@possumtech/rummy 0.5.0 → 2.0.0

package/src/agent/TurnExecutor.js

@@ -1,130 +1,19 @@
 import RummyContext from "../hooks/RummyContext.js";
-import ContextAssembler from "./ContextAssembler.js";
-import ResponseHealer from "./ResponseHealer.js";
-import { countTokens } from "./tokens.js";
+import { ContextExceededError } from "../llm/errors.js";
+import materializeContext from "./materializeContext.js";
 import XmlParser from "./XmlParser.js";
 
-const ACTION_SCHEMES = new Set([
-	"get",
-	"set",
-	"rm",
-	"mv",
-	"cp",
-	"sh",
-	"env",
-	"search",
-]);
-const MUTATION_SCHEMES = new Set(["set", "rm", "sh", "mv", "cp"]);
-const READ_SCHEMES = new Set(["get", "env", "search"]);
-
 export default class TurnExecutor {
 	#db;
 	#llmProvider;
 	#hooks;
-	#knownStore;
+	#entries;
 
-	constructor(db, llmProvider, hooks, knownStore) {
+	constructor(db, llmProvider, hooks, entries) {
 		this.#db = db;
 		this.#llmProvider = llmProvider;
 		this.#hooks = hooks;
-		this.#knownStore = knownStore;
-	}
-
-	/**
-	 * Rebuild turn_context from v_model_context, then assemble messages.
-	 * Called at turn start and again after any fidelity demotion within the turn.
-	 */
-	async #materializeTurnContext({
-		runId,
-		loopId,
-		turn,
-		systemPrompt,
-		mode,
-		toolSet,
-		contextSize,
-		demoted,
-	}) {
-		await this.#db.clear_turn_context.run({ run_id: runId, turn });
-		const viewRows = await this.#db.get_model_context.all({ run_id: runId });
-		for (const row of viewRows) {
-			const scheme = row.scheme || "file";
-			const projectedBody = await this.#hooks.tools.view(scheme, {
-				path: row.path,
-				scheme,
-				body: row.body,
-				attributes: row.attributes ? JSON.parse(row.attributes) : null,
-				fidelity: row.fidelity,
-				category: row.category,
-			});
-			await this.#db.insert_turn_context.run({
-				run_id: runId,
-				loop_id: loopId,
-				turn,
-				ordinal: row.ordinal,
-				path: row.path,
-				fidelity: row.fidelity,
-				status: row.status,
-				body: projectedBody ?? "",
-				// Full-body token count, not projected. This is the cost to
-				// promote the entry — the number the model needs to do Token
-				// Budget math. Projecting the demoted symbol-preview (145
-				// tokens for a 2108-token file) was misleading the model into
-				// promotes that blew the Token Budget by 10-30× per entry.
-				tokens: countTokens(row.body ?? ""),
-				attributes: row.attributes,
-				category: row.category,
-				source_turn: row.turn,
-			});
-		}
-		const rows = await this.#db.get_turn_context.all({ run_id: runId, turn });
-		const lastCtx = await this.#db.get_last_context_tokens.get({
-			run_id: runId,
-		});
-		const lastContextTokens = lastCtx?.context_tokens ?? 0;
-
-		// Baseline materialization — assemble with model's promoted spending
-		// removed (promoted data, promoted logging). The resulting size is the
-		// fixed overhead the model can't reduce without further demotion.
-		const baselineRows = rows.filter(
-			(r) =>
-				!(
-					(r.category === "data" || r.category === "logging") &&
-					r.fidelity === "promoted"
-				),
-		);
-		const baselineMessages = await ContextAssembler.assembleFromTurnContext(
-			baselineRows,
-			{
-				type: mode,
-				systemPrompt,
-				contextSize,
-				demoted,
-				toolSet,
-				lastContextTokens,
-				turn,
-			},
-			this.#hooks,
-		);
-		const baselineTokens = baselineMessages.reduce(
-			(sum, m) => sum + countTokens(m.content),
-			0,
-		);
-
-		const messages = await ContextAssembler.assembleFromTurnContext(
-			rows,
-			{
-				type: mode,
-				systemPrompt,
-				contextSize,
-				demoted,
-				toolSet,
-				lastContextTokens,
-				turn,
-				baselineTokens,
-			},
-			this.#hooks,
-		);
-		return { rows, messages, lastContextTokens };
+		this.#entries = entries;
 	}
 
