS GATE NEVER ACCEPTS A RELAYED ANSWER WHILE A HUMAN CHANNEL IS MOUNTED. That is the rule that makes
- * an approval a human act rather than an inference: when `ctx.userQuestions` is present, the question is
- * PUT TO THE PERSON and nothing else can settle it — not the caller's own `answer` argument, and not a
- * fabrication, because `ask()` resolves only with a real selection. A person's decline is likewise final
- * for that call and cannot be overridden by a model that asked for `Start run` in the same breath.
- *
- * ⚠ AND WHEN NO CHANNEL IS MOUNTED, THE RELAYED ANSWER IS THE ONLY POSSIBLE SOURCE, so it is used — that
- * is the same contract the other three gates have always had, and refusing it would leave a composition
- * without the channel unable to start any run at all. The question is surfaced first by the ASK branch
- * (the card data the host renders), and the model's `answer` is that person's selection coming back.
- *
- * ⚠ WHAT THE GATE THEREFORE DOES *NOT* CLAIM, stated rather than implied: in a composition with no
- * `userQuestions` channel, a plugin cannot verify that a person was really asked, so a model could in
- * principle relay a label nobody gave. That is a property of the relay, not of this gate — and it is the
- * reason the channel is consulted in preference whenever it exists. See the header of `run-start.ts`.
+ * ⚠ THREE OUTCOMES, AND TELLING THEM APART IS THE FIX. The first version of this function collapsed all of
+ * them into one `null`: "the channel threw", "the channel resolved with something unrecognisable", and
+ * "nobody answered" produced the same refusal, whose text asserted a cause ("so no person was asked") the
+ * plugin had already thrown away. A live session paid for that: the call failed after 22.9 s, the operator
+ * could not be told why, and the refusal's own advice prescribed the call that had just failed. So:
+ *
+ *   1. A PERSON WAS REACHED (`unusable`): the channel resolved, so somebody answered, and their answer is
+ *      not a label this gate offered — a skip, a custom value, several labels at once. That is a DECISION
+ *      the gate cannot record, and it is final: neither the caller's `answer` nor `relay=true` may replace
+ *      it. (RM5504)
+ *   2. THE CHANNEL FAILED (`unavailable`): no decision came back at all, and the refusal NAMES THE CAUSE
+ *      from the error the channel threw. Here the run can still be started, because a composition whose
+ *      channel cannot deliver the question would otherwise be unable to start any run — but only by the
+ *      caller asking for the relay in so many words (`relay=true`), which the result reports as
+ *      `source: "relayed"` rather than as a person's own selection. (RM5503)
+ *   3. NO CHANNEL IS MOUNTED: the relayed answer is the only possible source, exactly as before. (RM5502
+ *      when there is no answer either)
+ *
+ * ⚠ AND A CANCELLED OR CLOSED QUESTION IS NEVER RELAYABLE. `ASK_CANCELLED`, `ASK_ABORTED` and
+ * `ASK_TIMED_OUT` are the codes that mean the question was settled from outside this gate — the card was
+ * dismissed, the turn was cancelled, or a foreground window ended. The relay is refused for those, so a
+ * question the operator stopped cannot be turned into an approval by asking again in the same breath.
+ * Every other failure is a composition or capability failure — the question reached nobody — which is the
+ * class the relay exists for.
+ *
+ * ⚠ WHAT THE GATE STILL DOES *NOT* CLAIM: a relayed approval is a relayed approval. The plugin cannot
+ * verify that a person gave the label, and it does not pretend otherwise — the result's `source` and
+ * `channel` fields say where the decision came from, and a direct selection is preferred whenever the
+ * channel can produce one.
  */
 export declare function recordRunStartAnswer(recursive: RecursiveRuntime, root: string, runId: string, answer: string | undefined, exec: {
     agent?: unknown;
     signal?: unknown;
     callId?: unknown;
-}): Promise<Record<string, unknown>>;
+}, relay?: boolean): Promise<Record<string, unknown>>;
+/**
+ * PHASE 0 — WHAT THE BLOCKING CHANNEL ACTUALLY DID.
+ *
+ * ⚠ THIS TYPE EXISTS BECAUSE ITS ABSENCE WAS THE DEFECT. The first version returned a bare `null` from a
+ * `catch {}` for every failure and for every unrecognisable selection, so "the channel threw NO_PROVIDER",
+ * "the caller is not the live root agent", "the person skipped the question" and "the person typed a
+ * custom value" were ONE value. The refusal built from it then asserted the one cause it could not know.
+ * An outcome carries the cause, the raw message, and whether the failure is the class a relay may answer.
+ */
+export type RunStartChannelOutcome = 
+/** The person answered, and their selection is exactly one of the gate's own labels. */
+{
+    kind: 'answered';
+    answer: string;
+}
+/** The channel RESOLVED, so a person was reached — but the answer is not a label this gate offered. */
+ | {
+    kind: 'unusable';
+    detail: string;
+}
+/** The channel THREW: no decision came back, and the cause is named rather than discarded. */
+ | {
+    kind: 'unavailable';
+    cause: string;
+    detail: string;
+    relayable: boolean;
+};
+/**
+ * ⚠ THE CODES THAT MEAN THE QUESTION WAS CANCELLED OR CLOSED rather than never delivered: the person
+ * dismissed the card, their turn was cancelled, or a foreground window ended. A caller may not convert any
+ * of those into an approval by asking for the relay in the same breath. Every other failure means the
+ * question reached nobody — a composition or capability failure, which is the class the relay exists for.
+ */
+export declare const NON_RELAYABLE_CHANNEL_CODES: readonly ["ASK_CANCELLED", "ASK_ABORTED", "ASK_TIMED_OUT"];
+/**
+ * Name the failure of one `ask()` call, without inventing anything about it.
+ *
+ * The cause is the error's own `code` when it has one (the harness's `UserQuestionError` carries
+ * `NO_PROVIDER`, `CALLER_NOT_LIVE`, `DELEGATED_CALLER`, `ASK_ABORTED`, …), else its `name`, else its
+ * JavaScript type. `detail` keeps the message verbatim so a reader sees the channel's own words rather
+ * than this plugin's paraphrase — the paraphrase is exactly how the previous version came to assert a
+ * cause nobody had.
+ */
+export declare function classifyChannelFailure(err: unknown): {
+    cause: string;
+    detail: string;
+    relayable: boolean;
+};
+/**
+ * Ask the run-start question through the blocking channel and report WHAT HAPPENED.
+ *
+ * ⚠ THE CATCH IS THE POINT. It used to be `catch { return null }` — a blocking human question whose
+ * failure cause was erased at the exact moment the cause was the only thing worth knowing. Every path out
+ * of this function now says which path it was.
+ *
+ * The selection is filtered to the gate's OWN labels: a question a UI answered with a free-text custom
+ * value must not become an approval just because it arrived on the right channel.
+ */
+export declare function askRunStartDirectly(channel: NonNullable<RecursiveRuntime['userQuestionsChannel']>, exec: {
+    agent?: unknown;
+    signal?: unknown;
+    callId?: unknown;
+}): Promise<RunStartChannelOutcome>;

