@lawaty/lacode 0.5.0 → 0.7.0

package/dist/index.js

@@ -611,8 +611,9 @@ function coolProvider(s, provider, until, reason) {
   return affected;
 }
 
-// src/lib/manager-directive.ts
-var MANAGER_AGENT_NAME = "manager";
+// src/lib/agent-instructions.ts
+var AGENT_INSTRUCTIONS_COMMAND = "agent-instructions";
+var AGENT_INSTRUCTIONS_WRITER = "instructions";
 
 // src/agents.ts
 var POOLED_BASES = ["explore", "implement", "context"];
@@ -650,14 +651,28 @@ different number.
 
 The pooled agents are the ONLY route to investigation and implementation. If the
 pool cannot serve a spawn \u2014 every slot cooling or exhausted, spawns repeatedly
-failing, or a runtime message saying no slot is available \u2014 that is a STOP
-condition, not a reason to do the work yourself.
+failing, or a runtime message saying no slot is available \u2014 that is not a reason
+to do the work yourself, and not a reason to stop and ask.
+
+**Retry.** A refused or cooling slot is transient and costs you nothing but a
+turn. Spawn the bare base again. Do not escalate a \`-N\` variant when a refusal
+names one \u2014 the refusal means the pool has a routing problem, not that you
+picked the wrong number.
+
+Retry is always available and never wrong, which is why it is the default. But
+"never wrong" is not "sufficient": looping silently against a pool with every
+model cooling goes nowhere and burns the slot doing it. So separate the two
+failures. **Transient** \u2014 a cooldown, a rate limit, one slot at its ceiling \u2014
+means retry. **Persistent or structural** \u2014 every slot cooling, the same spawn
+failing repeatedly with the same error, a malformed model id \u2014 means stop
+retrying and *report* it: say what failed (slot, model, error) and continue
+with whatever else the task needs, or ask the user. Surface the failure; do not
+spin on it and do not let it end the whole task.
 
 **Never fall back to reading, grepping, exploring or implementing on your own.**
-Do not "just quickly" inspect a file because a spawn failed. Stop, say what
-failed (slot, model, error), and ASK the user how to proceed: wait for a slot to
-recover, change the pool, or obtain explicit permission to work without
-subagents. A failed spawn is never permission, and silence is not consent.
+Do not "just quickly" inspect a file because a spawn failed. A failed spawn is
+never permission, and silence is not consent \u2014 that part does not change with
+cooldowns, and no amount of retrying grants it.
 
 \`edit\`/\`write\`/\`bash\` stay yours for applying and verifying what a subagent
 returned \u2014 never as a substitute for a spawn that could not run.
@@ -736,7 +751,12 @@ missing). Never paste context file contents into the report.`,
       "edit": {
         "*": "allow",
         "*/.opencode/context/**": "deny",
-        ".opencode/context/**": "deny"
+        ".opencode/context/**": "deny",
+        // AGENTS.md has exactly one writer, the hidden `instructions` agent.
+        // opencode resolves a file permission with findLast over the map's own
+        // entry order, so this deny has to come AFTER the broad "*" allow above.
+        "AGENTS.md": "deny",
+        "*/AGENTS.md": "deny"
       },
       "bash": "allow",
       "webfetch": "deny",
@@ -979,7 +999,11 @@ if the blocker is understanding rather than execution, report that instead.`,
       "edit": {
         "*": "allow",
         "*/.opencode/context/**": "deny",
-        ".opencode/context/**": "deny"
+        ".opencode/context/**": "deny",
+        // AGENTS.md is written by the hidden `instructions` agent alone; LAST
+        // match wins, so this deny must sit after the broad allow above.
+        "AGENTS.md": "deny",
+        "*/AGENTS.md": "deny"
       },
       "bash": "allow",
       "webfetch": "allow",
@@ -1023,29 +1047,14 @@ You own exactly these five files and nothing else:
   during work that are not already recorded in the project's own decision
   log. Concise; variable length.
 