 	async execute({
@@ -138,24 +27,15 @@ export default class TurnExecutor {
 		loopPrompt,
 		loopIteration,
 		noRepo,
+		noWeb,
+		noInteraction,
+		noProposals,
 		toolSet,
-		inRecovery = false,
 		contextSize,
 		options,
 		signal,
 	}) {
-		const RECOVERY_EXCLUDED = new Set([
-			"sh",
-			"env",
-			"search",
-			"ask_user",
-			"set",
-		]);
-		const effectiveToolSet = inRecovery
-			? new Set([...toolSet].filter((t) => !RECOVERY_EXCLUDED.has(t)))
-			: toolSet;
-
-		const turn = await this.#knownStore.nextTurn(currentRunId);
+		const turn = await this.#entries.nextTurn(currentRunId);
 
 		const turnRow = await this.#db.create_turn.get({
 			run_id: currentRunId,
@@ -179,7 +59,7 @@ export default class TurnExecutor {
 			{
 				hooks: this.#hooks,
 				db: this.#db,
-				store: this.#knownStore,
+				store: this.#entries,
 				project,
 				type: mode,
 				sequence: turn,
@@ -187,13 +67,16 @@ export default class TurnExecutor {
 				loopId: currentLoopId,
 				turnId: turnRow.id,
 				noRepo,
-				toolSet: effectiveToolSet,
+				noWeb,
+				noInteraction,
+				noProposals,
+				toolSet,
 				contextSize,
 				systemPrompt: null,
 				loopPrompt,
 			},
 		);
-		// Plugins write prompt/progress/instructions entries
+		// Plugins write prompt/instructions entries
 		await this.#hooks.turn.started.emit({
 			rummy,
 			mode,
@@ -205,44 +88,32 @@ export default class TurnExecutor {
 		await this.#hooks.processTurn(rummy);
 
 		// Project instructions://system through the instructions tool's projection
-		const instrEntry = await this.#knownStore.getEntriesByPattern(
-			currentRunId,
-			"instructions://system",
-			null,
-		);
-		const instrAttrs = instrEntry[0]
-			? await this.#knownStore.getAttributes(
-					currentRunId,
-					"instructions://system",
-				)
-			: null;
-		const systemPrompt = await this.#hooks.tools.view("instructions", {
-			path: "instructions://system",
-			scheme: "instructions",
-			body: instrEntry[0]?.body || "",
-			attributes: instrAttrs,
-			fidelity: "promoted",
-			category: "system",
-		});
+		const systemPrompt =
+			await this.#hooks.instructions.resolveSystemPrompt(rummy);
 
 		// Materialize turn_context: VIEW rows projected through tools
 		const demoted = [];
-		let { rows, messages, lastContextTokens } =
-			await this.#materializeTurnContext({
-				runId: currentRunId,
-				loopId: currentLoopId,
-				turn,
-				systemPrompt,
-				mode,
-				toolSet: effectiveToolSet,
-				contextSize,
-				demoted,
-			});
+		const budgetCtx = {
+			runId: currentRunId,
+			loopId: currentLoopId,
+			turn,
+			systemPrompt,
+			mode,
+			toolSet,
+			demoted,
+			loopIteration,
+		};
+		const initial = await materializeContext({
+			db: this.#db,
+			hooks: this.#hooks,
+			contextSize,
+			...budgetCtx,
+		});
 
 		await this.#hooks.context.materialized.emit({
 			runId: currentRunId,
 			turn,
-			rowCount: rows.length,
+			rowCount: initial.rows.length,
 		});
 
 		await this.#hooks.run.progress.emit({
@@ -254,73 +125,26 @@ export default class TurnExecutor {
 