package/lib/recursive_closeout.tool.d.ts

@@ -4,5 +4,15 @@ import type { RecursiveRuntime } from './runtime.ts';
  * SESSION's workspace only (R1 workspace-scoping invariant). The run is resolved
  * via the session agent's cwd -> workspace registry; a runId outside the current
  * workspace is rejected.
+ *
+ * A RUN ID IS A NAME, NOT A PATH HERE TOO, and "outside the current workspace is rejected" is NOT enough
+ * on its own: `closeoutRun` joins the id onto the run layer and then writes a receipt under it, and its
+ * scoping check is `runDir.startsWith(runRoot)` — a string prefix test, not containment. A `..\` segment
+ * that lands on a SIBLING of the run layer passes that check whenever the sibling's name begins with
+ * `run`, so a path-shaped id can still receive a write. MEASURED pre-fix: `..\run-away` and `../run-away`
+ * were accepted and reached the report; the other shapes below were stopped only by the sibling not
+ * existing, which is the operator's filesystem deciding, not the tool. The rule is `run-id.ts`; this
+ * boundary is where the name enters, so this is where it is refused, with `recursive_init`'s refusal shape
+ * — `BAD_RUN_ID` (RM1107), same detail sentence.
  */
 export declare function createRecursiveCloseoutTool(recursive: RecursiveRuntime): import("@deepseek-ai/dsh-tools").ToolDefinition;

package/lib/recursive_init.tool.d.ts

@@ -13,5 +13,13 @@ import type { RecursiveRuntime } from './runtime.ts';
  * rather than "this call created something": re-initialising an APPROVED run reports `approved: true`
  * (and the run keeps its goal), which is what a caller re-scaffolding a run it already started needs to
  * see.
+ *
+ * AND A RUN ID IS A NAME, NOT A PATH. This is the boundary where the name enters, so it is the boundary
+ * that refuses a path-shaped one — loudly, and before `initRun` can mkdir anything. The check lives here
+ * (through the shared `run-id.ts` rule) rather than in `runtime.ts` because `initRun` is not the only
+ * caller and because the runtime's `join(root, '.recursive', 'run', runId)` is CORRECT for a name; what
+ * was missing was a gate on the name. Teaching the runtime to accept a path would silently relocate the
+ * run layer instead of rejecting the call. See `run-id.ts` for the rule, the evidence behind it, and the
+ * "do not fix this back" note.
  */
 export declare function createRecursiveInitTool(recursive: RecursiveRuntime): import("@deepseek-ai/dsh-tools").ToolDefinition;

