@possumtech/rummy 2.0.0 → 2.1.0

package/src/plugins/instructions/instructions_105.md

@@ -1,46 +1,39 @@
-# Discovery Stage
+# Distillation Stage: YOU MUST select an unknown:// entry, then discover its source entries and distill them into known:// entries
 
-YOU MUST ONLY perform discovery actions (Discover -> Distill -> Demote) during the Discovery Stage.
-YOU MUST discover and get source entries with information relevant to the unknown:// entries.
-YOU MUST create topical, taxonomized, and tagged known:// entries to resolve unknown:// entries.
-YOU MUST include at least 1 link to a relevant unknown and at least 1 link to a relevant source entry in every known:// entry.
-YOU MUST demote source entries to "summarized" after extracting and decomposing their relevant information into known:// entries.
-YOU MUST demote the unknown:// entries to "summarized" after they are referenced or resolved by known:// entries.
-YOU MUST demote all irrelevant source entries and log events to maximize FCRM.
-Tip: Source entry "summarized" information is not reliable. Only place "visible" source entry information in known:// entries.
-Tip: A "relevant" source entry that has been successfully distilled into known:// entries is no longer relevant.
-Tip: Discover, Distill, and Demote per source entry, not globally, to maximize FCRM.
+YOU MUST create topical, taxonomized, and tagged known:// entries to resolve the selected unknown:// entry.
+YOU MUST reference all related source entries and prompts in the `# Related` list
+YOU MUST ONLY populate known entries with promoted information, NOT from your own training data or opinion.
+YOU MUST immediately demote unknowns, source entries, prompts, and log events after they are distilled, irrelevant, or resolved.
 
-## Discovery Lifecycle: Promoting a source entry, creating a known entry, demoting the source entry, then archiving the resolved unknown
+* Check the `tokens="N"` of the source entries against the `tokensFree="N"` constraint before promoting entries.
+* You can use <get path="..." manifest/> to list paths and their token amounts for bulk operations without performing them.
+* You can use <get path="..." line="X" limit="Y"/> to read subsets of entries that would exceed your `tokensFree` budget.
+* Don't accidentally set the current prompt to `archived`.
 
-### Discover
+Example:
+	<get path="**" manifest>capital</get>
+	<get path="prompt://3" line="1" limit="100"/>
 
-<set path="trivia/capitals.csv" visibility="visible"/>
+	<set path="trivia/capitals.csv" visibility="visible"/>
 
-### Distill
-<set path="known://countries/france/capital" summary="countries,france,capital,geography,trivia">
-    # Capital of France
-    The capital of France is Paris.
+	<set path="known://countries/france/capital" summary="countries,france,capital,geography,trivia">
+		# Related
+		[trivia question](prompt://3)
+		[unknown resolving](unknown://countries/france/capital)
+		[source entry](trivia/capitals.csv)
 
-    { ... }
+		# Capital of France
+		The capital of France is Paris.
 
-    # References
-    [unknown resolving](unknown://countries/france/capital)
-    [source entry](trivia/capitals.csv)
-</set>
+		{...}
+	</set>
 
-### Demote
-
-<set path="trivia/capitals.csv" visibility="summarized"/>
-<set path="unknown://countries/france/capital" visibility="summarized"/>
-<set path="unknown://countries/poland/capital" summary="REJECTED: Irrelevant" visibility="summarized"/>
-<set path="https://en.wikipedia.org/wiki/Paris,_Texas" summary="REJECTED: Wrong Paris" visibility="summarized"/>
-<set path="log://turn_1/set/*" visibility="archived"/>
-<set path="log://turn_1/get/trivia/*" visibility="archived"/>
-<set path="log://turn_2/get/capital%20of%20france" visibility="archived"/>
+	<set path="prompt://3" visibility="summarized"/>
+	<set path="unknown://countries/france/capital" visibility="summarized"/>
+	<set path="unknown://countries/france/seat_of_government" summary="RESOLVED: Not necessary" visibility="summarized"/>
+	<set path="trivia/capitals.csv" visibility="summarized"/>
 