-A sixth file sits in the same directory and is not yours:
-\`.opencode/context/manager.md\`. It is the client's own statement of intent
-\u2014 what is being worked on, what comes next, what they have settled, and
-what every agent should keep in mind. It deliberately says nothing about
-where code lives: your five files and the source itself are the only
-authority on that, and this file never contradicts them. It has no fixed
-schema \u2014 read the whole file, because every heading in it is direction.
-
-You never create, edit, rename, or delete it \u2014 not at bootstrap, not in an
-update, not in a review, and not when a run of your own concludes the work
-has moved on. The client edits it, or asks a live agent to edit it on
-their explicit instruction. Either way it is their hand, never yours.
-
-Read it before you write anything. Its \`## Current\` and \`## Next\`
-sections tell you what is actually being worked on, and \`## Confirmed\`
-tells you which of your facts the client has ratified. A fact you are
-writing that appears under \`## Confirmed\` may be stated plainly, without
-a confidence marker; anything else follows your normal rules \u2014 infer
-carefully, mark uncertainty. Never copy its content into the five files.
-
-A missing or empty section is not an instruction. It means the client has
-not said, not that you may assume anything \u2014 proceed exactly as you would
-without the file.
+# Current-state only
+
+Every file you write states only the current arrangement and the reason it
+holds, in the present tense. No "used to", "previously", "before this",
+migration notes or deprecation narrative \u2014 history belongs in the project's
+changelog, not in your map. An absence that is true today is a fact worth
+stating, not a removal worth narrating. When a past state is worth keeping,
+record it briefly in \`changelog.md\` instead.
 
 Never create, modify, or delete anything outside \`.opencode/context/\`. The
 \`edit\` permission enforces this. Do not attempt to work around it with bash.
@@ -1081,11 +1090,6 @@ Then write the five files. For each significant subsystem, give a short
 reason. This is the single most valuable thing you produce: where an agent
 should open its editor next.
 
-manager.md is created for the client by the plugin, not by you; if it is
-missing anyway, mention once in your final report that they can add it
-themselves to record what is being worked on, what comes next, and what
-they have ratified. Never create it.
-
 # Incremental update
 
 Trigger: \`.opencode/context/\` exists and work has just been done here.
@@ -1113,14 +1117,6 @@ that is already stated in another context file \u2014 one fact has one home, and
 the other files link to it. If nothing material changed, make no edits and
 report "no update needed".
 
-Before editing, re-read manager.md: its \`## Current\` and \`## Next\`
-sections tell you what the client is actually working on, and its
-\`## Confirmed\` section settles which of your facts are ratified. If what
-you are about to write would place the code somewhere the client never
-described, trust the code, make the edit the code justifies, and report
-the gap in your final message \u2014 never reconcile the two sides silently,
-and never edit manager.md to close it.
-
 # Minimal exploration
 
 Default shape:
@@ -1215,14 +1211,10 @@ Audit the five files for:
 - duplicated information \u2014 one canonical place per fact
 - missing canonical entry points for significant subsystems
 - unsupported assumptions \u2014 anything unmarked that you cannot verify
-- manager.md, if present: is \`## Current\` and \`## Next\` still true of
-  recent work, is every \`## Confirmed\` entry still true of the code, has
-  it outgrown roughly 150 lines \u2014 report findings, but never edit, prune,
-  or reorganize it; that is the client's call
 
-Fix problems only in the five files you own \u2014 never in manager.md, which
-belongs to the client. Never modify source code during a review. Report
-what you found, what you fixed, and what you could not verify.
+Fix problems only in the five files you own. Never modify source code
+during a review. Report what you found, what you fixed, and what you
+could not verify.
 
 # Reporting
 
@@ -1240,9 +1232,7 @@ Finish with a short report:
       "edit": {
         "*": "deny",
         "*/.opencode/context/**": "allow",
-        ".opencode/context/**": "allow",
-        "*/.opencode/context/manager.md": "deny",
-        ".opencode/context/manager.md": "deny"
+        ".opencode/context/**": "allow"
       },
       "bash": {
         "*": "deny",
@@ -1290,58 +1280,60 @@ Finish with a short report:
     }
   }
 };
-var MANAGER_AGENT = {
-  description: "Records and edits the client's .opencode/context/manager.md. Run only by the /manager-directive command; no agent or subagent can spawn it.",
+var INSTRUCTIONS_AGENT = {
+  description: "Edits the project's AGENTS.md. Run only by the /agent-instructions command; no agent or subagent can spawn it.",
   mode: "subagent",
   hidden: true,
   temperature: 0.1,
   prompt: `# Role
 
-You maintain one file and nothing else: \`.opencode/context/manager.md\`, the
-client's own statement of intent. You are not a coding agent and you touch no
-other path.
+You maintain one file and nothing else: the project's \`AGENTS.md\`. You are not a
+coding agent and you touch no other path.
 