package/lib/recursive_phase.tool.d.ts

@@ -5,5 +5,13 @@ import type { RecursiveRuntime } from './runtime.ts';
  * (runtime.phaseRules -> phaseRulesFor) as the once-per-phase pre-step
  * reminder, so the agent can re-ask for the rules without re-injecting them on
  * every step. Returns { error } when no active phase is found.
+ *
+ * A RUN ID IS A NAME, NOT A PATH HERE TOO, INCLUDING WHEN IT IS OMITTED. Omitted and path-shaped are
+ * different cases and must stay different: omitted means "the latest run by mtime", which `resolveRunDir`
+ * answers by DISCOVERY rather than by joining anything, and that case is untouched below. A path-shaped id,
+ * by contrast, is joined by `phaseRules` -> `resolveRunDir` and then read, and `recordInjection` WRITES
+ * `memory-injections.json` under whatever directory it resolved to — with no scoping check at all on this
+ * path. So the gate fires only on an id that was actually supplied, and `run-id.ts` owns the rule. The
+ * refusal is `recursive_init`'s, `BAD_RUN_ID` (RM1107), composed identically.
  */
 export declare function createRecursivePhaseTool(recursive: RecursiveRuntime): import("@deepseek-ai/dsh-tools").ToolDefinition;

package/lib/recursive_scratch.tool.d.ts

@@ -3,5 +3,15 @@ import type { RecursiveRuntime } from './runtime.ts';
  * `recursive_scratch` — read/write/append the run-scoped disposable scratchpad
  * (R5) under the CURRENT session workspace only (R1). Scratch is git-ignored
  * and never citable as an Input.
+ *
+ * A RUN ID IS A NAME, NOT A PATH HERE TOO. `scratchRun` joins it onto the run layer, and it then WRITES
+ * (write/append) into the directory it landed on, so a path-shaped id does not merely fail to find a run:
+ * with a `..\` segment it can find and write into a SIBLING of the run layer, because the runtime's
+ * workspace-scoping check is `runDir.startsWith(runRoot)` — a string prefix test, not containment — and a
+ * sibling directory whose name begins with `run` passes it. MEASURED pre-fix: `..\run-away` was accepted
+ * and `scratchRun` wrote `scratch.md` into `<workspace>\.recursive\run-away\scratch\`. The rule is
+ * `run-id.ts`; the gate sits here, at the boundary where the name enters, rather than in `scratchRun`, for
+ * the same reason `recursive_init`'s does. Refusal shape is `recursive_init`'s, `BAD_RUN_ID` (RM1107),
+ * composed identically: one rule, one message.
  */
 export declare function createRecursiveScratchTool(recursive: RecursiveRuntime): import("@deepseek-ai/dsh-tools").ToolDefinition;

package/lib/recursive_worktree.tool.d.ts

@@ -3,5 +3,21 @@ import type { RecursiveRuntime } from './runtime.ts';
  * `recursive_worktree` — create a linked git worktree for a run and/or
  * promote a branch up the dev/stage/main chain. Workspace-scoped: the
  * operations run under the SESSION's control-plane root only.
+ *
+ * A RUN ID IS A NAME, NOT A PATH HERE TOO, and this is the worst place to be without the rule: a `create`
+ * builds TWO things out of the id — the linked worktree directory `.worktrees/<runId>` AND the git branch
+ * `recursive/<runId>` (git accepts '/' inside a ref) — so a path-shaped id used to leave a worktree and a
+ * ref behind, not just a folder. MEASURED pre-fix, per id, against a fresh repo: `nested/child-run`
+ * returned ok:true and created BOTH `.worktrees/nested/child-run` and
+ * `refs/heads/recursive/nested/child-run`, while the shapes git itself refuses as ref syntax
+ * (`recursive//tmp/x`, `recursive/C:…`, `.hidden-run`, a trailing space) failed the worktree add and
+ * created neither. The rule is `run-id.ts` and is not restated here; the gate sits at this boundary, ahead
+ * of `createRunWorktree`, so the refusal no longer depends on git happening to dislike the ref name.
+ *
+ * The refusal is `BAD_RUN_ID` (RM1107), composed exactly as `recursive_init` composes it — same code, same
+ * detail, same sentence. One message for one rule is what keeps a caller from having to learn a second
+ * vocabulary for the same defect, and the shared remedy it carries ("call recursive_init again") is right
+ * for this tool as well: a `create` for a run that does not exist yet is exactly what `recursive_init`
+ * with `createWorktree: true` does.
  */
 export declare function createRecursiveWorktreeTool(recursive: RecursiveRuntime): import("@deepseek-ai/dsh-tools").ToolDefinition;