 ## Turn Termination (CHOOSE ONLY ONE):
-
-Definition Stage Return: <update status="154">returning to definition stage</update>
-Discovery Stage Continuation: <update status="155">referencing and resolving more unknowns</update>
-Discovery Stage Completion: <update status="158">all unknowns (if any) referenced or resolved by known entries</update>
+* Decomposition Stage Return: <update status="154">additional unknowns identified; returning to Decomposition Stage</update>
+* Distillation Stage Continuation: <update status="155">discovering and distilling more for the selected unknown</update>
+* Distillation Stage Completion: <update status="156">this unknown's known entries written</update>

package/src/plugins/instructions/instructions_106.md

@@ -0,0 +1,22 @@
+# Demotion Stage: YOU MUST demote all source entries, prompts, and log events that are now distilled or no longer relevant
+
+Example:
+	<set path="prompt://2" summary="All information distilled into knowns" visibility="summarized"/>
+	<set path="trivia/capitals.csv" visibility="summarized"/>
+	<set path="unknown://countries/france/capital" visibility="summarized"/>
+	<set path="unknown://countries/poland/capital" summary="REJECTED: Irrelevant" visibility="summarized"/>
+	<set path="https://en.wikipedia.org/wiki/Paris,_Texas" summary="REJECTED: Wrong Paris" visibility="summarized"/>
+	<set path="log://turn_1/**" visibility="archived"/>
+	<set path="log://turn_2/**" visibility="archived"/>
+	<set path="log://turn_3/set/**" visibility="archived"/>
+	<set path="log://turn_3/get/**" visibility="archived"/>
+	<set path="log://turn_3/search/**" visibility="archived"/>
+
+* You need room to think. Demote large prompts and source entries, then iterate them with <get path="..." line="N" limit="N"/> as necessary.
+* When demoting prompts, prefer "summarized" to "archived" to avoid losing necessary context.
+
+## Turn Termination (CHOOSE ONLY ONE):
+* Decomposition Stage Return: <update status="164">additional unknowns identified; returning to Decomposition Stage</update>
+* Distillation Stage Return: <update status="165">more unknowns remain; returning to Distillation Stage</update>
+* Demotion Stage Continuation: <update status="166">demoting more distilled or irrelevant entries, prompts, and log events</update>
+* Demotion Stage Completion: <update status="167">all unknowns resolved and demoted; ready for Deployment Stage</update>

package/src/plugins/instructions/instructions_107.md

@@ -0,0 +1,17 @@
+# Deployment Stage: YOU MUST act on the prompt.
+
+YOU MUST attempt to deterministically verify your actions, outputs, or answers before declaring completion, if possible.
+
+Example: verifying deliverable before completion
+	<set path="sum.js">console.log(process.argv.slice(2).reduce((a, b) => a + Number(b), 0));</set>
+	<sh>[ -f sum.js ] && node --version && node sum.js 2 2 | grep -qx 4</sh>
+	<update status="177">sum.js written, node available, ran cleanly, correct output?</update>
+
+Example: <update status="200">Paris</update>
+
+## Turn Termination (CHOOSE ONLY ONE):
+* Decomposition Stage Return: <update status="174">additional unknowns identified; returning to Decomposition Stage</update>
+* Distillation Stage Return: <update status="175">selected unknown not yet resolved; returning to Distillation Stage</update>
+* Demotion Stage Return: <update status="176">context not yet sufficiently demoted; returning to Demotion Stage</update>
+* Deployment Stage Continuation: <update status="177">performing more actions</update>
+* Deployment Stage Completion: <update status="200">{direct answer (summary of actions performed if prompt not a question)}</update>

package/src/plugins/instructions/instructions_108.md

@@ -1,8 +0,0 @@
-YOU MUST act on the prompt.
-
-Turn Termination (CHOOSE ONLY ONE):
-
-Definition Stage Return: <update status="184">returning to definition stage</update>
-Discovery Stage Return: <update status="185">returning to discovery stage</update>
-Deployment Stage Continuation: <update status="188">performing more actions</update>
-Deployment Stage Completion: <update status="200">summary of actions performed, or direct answer</update>

package/src/plugins/known/README.md

@@ -1,13 +1,17 @@
 # known {#known_plugin}
 
-Writes knowledge entries into the store at full visibility.
+Writes knowledge entries into the store at full visibility, and renders
+the project's data surface as the bifurcated `<summarized>` /
+`<visible>` blocks at the top of the user message.
 