 		const budgetResult = await this.#hooks.budget.enforce({
 			contextSize,
-			messages,
-			rows,
-			lastPromptTokens: lastContextTokens,
+			messages: initial.messages,
+			rows: initial.rows,
+			lastPromptTokens: initial.lastContextTokens,
+			ctx: budgetCtx,
+			rummy,
 		});
-		messages = budgetResult.messages;
-		rows = budgetResult.rows;
-		let assembledTokens =
-			budgetResult.assembledTokens ??
-			messages.reduce((sum, m) => sum + countTokens(m.content), 0);
-
-		if (budgetResult.status === 413) {
-			if (loopIteration === 1) {
-				// Prompt Demotion: first-turn overflow — demote incoming prompt to summary
-				const promptRow = rows.findLast(
-					(r) => r.category === "prompt" && r.scheme === "prompt",
-				);
-				if (promptRow) {
-					await this.#knownStore.setFidelity(
-						currentRunId,
-						promptRow.path,
-						"demoted",
-					);
-				}
-				const reMat = await this.#materializeTurnContext({
-					runId: currentRunId,
-					loopId: currentLoopId,
-					turn,
-					systemPrompt,
-					mode,
-					toolSet: effectiveToolSet,
-					contextSize,
-					demoted,
-				});
-				rows = reMat.rows;
-				messages = reMat.messages;
-				const recheck = await this.#hooks.budget.enforce({
-					contextSize,
-					messages,
-					rows,
-					lastPromptTokens: reMat.lastContextTokens,
-				});
-				messages = recheck.messages;
-				rows = recheck.rows;
-				assembledTokens =
-					recheck.assembledTokens ??
-					messages.reduce((sum, m) => sum + countTokens(m.content), 0);
-				if (recheck.status === 413) {
-					return {
-						turn,
-						turnId: turnRow.id,
-						status: 413,
-						assembledTokens,
-						contextSize,
-						overflow: recheck.overflow,
-					};
-				}
-			} else {
-				// Base context too large even without new prompt — genuine failure
-				return {
-					turn,
-					turnId: turnRow.id,
-					status: 413,
-					assembledTokens,
-					contextSize,
-					overflow: budgetResult.overflow,
-				};
-			}
+		const messages = budgetResult.messages;
+		const assembledTokens = budgetResult.assembledTokens;
+
+		if (!budgetResult.ok) {
+			return {
+				turn,
+				turnId: turnRow.id,
+				recorded: [],
+				summaryText: null,
+				updateText: null,
+				assembledTokens,
+				contextSize,
+				overflow: budgetResult.overflow,
+			};
 		}
 
 		const runRow = await this.#db.get_run_by_id.get({ id: currentRunId });
@@ -332,49 +156,37 @@ export default class TurnExecutor {
 			turn,
 		});
 