package/lib/run-id.d.ts

@@ -0,0 +1,62 @@
+/**
+ * A RUN ID IS A NAME, NOT A PATH.
+ *
+ * WHY THIS MODULE EXISTS. Every consumer of a run id JOINS it onto a directory
+ * that already carries the meaning "the run layer":
+ *
+ *     join(root, '.recursive', 'run', runId)   // runtime.ts, run.ts, handoff.ts, scratch.ts
+ *     join(repoRoot, '.worktrees', runId)      // worktree.ts (a linked worktree)
+ *     'recursive/' + runId                     // worktree.ts (the run's git branch)
+ *
+ * `join` is a PATH operation: absolute paths, drive specifiers and `..` segments
+ * are all legal input to it, and each one silently changes what the call means.
+ * A caller who passes `E:\tmp\rm-live-diagnostics\01-calculator-lib` is asking
+ * for a run "on another drive"; what they get is a `mkdir` of
+ *
+ *     <workspace>\.recursive\run\E:\tmp\rm-live-diagnostics\01-calculator-lib
+ *
+ * which is not drive-qualified at all — on POSIX and Windows alike the colon is
+ * just another character in a relative component. The result is a bogus nested
+ * folder INSIDE the workspace, created before anything can refuse it, surfacing
+ * far away as an ENOENT-shaped runtime failure (RM5501) with the operator's
+ * filesystem already dirty.
+ *
+ * SO THE RULE IS ENFORCED WHERE THE NAME ENTERS, and NOT by teaching the runtime
+ * to accept a path. The joins in `runtime.ts` are CORRECT for a name; what was
+ * missing was a gate on the name. Do not "fix" this back: a run on another drive
+ * or in a worktree is reached through the session's control-plane root
+ * (`recursive_worktree`, `00-worktree.md`) — the run layer is never relocated by
+ * smuggling a path into the id.
+ *
+ * The charset below is deliberately the SAME one the read path already uses
+ * (`live-route.ts` `DOC_SAFE_RE`) so a name this gate accepts is a name that
+ * route can serve.
+ */
+/**
+ * The accepted shape, as prose that can be embedded in a model-facing parameter
+ * description and in a refusal detail, so the rule is stated once.
+ */
+export declare const RUN_ID_RULE = "letters, digits, dot, underscore or dash only, no leading or trailing dot, no path separator, no drive specifier and no \"..\" segment";
+/** Two ids in the shapes the scaffold convention actually produces. */
+export declare const RUN_ID_EXAMPLES = "01-calculator-lib, fixture-run";
+/**
+ * Longest run id accepted. Directory-name components cap at 255 bytes on NTFS
+ * and ext4; a run id also becomes a git ref component (`recursive/<runId>`) and
+ * a prefix of every lock/receipt filename inside the run, so the ceiling is set
+ * well below the filesystem limit rather than at it.
+ */
+export declare const RUN_ID_MAX_LENGTH = 100;
+/**
+ * Why a run id is refused, or `null` when it is a usable NAME.
+ *
+ * The returned string is the SPECIFIC problem (which rule the id broke), with no
+ * trailing punctuation and no sentence of its own, so a caller can hand it to
+ * `toolError('BAD_RUN_ID', …)` as the detail. `RUN_ID_RULE` states the shape.
+ *
+ * The order of the checks is part of the message quality: a Windows absolute
+ * path is reported as a drive-qualified path (what the caller passed) rather
+ * than as a separator complaint (what that path is made of).
+ */
+export declare function runIdProblem(raw: string): string | null;
+/** True when `raw` is a usable run NAME. Convenience for callers that only branch. */
+export declare function isValidRunId(raw: string): boolean;