 ## Registration
 
 - **Tool**: `known`
 - **Category**: `data`
 - **Handler**: Upserts the entry body at the target path with status 200.
-- **Filter**: `assembly.system` at priority 100 — renders `<knowns>` section.
+- **Filters**:
+  - `assembly.user` priority 50 — renders `<summarized>`.
+  - `assembly.user` priority 75 — renders `<visible>`.
 
 ## Projection
 
@@ -15,7 +19,23 @@ Shows `# known {path}` followed by the entry body.
 
 ## Assembly
 
-Filters turn_context rows where `category === "data"`. Renders all
-data entries (files, knowledge, skills, URLs) into the `<knowns>` section
-of the system message. Third-party plugins that register with
-`category: "data"` automatically appear here.
+Filters `ctx.rows` where `category === "data"`. Two separate blocks
+emit at the top of the user message in this order:
+
+- `<summarized>` — each data entry whose visibility is `visible` or
+  `summarized`, rendered under its scheme tag with the plugin's
+  summary projection as body (truncated knowns, code symbols,
+  page abstracts — whatever the plugin's `summary()` hook produces).
+  Plus the named carve-out: archived prompts pass through
+  (visibility="archived") so the model can `<get>` the active prompt
+  back after demotion.
+- `<visible>` — each data entry whose visibility is `visible`,
+  rendered with the plugin's visible projection (full body) as the
+  tag body. A visible entry appears in *both* blocks: summary
+  projection up top, full body below.
+
+This split lets `<summarized>` stay cache-stable across promote/demote
+operations — only `<visible>` mutates when the model promotes a
+summary or demotes a visible entry. Third-party plugins that register
+with `category: "data"` automatically appear in both blocks under
+their scheme tag.

package/src/plugins/known/known.js

@@ -12,10 +12,9 @@ export default class Known {
 		core.on("handler", this.handler.bind(this));
 		core.on("visible", this.full.bind(this));
 		core.on("summarized", this.summary.bind(this));
-		core.filter("assembly.system", this.assembleContext.bind(this), 100);
-		// <known> is internal — written via <set path="known://...">. Hidden
-		// from all model-facing tool lists. Handler still dispatches if the
-		// model emits <known> directly out of habit.
+		core.filter("assembly.user", this.assembleSummarized.bind(this), 50);
+		core.filter("assembly.user", this.assembleVisible.bind(this), 75);
+		// Hidden tool: written via <set path="known://...">; handler tolerates direct <known>.
 		core.markHidden();
 	}
 
@@ -23,23 +22,20 @@ export default class Known {
 		const { entries: store, sequence: turn, runId, loopId } = rummy;
 		if (!entry.body) return;
 
-		// Size gate
 		const entryTokens = countTokens(entry.body);
 		if (entryTokens > MAX_ENTRY_TOKENS) {
-			const rejectPath = await store.slugPath(runId, "known", entry.body);
 			await store.set({
 				runId,
 				turn,
-				path: rejectPath,
+				loopId,
+				path: entry.resultPath,
 				body: `Entry too large (${entryTokens} tokens, max ${MAX_ENTRY_TOKENS}). Sort the information, ideas, or plans carefully into multiple entries.`,
 				state: "failed",
 				outcome: `overflow:${entryTokens}`,
-				loopId,
 			});
 			return;
 		}
 