-You are run only by the \`/manager-directive\` command. The client tells you what
+You are run only by the \`/agent-instructions\` command. The client tells you what
 they want recorded or changed in the file; you apply that to the file and report
 what you changed. You have no permission to edit anything else, and you need
 none.
 
 # The file
 
-It records what is being worked on (\`## Current\`), what comes next
-(\`## Next\`), what the client has ratified (\`## Confirmed\`), and standing
-direction (\`## Instructions\`). The client may use any headings \u2014 every heading
-is direction, and the four above are a convenience, not a schema.
+\`AGENTS.md\` is the project's standing instructions \u2014 the rules every agent and
+subagent runs under. opencode loads it into every session automatically, so what
+you write here steers all work in this project. Keep it about how to work:
+conventions, constraints, priorities, things to always do and never do. It says
+nothing about where code lives \u2014 the source and the context map are the authority
+on that.
 
 # How to apply a directive
 
-- Do what the client asked. "Append", "rewrite Current", "move X to Confirmed",
-  "add a heading" all mean exactly that; a bare sentence goes under the most
-  fitting existing heading, and you create one only if none fits.
+- Do what the client asked: add, rewrite, move or remove exactly what they asked
+  for. A bare sentence goes under the most fitting existing heading, and you
+  create a heading only if none fits.
 - Never invent facts, never record your own inference as the client's, and never
   tidy, reorder or delete content the client did not ask you to touch.
-- \`## Current\` is rewritten, not appended to, so it does not become a running
-  history.
+- Prefer current-state wording: state how things are, not how they changed.
 - Keep the file concise. If it is growing past roughly 150 lines, say so.
+- If \`AGENTS.md\` does not exist, stop and report that there is nothing to edit.
+  Do not create it.
 
 # Reporting
 
 Report briefly what you changed and where. Do not edit any other file.`,
   permission: {
-    // Read is scoped to the one file too: the directive is about manager.md, and
+    // Read is scoped to the one file too: the directive is about AGENTS.md, and
     // nothing else is any of this agent's business.
     read: {
       "*": "deny",
-      ".opencode/context/manager.md": "allow",
-      "*/.opencode/context/manager.md": "allow"
+      "AGENTS.md": "allow",
+      "*/AGENTS.md": "allow"
     },
     glob: "deny",
     grep: "deny",
     list: "deny",
     edit: {
       "*": "deny",
-      ".opencode/context/manager.md": "allow",
-      "*/.opencode/context/manager.md": "allow"
+      "AGENTS.md": "allow",
+      "*/AGENTS.md": "allow"
     },
     bash: "deny",
     webfetch: "deny",
@@ -1367,7 +1359,7 @@ function buildAgentSet() {
       set[`${base}-${slot}`] = variant;
     }
   }
-  set[MANAGER_AGENT_NAME] = MANAGER_AGENT;
+  set[AGENT_INSTRUCTIONS_WRITER] = INSTRUCTIONS_AGENT;
   return set;
 }
 function applyAgents(cfg, agents = buildAgentSet(), opts = {}) {
@@ -1407,13 +1399,7 @@ Remember: write conclusions, not inventories. Link to the project's existing
 documentation instead of duplicating it.
 
 Report what you created or updated, what you deliberately left out, and
-anything that contradicts the code or existing docs.
-
-If \`.opencode/context/manager.md\` does not exist, mention in your final
-report that the client can add it themselves \u2014 their own file with
-\`## Current\`, \`## Next\`, \`## Confirmed\` and \`## Instructions\` \u2014 for what
-is being worked on, what comes next, what they have ratified, and standing
-direction. Never create it.`
+anything that contradicts the code or existing docs.`
   },
   "context-update": {
     description: "Incrementally update .opencode/context/ after recent work via the context agent.",
@@ -1444,9 +1430,8 @@ timestamp-only edits. If nothing material changed, make no edits and report
     subtask: true,
     template: `Context review.
 