-		// Call LLM
+		// Call LLM. Transient-error retry + context-exceeded detection live
+		// in LlmProvider; context-exceeded surfaces as ContextExceededError.
 		await this.#hooks.llm.request.started.emit({ model: requestedModel, turn });
 		let rawResult;
-		const isTransient = (e) =>
-			/\b(503|429|timeout|ECONNREFUSED|ECONNRESET|unavailable)\b/i.test(
-				e.message,
-			);
-		const isContextExceeded = (e) =>
-			/\b(context.*(size|length|limit)|token.*(limit|exceed)|too.*(long|large))\b/i.test(
-				e.message,
+		try {
+			rawResult = await this.#llmProvider.completion(
+				filteredMessages,
+				requestedModel,
+				{ temperature: options?.temperature, signal },
 			);
-
-		for (let llmAttempt = 0; ; llmAttempt++) {
-			try {
-				rawResult = await this.#llmProvider.completion(
-					filteredMessages,
-					requestedModel,
-					{ temperature: options?.temperature, signal },
-				);
-				break;
-			} catch (err) {
-				if (isTransient(err) && llmAttempt < 3) {
-					const delay = 1000 * 2 ** llmAttempt;
-					console.warn(
-						`[RUMMY] Transient LLM error (attempt ${llmAttempt + 1}/3): ${err.message.slice(0, 120)}. Retrying in ${delay}ms.`,
-					);
-					await new Promise((r) => setTimeout(r, delay));
-					continue;
-				}
-				if (isContextExceeded(err)) {
-					console.warn(
-						`[RUMMY] LLM context exceeded: ${err.message.slice(0, 120)}. Returning 413.`,
-					);
-					return {
-						turn,
-						turnId: turnRow.id,
-						status: 413,
-						assembledTokens,
-						contextSize,
-					};
-				}
-				throw err;
+		} catch (err) {
+			if (err instanceof ContextExceededError) {
+				await this.#hooks.error.log.emit({
+					store: this.#entries,
+					runId: currentRunId,
+					turn,
+					loopId: currentLoopId,
+					message: `LLM context exceeded: ${err.message}`,
+					status: 413,
+				});
+				return {
+					turn,
+					turnId: turnRow.id,
+					recorded: [],
+					summaryText: null,
+					updateText: null,
+					assembledTokens,
+					contextSize,
+				};
 			}
+			throw err;
 		}
 		const result = await this.#hooks.llm.response.filter(rawResult, {
 			model: requestedModel,
@@ -387,7 +199,10 @@ export default class TurnExecutor {
 			usage: result.usage,
 		});
 		const responseMessage = result.choices?.[0]?.message;
-		const content = responseMessage?.content || "";
+		// A valid completion response always carries content (possibly
+		// empty) on the message; protect against that specific case so
+		// downstream parsers see a string.
+		const content = responseMessage?.content ? responseMessage.content : "";
 
 		await this.#hooks.run.progress.emit({
 			projectId,
@@ -397,19 +212,38 @@ export default class TurnExecutor {
 		});
 
 		// Parse and emit — plugins handle audit storage
-		const { commands, unparsed } = XmlParser.parse(content);
+		const { commands, warnings, unparsed } = XmlParser.parse(content);
+		for (const w of warnings) {
+			await this.#hooks.error.log.emit({
+				store: this.#entries,
+				runId: currentRunId,
+				turn,
+				message: w,
+				loopId: currentLoopId,
+				status: 422,
+			});
+		}
+		if (commands.length === 0 && !!unparsed?.trim() && warnings.length === 0) {
+			await this.#hooks.error.log.emit({
+				store: this.#entries,
+				runId: currentRunId,
+				turn,
+				loopId: currentLoopId,
+				message: "Response contained no actionable tags.",
+				status: 422,
+			});
+		}
 
-		// Ensure reasoning_content captures both API field and <think> tag
+		// Merge reasoning contributions from subscribers (think plugin's
+		// <think> tag, other plugin reasoning sources). Filter starts with
+		// the API-provided reasoning_content and layers on each plugin's
+		// contribution.
 		if (responseMessage) {
-			const thinkCmds = commands.filter((c) => c.name === "think");
-			const thinkText = thinkCmds
-				.map((c) => c.body)
-				.filter(Boolean)
-				.join("\n");
-			const apiReasoning = responseMessage.reasoning_content || "";
-			const parts = [apiReasoning, thinkText].filter(Boolean);
-			responseMessage.reasoning_content =
-				parts.length > 0 ? parts.join("\n") : null;
+			const seed = responseMessage.reasoning_content
+				? responseMessage.reasoning_content
+				: "";
+			const merged = await this.#hooks.llm.reasoning.filter(seed, { commands });
+			responseMessage.reasoning_content = merged ? merged : null;
 		}
 
 		const systemMsg = filteredMessages.find((m) => m.role === "system");
@@ -445,171 +279,126 @@ export default class TurnExecutor {
 		// Sequential queue. Each tool completes before the next starts.
 		// On failure: abort remaining. On proposal: notify client, await
 		// resolution, continue.
-		let hasErrors = false;
+		// Narration text outside tags is fine when the turn also emitted
+		// at least one command — "OK", "Let me check:", reasoning prefixes
+		// are natural. Parse warnings and no-tags responses already emitted
+		// errors above; dispatch crashes and failed entries emit below.
 		let abortAfter = null;
 