-		// Resolve path: explicit or auto-generated slug
 		let knownPath = entry.attributes?.path;
 		if (knownPath && !knownPath.includes("://")) {
 			knownPath = `known://${knownPath}`;
@@ -53,9 +49,7 @@ export default class Known {
 			);
 		}
 
-		// Dedup: if path exists, update rather than duplicate. An empty
-		// new body means "preserve the existing entry's body" (e.g. the
-		// model is updating attributes only).
+		// Dedup: existing path → update; empty body preserves existing body.
 		const existing = await store.getEntriesByPattern(runId, knownPath, null);
 		if (existing.length > 0) {
 			const nextBody = entry.body === "" ? existing[0].body : entry.body;
@@ -86,31 +80,45 @@ export default class Known {
 		return entry.body;
 	}
 
-	// Summarized knowns keep the first 500 characters so the model
-	// doesn't lose the plot when budget auto-demotion kicks in on its
-	// own work. Anything larger gets capped so a pathologically big
-	// known doesn't saturate the packet at summarized visibility
-	// either. Matches the pattern on `<prompt>` summarized view.
+	// Summarized: first 500 chars; matches <prompt> summarized.
 	summary(entry) {
 		if (!entry.body) return "";
 		if (entry.body.length <= 500) return entry.body;
 		return `${entry.body.slice(0, 500)}\n[truncated — promote to see the full body]`;
 	}
 
-	async assembleContext(content, ctx) {
-		const entries = ctx.rows.filter((r) => r.category === "data");
+	// Identity-keyed summary lines: every data entry the run is tracking
+	// at visibility=visible or visibility=summarized.
+	async assembleSummarized(content, ctx) {
+		const entries = ctx.rows.filter(
+			(r) =>
+				r.category === "data" &&
+				(r.visibility === "visible" || r.visibility === "summarized"),
+		);
+		if (entries.length === 0) return content;
+		const lines = entries.map((e) =>
+			renderContextTag(e, e.sBody != null ? e.sBody : e.body),
+		);
+		return `${content}<summarized>\n${lines.join("\n")}\n</summarized>\n`;
+	}
+
+	async assembleVisible(content, ctx) {
+		const entries = ctx.rows.filter(
+			(r) => r.category === "data" && r.visibility === "visible",
+		);
 		if (entries.length === 0) return content;
-		const demotedSet = new Set(ctx.demoted);
-		const lines = entries.map((e) => renderContextTag(e, demotedSet));
-		return `${content}\n\n<context>\n${lines.join("\n")}\n</context>`;
+		const lines = entries.map((e) =>
+			renderContextTag(e, e.vBody != null ? e.vBody : e.body),
+		);
+		return `${content}<visible>\n${lines.join("\n")}\n</visible>\n`;
 	}
 }
 
-function renderContextTag(entry, demotedSet) {
-	// schemeOf() returns NULL / "" for bare file paths; translate for the tag.
+function renderContextTag(entry, projectedBody) {
 	const tag = entry.scheme ? entry.scheme : "file";
 	const turn = entry.source_turn ? ` turn="${entry.source_turn}"` : "";
 	const tokens = entry.aTokens != null ? ` tokens="${entry.aTokens}"` : "";
+	const lines = entry.vLines != null ? ` lines="${entry.vLines}"` : "";
 	const attrs =
 		typeof entry.attributes === "string"
 			? JSON.parse(entry.attributes)
@@ -128,21 +136,16 @@ function renderContextTag(entry, demotedSet) {
 	const stateAttr =
 		entry.state && entry.state !== "resolved" ? ` state="${entry.state}"` : "";
 	const outcomeAttr = entry.outcome ? ` outcome="${entry.outcome}"` : "";
-	const visibility = entry.visibility
-		? ` visibility="${entry.visibility}"`
-		: "";
-	const flag = demotedSet?.has(entry.path) ? " demoted" : "";
-	// Always render summary attribute on knowns — empty value hints the model
-	// it forgot to add searchable keywords.
+	const visibility =
+		entry.visibility === "archived" ? ` visibility="archived"` : "";
 	const summaryText =
 		typeof attrs?.summary === "string"
 			? attrs.summary.replace(/"/g, "'").slice(0, 80)
 			: "";
 	const summary = ` summary="${summaryText}"`;
-
-	const attrStr = `${turn}${status}${stateAttr}${outcomeAttr}${summary}${visibility}${tokens}${flag}`;
-	if (entry.body) {
-		return `<${tag} path="${entry.path}"${attrStr}>${entry.body}</${tag}>`;
+	const attrStr = `${turn}${status}${stateAttr}${outcomeAttr}${summary}${visibility}${tokens}${lines}`;
+	if (projectedBody) {
+		return `<${tag} path="${entry.path}"${attrStr}>${projectedBody}</${tag}>`;
 	}
 	return `<${tag} path="${entry.path}"${attrStr}/>`;
 }