package/lib/run-spec.d.ts

@@ -0,0 +1,78 @@
+/**
+ * IS THIS RUN SPEC STILL A HOLLOW TEMPLATE? — one answer, two callers.
+ *
+ * WHY THIS MODULE EXISTS. `recursive_init` scaffolds Phase 0 as a TEMPLATE: the requirement block
+ * still reads `### \`R1\` <short title>`, the acceptance criteria are still `[observable condition 1]`,
+ * the checklists are still unchecked and both gates still read `FAIL`. The owner's defect report is that
+ * `recursive_ask gate=run-start` raised the "start this run or hold?" gate while the document was in
+ * exactly that state — *"i was never shown the spec before that so how could i approve if i havent seen
+ * it"*. Approving an unfilled template is not a decision about a spec; there is no spec yet.
+ *
+ * So the QUESTION "is this document still the template?" must have ONE answer, and two consumers need it:
+ *
+ *   1. the SERVER gate (`recursive_ask`), which must REFUSE to raise the run-start question while the
+ *      answer is yes, and
+ *   2. the CLIENT sheet (`client/spec-sheet.tsx`), which must SAY SO plainly rather than dress a hollow
+ *      document up as an approvable one — the honesty rule of `client/settings-view.ts`, applied to a
+ *      document instead of a value.
+ *
+ * ⚠ THIS MODULE IS NODE-FREE ON PURPOSE. It is reached by both halves of the bundle, and the client bundle
+ * is a BROWSER closure: a `node:fs` import anywhere in its graph breaks the page. So the file reading stays
+ * in `run-start.ts` (which owns the artifact path) and this module takes TEXT and returns a VERDICT.
+ *
+ * ⚠ AND IT PINS THE TEMPLATE, NOT A COPY OF IT. The markers below are the literal lines
+ * `init-templates.ts::requirementsContent` writes. A checker that instead carried its own copy of the whole
+ * template would silently stop matching the moment the template changed; a checker that carried a CHECKSUM
+ * would call every edited document filled and every untouched one unfilled on the strength of a byte count.
+ * Naming the placeholder markers is the check that keeps meaning what it says.
+ */
+/** What the checker decided about one artifact's text. */
+export type ArtifactVerdict = 'unfilled' | 'filled';
+/** One piece of evidence found in the document, quoted with its line number. */
+export interface ArtifactMarkerHit {
+    /** Stable id of the marker that matched. */
+    id: string;
+    /** 1-based line number in the document as it was read. */
+    line: number;
+    /** The line verbatim, trimmed of surrounding whitespace. */
+    text: string;
+}
+/** The verdict plus the evidence for it. */
+export interface ArtifactVerdictResult {
+    verdict: ArtifactVerdict;
+    /** Marker hits, in line order. Empty exactly when the verdict is `filled`. */
+    hits: ArtifactMarkerHit[];
+}
+/** The named evidence classes, so a reader can tell a placeholder from an unmet gate. */
+export declare const ARTIFACT_MARKER_IDS: {
+    readonly placeholder: "placeholder";
+    readonly uncheckedTodo: "unchecked-todo";
+    readonly failedGate: "failed-gate";
+};
+export type ArtifactMarkerId = (typeof ARTIFACT_MARKER_IDS)[keyof typeof ARTIFACT_MARKER_IDS];
+/**
+ * Which marker a single line carries, or null.
+ *
+ * Exported because the client prints the marker NAMES beside the quoted lines, and a second classifier that
+ * re-derived them would be a second answer to the same question.
+ */
+export declare function markerIdsOnLine(line: string): ArtifactMarkerId[];
+/**
+ * Classify one artifact's text.
+ *
+ * A verdict of `unfilled` means the document still carries the template's own placeholder text — the
+ * evidence travels with it, line by line, so the refusal (and the client notice) can name what is missing
+ * instead of asserting a state the reader cannot check.
+ */
+export declare function classifyArtifact(text: string): ArtifactVerdictResult;
+/** The unfilled evidence only (what a refusal names). */
+export declare function unfilledEvidence(result: ArtifactVerdictResult): ArtifactMarkerHit[];
+/** The weak, contextual markers (unchecked boxes, FAIL gates). */
+export declare function contextEvidence(result: ArtifactVerdictResult): ArtifactMarkerHit[];
+/**
+ * One line naming what is missing, for a refusal sentence.
+ *
+ * The line number and the text are both quoted: "line 12: <short title>" is a thing a reader can go and
+ * look at, while "the requirements are not filled in" is an assertion they would have to take on trust.
+ */
+export declare function describeEvidence(hits: readonly ArtifactMarkerHit[], limit?: number): string;

package/lib/run-start.d.ts

@@ -53,3 +53,30 @@ export declare function readRunStartApproval(root: string, runId: string): {
  * the same string instead of re-typing it (a re-typed reason is a caller that silently stops matching).
  */
 export declare const RUN_START_NOT_APPROVED = "run not started: phase 0 approval has not been granted";
+/**
+ * PHASE 0 — THE GATE CANNOT BE RAISED BEFORE THERE IS A SPEC TO DECIDE ABOUT.
+ *
+ * THE DEFECT. `recursive_init` scaffolds Phase 0 as a TEMPLATE, and `recursive_ask gate=run-start` raised
+ * "start this run or hold?" over it immediately — while every requirement was still `<short title>`, every
+ * acceptance criterion was still `[observable condition 1]`, and nothing put the document in front of the
+ * person at all. The owner: *"the card ui for accepting the spec appeared, but i was never shown the spec
+ * before that so how could i approve if i havent seen it"*. Approving an unfilled template is not a decision
+ * about a spec; there is no spec yet, and a card that asks the question anyway teaches a person to answer
+ * without reading.
+ *
+ * ⚠ WHAT THIS DOES *NOT* TOUCH. It does not weaken the gate's own contract, it does not add a second way to
+ * start a run, and it does not make the plugin the decider: it only refuses to ASK. A person's own answer
+ * still wins (`recordRunStartAnswer` is unchanged), a spec still creates no goal, and cancellation / abort /
+ * timeout are still unrelayable. The check runs BEFORE the question is put to anybody, so no card is shown
+ * for a document that cannot be approved meaningfully.
+ *
+ * ⚠ AND IT IS A CHECK ON THE DOCUMENT, NOT ON THE CALLER. A `runId` that does not resolve is not this
+ * refusal's business — the ask path already reports that — so the guard says `ok: true` there and lets the
+ * existing route handle it.
+ */
+export declare function runStartSpecGuard(root: string, runId: string): {
+    ok: true;
+} | {
+    ok: false;
+    reason: string;
+};