-Everything outside the five context files is read-only for this run \u2014
-including \`manager.md\`, which belongs to the client. Do not modify source
-code under any circumstances.
+Everything outside the five context files is read-only for this run. Do not
+modify source code under any circumstances.
 
 Audit the five context files for:
 
@@ -1459,18 +1444,15 @@ Audit the five context files for:
 - missing canonical entry points for significant subsystems
 - unsupported assumptions \u2014 anything stated as fact that you cannot verify,
   or that lacks a confidence marker
-- \`manager.md\`, if present: is \`## Current\` and \`## Next\` still true of
-  recent work, is every \`## Confirmed\` entry still true of the code, has it
-  outgrown roughly 150 lines \u2014 report findings, but never edit it
 
 Fix what you find in \`.opencode/context/\` only, then report: findings, fixes,
 and anything you could not verify.`
   },
-  "manager-directive": {
-    description: "Apply a client instruction to .opencode/context/manager.md. Client-only \u2014 runs the manager agent, the only writer of that file.",
-    agent: MANAGER_AGENT_NAME,
+  [AGENT_INSTRUCTIONS_COMMAND]: {
+    description: "Apply a client instruction to the project's AGENTS.md. Client-only \u2014 runs the instructions agent, the only writer of that file.",
+    agent: AGENT_INSTRUCTIONS_WRITER,
     subtask: true,
-    template: `Apply this instruction to \`.opencode/context/manager.md\`, and edit only that file:
+    template: `Apply this instruction to the project's \`AGENTS.md\`, and edit only that file:
 
 $ARGUMENTS
 
@@ -1506,6 +1488,36 @@ Print one PASS/FAIL line per check below, with the exact fix for any FAIL.
    whether or not the gateway advertises it, so it needs no allowlist entry.
 