package/src/plugins/log/README.md

@@ -29,7 +29,7 @@ size. Resolution:
   own body tokens.
 - `sh` and `env` own multiple streaming channels (`sh://turn_N/{slug}_N`)
   — no single target to point at. `tokens=` is omitted; the channels
-  render their own tokens in `<context>`.
+  render their own tokens in `<visible>`.
 
 ## Behavior
 
@@ -43,6 +43,6 @@ Log entries (`log://turn_N/{action}/{slug}`) are audit records —
 summary, exit status, references to where the data lives — and never
 carry the payload itself. Payload for streaming actions lives under the
 producer's own scheme (`sh://`, `env://`, future `search://`, etc.) at
-`category=data`, and is rendered inside `<context>` by the known
+`category=data`, and is rendered inside `<visible>` by the known
 plugin. Scheme determines category; data and logging never share a
 scheme. See [scheme_category_split](#scheme_category_split).

package/src/plugins/log/log.js

@@ -1,11 +1,6 @@
 import { stateToStatus } from "../../agent/httpStatus.js";
 
-// Schemes whose log body is an action summary, not the cost-bearing
-// content. For these, the action's cost lives on a separate data entry
-// (sh/env: streaming channels; set/mv/cp: the target entry). Report
-// tokens from the target when we can resolve it (set/mv/cp via
-// attrs.path); omit entirely for sh/env (multiple channels, no single
-// target to point at).
+// sh/env span multiple channels; channels render their own tokens in <visible>.
 const STREAM_NO_TOKENS = new Set(["sh", "env"]);
 
 export default class Log {
@@ -17,10 +12,7 @@ export default class Log {
 	}
 
 	async assembleLog(content, ctx) {
-		// Log includes action entries (scheme=log) AND prior prompts. The
-		// most recent prompt is rendered separately by the prompt plugin
-		// as `<prompt>`; everything older lives in the log so the model
-		// can see the full question history across a sustained run.
+		// Includes prior prompts; the latest prompt is rendered separately as <prompt>.
 		const latestPrompt = ctx.rows.findLast(
 			(r) => r.category === "prompt" && r.scheme === "prompt",
 		);
@@ -39,10 +31,7 @@ export default class Log {
 	}
 }
 