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@try-works/dsh-recursive-mode",
   "description": "recursive-mode workflow as a DeepSeek Harness bundle: RecursiveRuntime service + 13 recursive_* tools (recursive_status, recursive_init, recursive_lock, recursive_lint, recursive_closeout, recursive_scratch, recursive_worktree, recursive_phase, recursive_audit_team, recursive_review, recursive_delegate, recursive_ask, recursive_preview)",
-  "version": "0.4.5",
+  "version": "0.4.7",
   "type": "module",
   "main": "lib/index.js",
   "types": "lib/index.d.ts",

package/src/client/contract.ts

@@ -62,6 +62,16 @@ export interface ClientSlots {
 /** A registered slot's options. */
 export interface SlotOptions {
   name: string
+  /**
+   * The dispatch key of a KEYED seat (`tool.call.toolview` is one), required there and ignored elsewhere.
+   *
+   * ⚠ THIS FIELD WAS MISSING AND ITS ABSENCE WAS NOT COSMETIC: `slots.ts` registers the run-start spec sheet
+   * on `tool.call.toolview` under the tool's wire name, and the harness's own options type makes `key`
+   * mandatory for a keyed slot (`KindOptions` in `packages/client/ui-slots`, where a keyed registration with
+   * no key THROWS `keyed slot "<name>" requires options.key`). Without the field here the face did not match
+   * the API it describes, so `tsc --noEmit` rejected the registration and the sheet was unreachable code.
+   */
+  key?: string
   id?: string
   order?: number
   label?: string

package/src/client/doc-viewer.tsx

@@ -25,6 +25,27 @@ export interface DocLine {
   text: string;
   /** table: data rows (header separator row dropped); first row is the header. */
   cells?: string[][];
+  /**
+   * li: the item WAS a `- [ ]` / `- [x]` task box, and whether it was ticked.
+   *
+   * ⚠ THIS IS A FIELD ON `li`, NOT A NEW KIND, ON PURPOSE. A task box is a list item — it keeps the bullet
+   * layout, the line index and the search behaviour every existing `li` has — and the ONE thing the reader
+   * must be able to see that a plain string cannot carry is the BOX ITSELF. The scaffolded Phase 0 template
+   * ships seven unticked boxes, so "is this still the template?" is partly a question about these marks.
+   *
+   * ⚠ AND THE BOX IS NOT LEFT IN `text`. `text` is the item's CONTENT, exactly the shape the base parser
+   * already produces for a plain bullet, so `search`/`copy` and every other consumer of a line's text see the
+   * words rather than the syntax. The tick survives as this field, and the renderer draws the mark back.
+   */
+  checked?: boolean;
+  /**
+   * plain: this line is a `Coverage:` / `Approval:` GATE reading, and how it reads.
+   *
+   * Same reasoning as `checked`: a gate line is a line of text, so it stays `plain` and gains the one fact
+   * the renderer cannot re-derive — whether it currently says PASS or FAIL, which is exactly what a person
+   * must be able to see before approving the document that contains it.
+   */
+  gate?: 'pass' | 'fail';
 }
 
 /** Inline segment parsed from line text (bold / code span / link). */
@@ -34,10 +55,33 @@ export interface InlineSegment {
   href?: string;
 }
 