 Change nothing. Print a final one-line summary of failures.`
+  },
+  "pool": {
+    description: "Modify this machine's LaCode pool from a short instruction \u2014 the shared free-model slot table in ~/.config/lacode/pool.txt. Restarts opencode afterwards.",
+    template: `Apply this change to the LaCode pool on this machine:
+
+$ARGUMENTS
+
+The pool is ONE global slot table shared by every pooled base (explore, implement, context) \u2014 there are no per-base model lists, so "put X on explore" means "give X a slot", not a per-base override.
+
+1. Establish the baseline. Read \`${USER_MODELS_TXT}\`. If it is absent, the pool is resolving from the bundled preset \`presets/free-tier.json\`; read that file to learn what is in effect. Call the \`pool_status\` tool only if you need live load or cooldown state.
+
+2. Resolve the instruction above into a target slot table. These are the rules the loader enforces \u2014 never write a file that breaks one:
+   - one slot per line: \`<provider>/<model-id>  [weight]\`
+   - at most ${MAX_SLOTS} slots, and a model may appear only once
+   - weight is optional (default 1) and must be a number in (0, ${MAX_WEIGHT}]. It is the soft ceiling on concurrent claims and the slot's share of the rotation \u2014 NOT a priority; pick order follows line order.
+   - \`#\` starts a comment (whole-line or trailing); blank lines are ignored.
+   If the instruction names a model ambiguously, or would need more than ${MAX_SLOTS} slots, stop and ask rather than guessing.
+
+3. Write the WHOLE file, not a patch. This file is an override, not an edit target: any slot you drop disappears, and slot numbers are line positions, so a partial write silently renumbers the rotation. Keep the existing comment header. If the write fails on a missing directory, create \`~/.config/lacode\` first and retry.
+
+4. Report the before \u2192 after slot table, the resolved path, and this reminder in substance: nothing hot-reloads, so every running opencode process must be restarted before the new pool takes effect.
+
+Two traps that fail silently, so check and state them explicitly:
+- A slot's provider must already exist in the user's opencode config, or that slot cannot resolve at all. If the instruction adds a model on a provider that is not configured, say so.
+- If a slot model is \`omniroute/<id>\`, that id must also appear in the plugin's \`omniroute.models\` allowlist (or be the \`combo\` id, which needs no entry) \u2014 otherwise the pool resolves and every spawn fails with \`ProviderModelNotFoundError\`.
+- A provider \`whitelist\` is an allowlist, and the plugin appends your model to it at config time. Nothing for you to do there.
+
+If the instruction is to RESET or REVERT rather than to change models, delete \`${USER_MODELS_TXT}\` instead, and report that the pool falls back to the bundled preset \u2014 with the same restart reminder.
+
+Change nothing else. In particular do not touch \`presets/free-tier.json\`, \`~/.config/lacode/pool.json\`, \`FALLBACK_SLOTS\` in \`src/lib/pool.ts\`, or \`tests/models-snapshot.json\`: this command owns the per-machine override and nothing else.`
   }
 };
 function applyCommands(cfg, commands = COMMANDS) {
@@ -1522,45 +1534,6 @@ function applyCommands(cfg, commands = COMMANDS) {
   }
   return { added, kept };
 }
-var INSTRUCTION_MARKER = "<!-- lacode:context-map-instruction:v1 -->";
-var INSTRUCTION_OPT_OUT = "<!-- lacode:context-map-instruction:off -->";
-var STANDING_INSTRUCTION = `## Repository context map
-
-This project may keep a navigation map in \`.opencode/context/\`. Before broad
-exploration, check whether it exists (\`ls .opencode/context\` \u2014 never the \`glob\`
-tool, which skips hidden directories). If \`manager.md\` exists there, read it
-first: it is the client's own statement of intent \u2014 what is being worked on,
-what comes next, \`## Confirmed\` lists what the client has ratified, and
-\`## Instructions\` holds standing managerial directions. It outranks every
-other context file and every inference of yours about what matters right now.
-It says nothing about where code lives, so it cannot contradict the code;
-treat its \`## Confirmed\` entries as settled decisions, not as a licence to
-skip reading the code. Any heading in it is direction, so read the whole
-file rather than only the sections above. The rest of the map remains usable,
-just unratified where \`## Confirmed\` does not cover it. Then read only the
-file relevant to the task \u2014 \`architecture.md\`, \`contexts.md\`,
-\`conventions.md\`, \`workflows.md\`, \`decisions.md\` \u2014 and start from the
-canonical entry points it lists.
-
-Source code is authoritative: if the map disagrees with the code,
-trust the code and say so; never edit source to match it.
-Never edit the five context files yourself \u2014 the context agent owns them.
-\`manager.md\` is the client's own file and the context agent is explicitly
-permission-denied from writing it. You may write it only when the client
-explicitly asks you to in this conversation, and then only what they asked
-for; never on your own initiative, and never to "keep it up to date". Say in
-your final report if your work revealed something that belongs to the map, or
-if \`manager.md\` and the code disagree.`;
-function applyInstructions(cfg) {
-  const existing = Array.isArray(cfg.instructions) ? cfg.instructions : [];
-  const texts = existing.filter((e) => typeof e === "string");
-  if (texts.some((e) => e.includes(INSTRUCTION_OPT_OUT))) return { added: false };
-  if (texts.some((e) => e.includes(INSTRUCTION_MARKER))) return { added: false };
-  cfg.instructions = [...existing, `${INSTRUCTION_MARKER}
-
-${STANDING_INSTRUCTION}`];
-  return { added: true };
-}
 
 // src/plugins/agent-pool.ts
 import { tool } from "@opencode-ai/plugin";
@@ -1896,7 +1869,7 @@ round robin in slot order, each slot taking as many turns as its weight says, so
 });
 
 // src/plugins/context-autoupdate.ts
-import { existsSync as existsSync2, readdirSync as readdirSync2, statSync as statSync2, mkdirSync as mkdirSync2, writeFileSync as writeFileSync2 } from "node:fs";
+import { existsSync as existsSync2, readdirSync as readdirSync2, statSync as statSync2 } from "node:fs";
 import * as path2 from "node:path";
 