-// Log paths are log://turn_N/action/slug. The second segment is the
-// action — the plugin/tool that produced this log entry (set, get,
-// search, update, error, etc.). Used as the XML tag name. Prompt
-// entries live at prompt://N; they render as <prompt> in history.
+// Action segment of log://turn_N/action/slug → XML tag.
 function actionFromPath(path) {
 	if (path?.startsWith("prompt://")) return "prompt";
 	const match = path?.match(/^log:\/\/turn_\d+\/([^/]+)\//);
@@ -63,23 +52,30 @@ function renderLogTag(entry, rowsByPath) {
 			: entry.state
 				? stateToStatus(entry.state, entry.outcome)
 				: null;
-	const status = statusValue != null ? ` status="${statusValue}"` : "";
+	// Suppress status on prompts; uniform 200 carries no signal.
+	const status =
+		statusValue != null && action !== "prompt"
+			? ` status="${statusValue}"`
+			: "";
 	const outcomeAttr = entry.outcome ? ` outcome="${entry.outcome}"` : "";
-	// `tokens=` is the promotion premium (aTokens) of the thing this tag
-	// represents — what the model would free by demoting it. For actions
-	// that reference a separate data entry (get/set/mv/cp), resolve via
-	// attrs.path and report the target's aTokens. For actions whose log
-	// body IS the cost-bearing content (search/update/error/ask_user,
-	// plus <get> slice reads), use the log entry's own aTokens. sh/env
-	// span multiple channel entries and are omitted — the channels
-	// render their own tokens in <context>.
+	// tokens = aTokens of the thing this tag represents (target via attrs.path, else self).
 	const isSlice = attrs?.lineStart != null;
 	const targetEntry = attrs?.path ? rowsByPath.get(attrs.path) : null;
 	let tokenSource = null;
-	if (STREAM_NO_TOKENS.has(action)) tokenSource = null;
-	else if (isSlice) tokenSource = entry.aTokens;
-	else if (targetEntry) tokenSource = targetEntry.aTokens;
-	else tokenSource = entry.aTokens;
+	let lineSource = null;
+	if (STREAM_NO_TOKENS.has(action)) {
+		tokenSource = null;
+		lineSource = null;
+	} else if (isSlice) {
+		tokenSource = entry.aTokens;
+		lineSource = entry.vLines;
+	} else if (targetEntry) {
+		tokenSource = targetEntry.aTokens;
+		lineSource = targetEntry.vLines;
+	} else {
+		tokenSource = entry.aTokens;
+		lineSource = entry.vLines;
+	}
 	const tokens = tokenSource != null ? ` tokens="${tokenSource}"` : "";
 	const summary =
 		typeof attrs?.summary === "string"
@@ -89,16 +85,13 @@ function renderLogTag(entry, rowsByPath) {
 		typeof attrs?.query === "string" ? ` query="${attrs.query}"` : "";
 	const command =
 		typeof attrs?.command === "string" ? ` command="${attrs.command}"` : "";
-	// target= is the path the action touched (e.g. the file/known that was
-	// set, the URL that was fetched). Plugins store it in attrs.path when
-	// they write the log entry.
 	const target = attrs?.path ? ` target="${attrs.path}"` : "";
-	// Slice reads tag the log entry with lineStart/lineEnd/totalLines so
-	// the <get> tag surfaces `lines="a-b/total"` — a concrete handle for
-	// the model to re-issue or compare against another slice.
+	// Slice reads emit lines="a-b/total"; others emit simple lines="N".
 	const lines = isSlice
 		? ` lines="${attrs.lineStart}-${attrs.lineEnd}/${attrs.totalLines}"`
-		: "";
+		: lineSource != null
+			? ` lines="${lineSource}"`
+			: "";
 
 	const attrStr = `${target}${status}${outcomeAttr}${query}${command}${summary}${lines}${tokens}`;
 

package/src/plugins/ollama/ollama.js

@@ -1,17 +1,13 @@
+import config from "../../agent/config.js";
 import msg from "../../agent/messages.js";
+import { chatCompletionStream } from "../../llm/openaiStream.js";
+import { retryWithBackoff } from "../../llm/retry.js";
 
-const FETCH_TIMEOUT = Number(process.env.RUMMY_FETCH_TIMEOUT);
-if (!FETCH_TIMEOUT) throw new Error("RUMMY_FETCH_TIMEOUT must be set");
+const { FETCH_TIMEOUT } = config;
 
 const PROVIDER = "ollama";
 
-/**
- * Ollama LLM provider plugin. Registers with hooks.llm.providers if
- * OLLAMA_BASE_URL is set; inert otherwise. Handles model aliases of the
- * form `ollama/{modelName}` — e.g. `ollama/llama3.1:8b` or
- * `ollama/library/qwen:7b` (Ollama accepts both bare and
- * registry-qualified model names).
- */
+// Inert unless OLLAMA_BASE_URL is set; ollama/{model[/registry]} aliases.
 export default class Ollama {
 	#baseUrl;
 