+/**
+ * A list item: the text AFTER its marker. The marker is consumed here exactly as the base parser consumed
+ * it — `text` is the item's CONTENT, and the renderer draws the `- ` / box back from `kind` + `checked`.
+ */
+const BULLET_RE = /^\s*[-*]\s+(.+)$/;
+
+/**
+ * A task box (`[ ]`, `[x]`, `[X]`) at the front of a list item, split into its tick and the rest of the item.
+ *
+ * The tick is the one fact a reader of the document cannot recover from the text alone once the box has been
+ * recognised, so it travels as `checked`; everything after it is the item's content, as for any other bullet.
+ */
+const TASK_BOX_RE = /^\[([ xX])\]\s*(.*)$/;
+
+/** A gate reading: `Coverage: FAIL` / `Approval: PASS`, as `run-spec.ts` reads the same lines. */
+const GATE_RE = /^\s*(?:Coverage|Approval)\s*:\s*(PASS|FAIL)\s*$/i;
+
 /**
  * Markdown -> line tokens. Base is parsePlan (blank/h1-h4/li/plain, MIT),
  * extended for fenced code blocks (one code line per block) and pipe tables
  * (one table line per block, header separator row dropped).
+ *
+ * AND EXTENDED FOR WHAT THE RUN ARTIFACTS ACTUALLY CONTAIN — the task boxes and gate readings the
+ * scaffolded `00-requirements.md` ships (`- [ ] …`, `Coverage: FAIL`, `Approval: FAIL`), because a preview
+ * that renders an unticked box and a FAIL gate as generic body text hides the two marks a person who is
+ * being asked to approve the document most needs to see. Both are additive FIELDS on the existing `li` and
+ * `plain` kinds, so no line is retyped, no character is dropped, and every existing caller keeps working.
  */
 export function parseDoc(plan: string): DocLine[] {
   const raw = String(plan == null ? '' : plan).split('\n');
@@ -83,9 +127,21 @@ export function parseDoc(plan: string): DocLine[] {
       i += 1;
       continue;
     }
-    // list item
-    const li = /^\s*[-*]\s+(.+)$/.exec(line);
-    if (li) { out.push({ kind: 'li', text: li[1] }); i += 1; continue; }
+    // list item — a task box is CLASSIFIED (its tick kept as `checked`) and its text is the item's CONTENT,
+    // which is the shape the base parser already produces for a bullet. The marker and the box are DRAWN by
+    // the renderer from `kind` + `checked`, so no line is retyped and the reader still sees `- [ ] item`.
+    const item = BULLET_RE.exec(line);
+    if (item) {
+      const box = TASK_BOX_RE.exec(item[1]);
+      out.push(box === null
+        ? { kind: 'li', text: item[1] }
+        : { kind: 'li', text: box[2], checked: box[1] !== ' ' });
+      i += 1;
+      continue;
+    }
+    // gate reading: `Coverage: FAIL` / `Approval: PASS`
+    const gate = GATE_RE.exec(line);
+    if (gate) { out.push({ kind: 'plain', text: line, gate: gate[1].toUpperCase() === 'FAIL' ? 'fail' : 'pass' }); i += 1; continue; }
     out.push({ kind: 'plain', text: line });
     i += 1;
   }
@@ -137,12 +193,41 @@ function inlineNodes(segments: InlineSegment[], baseKey: string): ReactNode[] {
   });
 }
 
+/**
+ * The mark of a task box.
+ *
+ * ⚠ THE MARK REPLACES `[ ]` / `[x]` IN PLACE, GLYPH FOR GLYPH. The parser hands over the item's content
+ * without the box, so what the reader sees is `- [ ] item` where the document says `- [x] item`: same line,
+ * same position, same length of reading — and the box is still legible as a box rather than as an assertion
+ * about the item.
+ *
+ * ⚠ AND THE TICK IS CARRIED TWICE — once as that glyph for the eye and once as text, visually hidden, for the
+ * ear — because the two bracket forms are read inconsistently by screen readers, and the whole point of the
+ * mark is that "this box is not ticked" survives every way of reading it. This span carries NO separator of
+ * its own: the item's text follows it directly, separated by the leading space on that text (see
+ * `lineElement`), so that neither side of the boundary has a trailing space to lose.
+ */
+function todoMark(line: DocLine, key: string): ReactNode {
+  const done = line.checked === true;
+  const cls = 'rec-doc-todo-check' + (done ? ' rec-doc-todo-check-on' : ' rec-doc-todo-check-off');
+  return createElement('span', { key, className: cls, title: done ? 'done' : 'not done' },
+    createElement('span', { 'aria-hidden': 'true' }, done ? '[x]' : '[ ]'),
+    createElement('span', { className: 'rec-doc-sr' }, done ? 'done:' : 'not done:'),
+  );
+}
+
 /**
  * Render one parsed line as a React element. Headings/bullets get parsePlan
  * sizing; code/table get block layout; inline markup applies to plain-ish text.
+ *
+ * ⚠ ONE RENDERER, TWO READERS. `DocViewer` and the run-start spec sheet's preview both come through here,
+ * so a mark that means "unticked box" or "gate reads FAIL" cannot mean one thing in the phase-doc viewer and
+ * another in the document a person is approving. Exported for that reason alone.
  */
-function lineElement(line: DocLine, i: number, isCurrent: boolean): ReactNode {
+export function lineElement(line: DocLine, i: number, isCurrent = false): ReactNode {
   const cls = 'rec-doc-line rec-doc-' + line.kind + (isCurrent ? ' rec-doc-line-current' : '');
+  const gateCls = line.gate === undefined ? '' : ' rec-doc-gate rec-doc-gate-' + line.gate;
+  const todoCls = line.checked === undefined ? '' : ' rec-doc-todo' + (line.checked ? ' rec-doc-todo-done' : ' rec-doc-todo-open');
   if (line.kind === 'blank') return createElement('div', { key: i, 'data-line': String(i), className: cls }, null);
   if (line.kind === 'code') return createElement('pre', { key: i, 'data-line': String(i), className: cls + ' rec-doc-pre' }, createElement('code', { className: 'rec-doc-code' }, line.text));
   if (line.kind === 'table') {
@@ -157,8 +242,43 @@ function lineElement(line: DocLine, i: number, isCurrent: boolean): ReactNode {
     );
   }
   const nodes = inlineNodes(parseInline(line.text), String(i));
-  if (line.kind === 'li') return createElement('div', { key: i, 'data-line': String(i), className: cls }, createElement('span', { className: 'rec-doc-bullet' }, '•'), createElement('span', { className: 'rec-doc-li-text' }, nodes));
-  return createElement('div', { key: i, 'data-line': String(i), className: cls }, nodes);
+  // ⚠ THE PREVIEW LINE IS THE DOCUMENT'S OWN LINE. The parser keeps list content MARKER-FREE (the base
+  // parser's shape), so the renderer puts the marker back: `- ` for a bullet, and a drawn box IN PLACE OF
+  // the `[ ]` / `[x]` for a task box. That is what lets a reader check the preview against the source and
+  // see at a glance that an unticked box is unticked.
+  //
+  // ⚠ AND THE SEPARATOR IS A NON-BREAKING SPACE WRITTEN AS AN ESCAPE, ON THE ITEM'S SIDE OF THE MARK.
+  // Measured, twice: a text node that ENDS in a space loses it (so a `'- '` bullet glyph renders as `-`,
+  // and a minifier carries that through to the shipped bundle, turning every bullet into `-item`), and a
+  // plain leading space is normalised away by the JSX transform before React ever sees it. An ESCAPED
+  // non-breaking space is neither trailing nor transformable, so it is what these separators are.
+  const SPACER = '\u00A0';
+  if (line.kind === 'li' && line.checked !== undefined) {
+    return createElement('div', { key: i, 'data-line': String(i), className: cls + todoCls },
+      createElement('span', { className: 'rec-doc-bullet' }, '-'),
+      todoMark(line, 'todo-' + String(i)),
+      createElement('span', { className: 'rec-doc-li-text' }, SPACER, nodes));
+  }
+  if (line.kind === 'li') return createElement('div', { key: i, 'data-line': String(i), className: cls }, createElement('span', { className: 'rec-doc-bullet' }, '-'), createElement('span', { className: 'rec-doc-li-text' }, SPACER, nodes));
+  return createElement('div', { key: i, 'data-line': String(i), className: cls + gateCls }, nodes);
+}
+
+/**
+ * The parsed lines as elements — the preview built from `parseDoc` + `lineElement`, with no shell of its own.
+ *
+ * ⚠ THIS IS WHAT MAKES A SECOND RENDERER UNNECESSARY. Any surface that wants to show a run artifact as a
+ * PREVIEW (the phase-doc viewer's body, the run-start spec sheet's document body) renders these nodes inside
+ * whatever frame it owns, so the markdown is parsed and drawn exactly once in the plugin. `keyBase` namespaces
+ * the React keys when several of these are on screen at once; `current` is the vim cursor line, which the
+ * spec sheet never sets.
+ */
+export function PreviewLines({ lines, keyBase = 'doc', current = -1 }: { lines: DocLine[]; keyBase?: string; current?: number }): ReactNode {
+  return createElement('div', { className: 'rec-doc-lines', 'data-preview-lines': String(lines.length) },
+    ...lines.map((line, i) => createElement(
+      'div',
+      { key: keyBase + '-line-' + String(i), className: 'rec-doc-line-wrap' },
+      lineElement(line, i, i === current),
+    )));
 }
 