 // src/lib/jsonc.ts
@@ -1951,7 +1924,7 @@ import { readFileSync as readFileSync3 } from "node:fs";
 import * as path from "node:path";
 var CONTEXT_DIR = path.join(".opencode", "context");
 var WRITER = "context";
-var MANAGER_WRITER = MANAGER_AGENT_NAME;
+var INSTRUCTIONS_WRITER = AGENT_INSTRUCTIONS_WRITER;
 var toPosix = (value) => value.split(path.sep).join("/");
 var compilePattern = (pattern) => {
   const escaped = toPosix(pattern).replace(/[.+^${}()|[\]\\]/g, "\\$&").replace(/\*/g, ".*").replace(/\?/g, ".");
@@ -1980,8 +1953,8 @@ var evaluatedForm = (mapDir, projectRoot) => {
   const absolute = toPosix(path.join(mapDir, "architecture.md"));
   return projectRoot ? toPosix(path.relative(projectRoot, absolute)) : absolute;
 };
-var managerForm = (mapDir, projectRoot) => {
-  const absolute = toPosix(path.join(mapDir, "manager.md"));
+var agentsForm = (mapDir, projectRoot) => {
+  const absolute = toPosix(path.join(mapDir, "..", "..", "AGENTS.md"));
   return projectRoot ? toPosix(path.relative(projectRoot, absolute)) : absolute;
 };
 var verifyWriterRule = (configPath, mapDir, projectRoot) => {
@@ -2015,17 +1988,17 @@ var verifyWriterRule = (configPath, mapDir, projectRoot) => {
       return `context map has more than one writer: ${offenders.join(", ")} would be granted the map by permission, evaluating "${form}". Only ${WRITER} may write the map. Narrow the grant and put the deny LAST -- opencode takes the LAST matching rule, so "${CONTEXT_DIR}/**": "deny" written after "*": "allow" is what denies it: "*": "allow", "*/${CONTEXT_DIR}/**": "deny", "${CONTEXT_DIR}/**": "deny". Both spellings are required because the evaluated path depends on the project root.`;
     }
   }
-  const managerPath = managerForm(mapDir, projectRoot);
+  const agentsPath = agentsForm(mapDir, projectRoot);
   const agentDefs = readAgents(raw);
   if (agentDefs) {
     const writers = [];
     for (const [name, def] of Object.entries(agentDefs)) {
-      if (name === MANAGER_WRITER) continue;
-      const win = winningRule(permissionOf(def), managerPath);
+      if (name === INSTRUCTIONS_WRITER) continue;
+      const win = winningRule(permissionOf(def), agentsPath);
       if (win?.action === "allow") writers.push(`${name} (edit rule "${win.pattern}": "${win.action}")`);
     }
     if (writers.length > 0) {
-      return `${CONTEXT_DIR}/manager.md is writable by permission: ${writers.join(", ")}. It is the client's own file \u2014 what they are working on, what is ratified, what every agent must respect \u2014 and only the \`${MANAGER_WRITER}\` agent (run solely by /manager-directive) may write it. Deny it LAST on every other agent, after any broad allow, because opencode takes the LAST matching rule: "*": "allow", "*/${CONTEXT_DIR}/manager.md": "deny", "${CONTEXT_DIR}/manager.md": "deny". Declare this on every agent above, including \`${WRITER}\`: an agent you define yourself replaces the shipped one wholesale, permission and all.`;
+      return `AGENTS.md is writable by permission: ${writers.join(", ")}. It is the project's own file \u2014 the standing instructions every agent runs under \u2014 and only the \`${INSTRUCTIONS_WRITER}\` agent (run solely by /agent-instructions) may write it. Deny it LAST on every other agent, after any broad allow, because opencode takes the LAST matching rule: "*": "allow", "AGENTS.md": "deny", "*/AGENTS.md": "deny". Declare this on every agent above, including \`${WRITER}\`: an agent you define yourself replaces the shipped one wholesale, permission and all.`;
     }
   }
   return void 0;
@@ -2054,51 +2027,6 @@ var IGNORED_SEGMENTS = /* @__PURE__ */ new Set([
   ".cache",
   ".git"
 ]);
-var MANAGER_TEMPLATE = `# Managerial directions
-
-This file is yours. No agent writes it: the context agent is explicitly denied
-permission to touch this path, and nothing changes it except you, or an agent
-acting on your explicit instruction in that conversation.
-
-It records what you want done and how you want it approached \u2014 what is being
-worked on, what comes next, what you have settled, and what every agent should
-keep in mind. It deliberately says nothing about where code lives: the source
-and the five generated context files are the only authority on that, so this
-file can never contradict the code.
-
-Add any headings you like. The four below are a convenience, not a schema \u2014
-every heading in this file is direction.
-
-## Current
-
-<!-- What is being worked on right now. Rewrite it; do not append a history. -->
-
-## Next
-
-<!-- What comes next, in order. -->
-
-## Confirmed
-
-<!-- What you have personally ratified. Nothing enters or leaves this list
-     except by your hand \u2014 no agent promotes, prunes or rewrites it. -->
-
-## Instructions
-
-<!-- Standing direction that applies beyond the current task: priorities, things
-     to care about, constraints that are not obvious from the code. -->`;
-var ensureManagerFile = (root) => {
-  try {
-    const target = path2.join(root, ".opencode", "context", "manager.md");
-    if (existsSync2(target)) return { created: false };
-    mkdirSync2(path2.dirname(target), { recursive: true });
-    writeFileSync2(target, MANAGER_TEMPLATE, { flag: "wx" });
-    return { created: true };
-  } catch (error) {
-    if (error?.code === "EEXIST") return { created: false };
-    console.warn(`[context-autoupdate] ensureManagerFile failed: ${error}`);
-    return { created: false };
-  }
-};
 var toPosix2 = (value) => value.split(path2.sep).join("/");
 var findProjectConfig = () => {
   let dir = import.meta.dir ?? path2.dirname(new URL(import.meta.url).pathname);
@@ -2142,7 +2070,6 @@ var contextAutoUpdateHooks = (async ({ client, directory }, options) => {
   });
   const root = path2.resolve(directory);
   const contextAbs = path2.join(root, CONTEXT_DIR2);
-  if (ensureManagerFile(root).created) void log("info", `created ${CONTEXT_DIR2}/manager.md`);
   const config = findProjectConfig();
   const writerProblem = config ? verifyWriterRule(config, contextAbs, root) : void 0;
   if (writerProblem) void log("error", writerProblem);
@@ -3102,7 +3029,7 @@ var omniRouteHooks = async (input, options, deps = {}) => {
 