@@ -41,70 +37,58 @@ export default class Ollama {
 			? AbortSignal.any([options.signal, timeoutSignal])
 			: timeoutSignal;
 
-		const response = await fetch(`${this.#baseUrl}/v1/chat/completions`, {
-			method: "POST",
-			headers: { "Content-Type": "application/json" },
-			body: JSON.stringify(body),
-			signal,
-		});
-
-		if (!response.ok) {
-			const error = await response.text();
-			throw new Error(
-				msg("error.ollama_api", { status: `${response.status} - ${error}` }),
-			);
-		}
-
-		const data = await response.json();
-
-		for (const choice of data.choices) {
-			const m = choice.message;
-			if (!m) continue;
-			const parts = [m.reasoning_content, m.reasoning, m.thinking].filter(
-				Boolean,
-			);
-			m.reasoning_content =
-				parts.length > 0 ? [...new Set(parts)].join("\n") : null;
+		try {
+			return await chatCompletionStream({
+				url: `${this.#baseUrl}/v1/chat/completions`,
+				headers: {},
+				body,
+				signal,
+			});
+		} catch (err) {
+			if (err.status) {
+				throw new Error(
+					msg("error.ollama_api", { status: `${err.status} - ${err.body}` }),
+				);
+			}
+			throw err;
 		}
-
-		return data;
 	}
 
 	async #getContextSize(model) {
-		for (let attempt = 0; attempt < 3; attempt++) {
-			try {
-				const response = await fetch(`${this.#baseUrl}/api/show`, {
-					method: "POST",
-					headers: { "Content-Type": "application/json" },
-					body: JSON.stringify({ model }),
-					signal: AbortSignal.timeout(FETCH_TIMEOUT),
-				});
-				if (!response.ok) {
-					throw new Error(
-						msg("error.ollama_show_failed", {
-							status: response.status,
-							baseUrl: this.#baseUrl,
-						}),
-					);
-				}
-				const data = await response.json();
-				if (data.model_info) {
-					for (const [key, value] of Object.entries(data.model_info)) {
-						if (key.endsWith(".context_length")) return value;
-					}
-				}
-				throw new Error(msg("error.ollama_no_context_length", { model }));
-			} catch (err) {
-				if (err.message.includes("Ollama")) throw err;
-				if (attempt < 2) {
-					await new Promise((r) => setTimeout(r, (attempt + 1) * 2000));
-					continue;
-				}
+		const fetchContext = async () => {
+			const response = await fetch(`${this.#baseUrl}/api/show`, {
+				method: "POST",
+				headers: { "Content-Type": "application/json" },
+				body: JSON.stringify({ model }),
+				signal: AbortSignal.timeout(FETCH_TIMEOUT),
+			});
+			if (!response.ok) {
 				throw new Error(
-					msg("error.ollama_unreachable", { baseUrl: this.#baseUrl }),
-					{ cause: err },
+					msg("error.ollama_show_failed", {
+						status: response.status,
+						baseUrl: this.#baseUrl,
+					}),
 				);
 			}
+			const data = await response.json();
+			if (data.model_info) {
+				for (const [key, value] of Object.entries(data.model_info)) {
+					if (key.endsWith(".context_length")) return value;
+				}
+			}
+			throw new Error(msg("error.ollama_no_context_length", { model }));
+		};
+		try {
+			return await retryWithBackoff(fetchContext, {
+				deadlineMs: FETCH_TIMEOUT,
+				isRetryable: (err) => !err.message.includes("Ollama"),
+			});
+		} catch (err) {
+			if (err.message.includes("Ollama")) throw err;
+			throw new Error(
+				msg("error.ollama_unreachable", { baseUrl: this.#baseUrl }),
+				{ cause: err },
+			);
 		}
 	}
 }