 		for (const entry of recorded) {
+			if (entry.state === "failed" || entry.state === "cancelled") continue;
+
 			if (abortAfter) {
 				const errorMsg = `Aborted — preceding <${abortAfter}> failed.`;
-				await this.#knownStore.upsert(
-					currentRunId,
+				await this.#entries.set({
+					runId: currentRunId,
 					turn,
-					entry.resultPath || entry.path,
-					errorMsg,
-					409,
-					{ attributes: { error: errorMsg }, loopId: currentLoopId },
-				);
-				hasErrors = true;
+					path: entry.resultPath || entry.path,
+					body: errorMsg,
+					state: "failed",
+					outcome: "aborted",
+					attributes: { error: errorMsg },
+					loopId: currentLoopId,
+				});
 				continue;
 			}
 
 			await this.#hooks.tool.before.emit({ entry, rummy });
-			await this.#hooks.tools.dispatch(entry.scheme, entry, rummy);
+			try {
+				await this.#hooks.tools.dispatch(entry.scheme, entry, rummy);
+			} catch (dispatchErr) {
+				await this.#hooks.error.log.emit({
+					store: this.#entries,
+					runId: currentRunId,
+					turn,
+					loopId: currentLoopId,
+					message: `Dispatch crash in ${entry.scheme}: ${dispatchErr.message}`,
+					status: 500,
+				});
+				abortAfter = entry.scheme;
+				continue;
+			}
 			await this.#hooks.tool.after.emit({ entry, rummy });
 			await this.#hooks.entry.created.emit(entry);
 
-			// Materialize proposals for this entry (set revisions → 202)
-			await this.#hooks.turn.proposing.emit({ rummy, recorded: [entry] });
+			// Plugins (e.g. set) materialize pending proposals from the
+			// recorded entry — e.g. search/replace revisions → set:// 202.
+			await this.#hooks.proposal.prepare.emit({ rummy, recorded: [entry] });
 
 			// Check for any proposals created by this entry's dispatch
-			const proposed = await this.#knownStore.getUnresolved(currentRunId);
+			const proposed = await this.#entries.getUnresolved(currentRunId);
 			for (const p of proposed) {
-				await this.#hooks.turn.proposal.emit({
+				await this.#hooks.proposal.pending.emit({
 					projectId,
 					run: currentAlias,
 					proposed: [p],
 				});
-				await this.#knownStore.waitForResolution(currentRunId, p.path);
-				const resolved = await this.#db.get_entry_state.get({
-					run_id: currentRunId,
-					path: p.path,
-				});
+				await this.#entries.waitForResolution(currentRunId, p.path);
+				const resolved = await this.#entries.getState(currentRunId, p.path);
 				if (resolved?.status >= 400) {
-					hasErrors = true;
+					await this.#hooks.error.log.emit({
+						store: this.#entries,
+						runId: currentRunId,
+						turn,
+						loopId: currentLoopId,
+						message: `Proposal ${p.path} rejected: status ${resolved.status}.`,
+						status: resolved.status,
+					});
 					abortAfter = entry.scheme;
 				}
 			}
 
-			// Also check the entry itself for direct failures
-			if (!hasErrors) {
+			if (!abortAfter) {
 				const entryPath = entry.resultPath || entry.path;
-				const row = await this.#db.get_entry_state.get({
-					run_id: currentRunId,
-					path: entryPath,
-				});
+				const row = await this.#entries.getState(currentRunId, entryPath);
 				if (row?.status >= 400) {
-					hasErrors = true;
+					await this.#hooks.error.log.emit({
+						store: this.#entries,
+						runId: currentRunId,
+						turn,
+						loopId: currentLoopId,
+						message: `Entry ${entryPath} failed: status ${row.status}.`,
+						status: row.status,
+					});
 					abortAfter = entry.scheme;
 				}
 			}
 		}
 