 /**
@@ -269,11 +389,10 @@ export function DocViewer({ runId, worktreeRoot, fileName, theme, onClose }: Doc
     }
   };
 
-  const matchSet = new Set(matches);
-  const activeLine = matches.length > 0 ? matches[activeMatch] : -1;
-
-  const lineEls = docLines.map((line, i) => lineElement(line, i, i === cursor));
-
+  // NOTE: `n`/`N` move the cursor to the matching line, which is marked `rec-doc-line-current` by
+  // `PreviewLines` below. The lines themselves are NOT individually match-highlighted — they were not
+  // before this file gained a shared preview renderer either, and inventing a highlight here would change
+  // how the phase-doc viewer draws a document as a side effect of the run-start spec sheet's work.
   const searchBar = searchOpen ? createElement('div', { className: 'rec-doc-search' },
     createElement('input', { className: 'rec-doc-search-input', value: query, placeholder: '/ search doc…', onChange: onSearchChange, onKeyDown: onSearchKey }),
     createElement('span', { className: 'rec-doc-search-count' }, matches.length > 0 ? (activeMatch + 1) + '/' + matches.length : (query ? '0' : '')),
@@ -296,7 +415,7 @@ export function DocViewer({ runId, worktreeRoot, fileName, theme, onClose }: Doc
     createElement('div', { className: 'rec-doc-body' },
       text === null && error === null ? createElement('p', { className: 'rec-doc-text' }, 'Loading doc…') : null,
       error !== null ? createElement('p', { className: 'rec-doc-text rec-doc-error' }, error) : null,
-      text !== null ? lineEls : null,
+      text !== null ? createElement(PreviewLines, { lines: docLines, keyBase: 'doc', current: cursor }) : null,
     ),
     createElement('footer', { className: 'rec-doc-footer' },
       createElement('div', { className: statusCls, role: 'status' }, statusText),