-		// Turn Demotion: if end-of-turn context exceeds ceiling, demote this
-		// turn's data entries and the incoming prompt to summary, then force a
-		// budget recovery phase before continuing.
-		let budgetRecovery = null;
-		// Use actual prompt_tokens from this turn's LLM response as the ground-truth
-		// Post-dispatch budget check — demotion handled by budget plugin
-		if (contextSize) {
-			const postMat = await this.#materializeTurnContext({
-				runId: currentRunId,
-				loopId: currentLoopId,
-				turn,
-				systemPrompt,
-				mode,
-				toolSet: effectiveToolSet,
-				contextSize,
-				demoted,
-			});
-			budgetRecovery = await this.#hooks.budget.postDispatch({
-				contextSize,
-				messages: postMat.messages,
-				rows: postMat.rows,
-				runId: currentRunId,
-				loopId: currentLoopId,
-				turn,
-				db: this.#db,
-				store: this.#knownStore,
-			});
-		}
-
-		const summaryEntry = recorded.findLast((e) => e.scheme === "summarize");
-		const updateEntry = recorded.findLast((e) => e.scheme === "update");
-		let summaryText = summaryEntry?.body || null;
-		let updateText = updateEntry?.body || null;
-
-		// If model sent both, last signal wins — respects the model's final intent
-		if (summaryText && updateText) {
-			const lastLifecycle = recorded.findLast(
-				(e) => e.scheme === "summarize" || e.scheme === "update",
-			);
-			if (lastLifecycle.scheme === "summarize") updateText = null;
-			else summaryText = null;
-		}
-
-		// If model says "done" but actions failed, override — the model's
-		// assertion that it's done is false if it failed to do what it tried.
-		if (summaryText && hasErrors) {
-			console.warn(
-				"[RUMMY] Overriding <summarize> — actions in this turn failed. Continuing.",
-			);
-			// Mark the recorded summarize entry as 409 so the model sees it was rejected
-			if (summaryEntry?.path) {
-				await this.#knownStore.resolve(
-					currentRunId,
-					summaryEntry.path,
-					409,
-					"Overridden — actions in this turn failed. Use <update/> until resolved.",
-				);
-			}
-			updateText = summaryText;
-			summaryText = null;
-		}
-
-		// If model sent neither, heal from content
-		let statusHealed = false;
-		if (!summaryText && !updateText) {
-			const healed = ResponseHealer.healStatus(content, commands);
-			summaryText = healed.summaryText;
-			updateText = healed.updateText;
-			statusHealed = true;
-		}
-
-		// --- Classify for return value ---
-
-		const actionCalls = recorded.filter((e) => ACTION_SCHEMES.has(e.scheme));
-		const writeCalls = recorded.filter(
-			(e) =>
-				e.scheme === "known" ||
-				(e.scheme === "set" && !e.attributes?.blocks && !e.attributes?.search),
-		);
-		const unknownCalls = recorded.filter((e) => e.scheme === "unknown");
+		// Turn Demotion: budget plugin re-materializes end-of-turn context
+		// and demotes this turn's promoted entries on overflow. Overflow
+		// emits an error (status 413) via the unified error channel.
+		await this.#hooks.budget.postDispatch({
+			contextSize,
+			ctx: budgetCtx,
+			rummy,
+		});
 
-		const hasAct = actionCalls.some((c) => MUTATION_SCHEMES.has(c.scheme));
-		const hasReads = actionCalls.some((c) => READ_SCHEMES.has(c.scheme));
-		const hasWrites = writeCalls.length > 0 || unknownCalls.length > 0;
-		const flags = { hasAct, hasReads, hasWrites };
+		const { summaryText, updateText } = await this.#hooks.update.resolve({
+			recorded,
+			content,
+			commands,
+			runId: currentRunId,
+			turn,
+			loopId: currentLoopId,
+			rummy,
+		});
 
 		const askUserEntry = recorded.find((e) => e.scheme === "ask_user");
 
 		const turnResult = {
 			turn,
 			turnId: turnRow.id,
-			actionCalls,
-			writeCalls,
-			unknownCalls,
+			recorded,
 			summaryText,
 			updateText,
-			statusHealed,
-			askUserCmd: askUserEntry || null,
-			flags,
-			model: result.model || requestedModel,
+			askUserCmd: askUserEntry,
+			model: result.model ? result.model : requestedModel,
 			modelAlias: requestedModel,
 			temperature: options?.temperature,
 			contextSize,
 			assembledTokens,
 			usage: result.usage,
-			budgetRecovery,
 		};
 
 		await this.#hooks.turn.completed.emit(turnResult);
@@ -623,59 +412,78 @@ export default class TurnExecutor {
 	 */
 	async #record(runId, loopId, turn, mode, cmd) {
 		const scheme = cmd.name;
-		const rawTarget = cmd.path || cmd.command || cmd.question || "";
+		// Each tool's XmlParser shape surfaces exactly one of these
+		// three fields as its addressable target. Treat absent as empty
+		// so the length/control-char validation below catches bad shapes
+		// rather than letting an undefined slip through.
+		let rawTarget = "";
+		if (cmd.path) rawTarget = cmd.path;
+		else if (cmd.command) rawTarget = cmd.command;
+		else if (cmd.question) rawTarget = cmd.question;
 		// Reject paths that are likely reasoning bleed — too long or contain non-printing chars
 		if (rawTarget.length > 512 || /\p{Cc}/u.test(rawTarget)) {
-			const rejectPath = await this.#knownStore.dedup(
+			const rejectPath = await this.#entries.logPath(
 				runId,
-				scheme,
-				`${scheme}://invalid`,
 				turn,
+				scheme,
+				"invalid",
 			);
-			await this.#knownStore.upsert(
+			await this.#entries.set({
 				runId,
 				turn,
-				rejectPath,
-				`Invalid path: too long or contains non-printing characters`,
-				400,
-				{ loopId },
-			);
+				path: rejectPath,
+				body: `Invalid path: too long or contains non-printing characters`,
+				state: "failed",
+				outcome: "validation",
+				attributes: { action: scheme },
+				loopId,
+			});
 			return {
 				scheme,
 				path: rejectPath,
 				body: "",
 				attributes: {},
-				status: 400,
+				state: "failed",
+				outcome: "validation",
 				resultPath: rejectPath,
 			};
 		}
 		const target = rawTarget;
-		const resultPath = await this.#knownStore.dedup(
-			runId,
-			scheme,
-			target,
-			turn,
-		);
+		const resultPath = await this.#entries.logPath(runId, turn, scheme, target);
 
 		// Pass parsed command fields through as attributes
 		const { name: _, ...attributes } = cmd;
 		if (cmd.path) attributes.path = target;
 
-		const body = cmd.body || cmd.command || cmd.question || "";
+		// Same per-shape resolution as rawTarget; the three sources are
+		// mutually exclusive per tool. Empty string when none set.
+		let body = "";
+		if (cmd.body) body = cmd.body;
+		else if (cmd.command) body = cmd.command;
+		else if (cmd.question) body = cmd.question;
 
 		// Filter: plugins can validate/transform before recording
 		const filtered = await this.#hooks.entry.recording.filter(
-			{ scheme, path: resultPath, body, attributes, status: 200 },
-			{ runId, turn, loopId, mode },
+			{
+				scheme,
+				path: resultPath,
+				body,
+				attributes,
+				state: "resolved",
+				outcome: null,
+			},
+			{ store: this.#entries, runId, turn, loopId, mode },
 		);
-		if (filtered.status >= 400) return filtered;
+		if (filtered.state === "failed" || filtered.state === "cancelled") {
+			return filtered;
+		}
 
 		return {
 			scheme: filtered.scheme,
 			path: filtered.path,
 			body: filtered.body,
 			attributes: filtered.attributes,
-			status: 200,
+			state: "resolved",
 			resultPath: filtered.path,
 		};
 	}