 // src/lib/selfupdate.ts
 import { spawn as spawn2 } from "node:child_process";
-import { closeSync, mkdirSync as mkdirSync3, openSync, readFileSync as readFileSync4, renameSync as renameSync2, rmSync, statSync as statSync3, unlinkSync as unlinkSync2, writeFileSync as writeFileSync3, writeSync } from "node:fs";
+import { closeSync, mkdirSync as mkdirSync2, openSync, readFileSync as readFileSync4, renameSync as renameSync2, rmSync, statSync as statSync3, unlinkSync as unlinkSync2, writeFileSync as writeFileSync2, writeSync } from "node:fs";
 import { homedir as homedir2 } from "node:os";
 import { dirname as dirname3, join as join6, resolve as resolve2 } from "node:path";
 var REGISTRY_LATEST = `https://registry.npmjs.org/${PACKAGE_NAME}/latest`;
@@ -3196,8 +3123,8 @@ function readState(path4 = STATE_PATH) {
 }
 function writeState(state2, path4 = STATE_PATH) {
   try {
-    mkdirSync3(dirname3(path4), { recursive: true });
-    writeFileSync3(path4, JSON.stringify(state2));
+    mkdirSync2(dirname3(path4), { recursive: true });
+    writeFileSync2(path4, JSON.stringify(state2));
   } catch {
   }
 }
@@ -3223,7 +3150,7 @@ function acquireLock(path4 = LOCK_PATH, now = Date.now()) {
     };
   };
   try {
-    mkdirSync3(dirname3(path4), { recursive: true });
+    mkdirSync2(dirname3(path4), { recursive: true });
     return claim();
   } catch {
     let orphaned = false;
@@ -3411,9 +3338,6 @@ var LaCode = (async (input, options) => {
       if (cmds.added.length > 0) {
         await log("info", `commands: added ${cmds.added.join(", ")}; kept ${cmds.kept.join(", ") || "none"} already defined`);
       }
-      if (applyInstructions(cfg).added) {
-        await log("info", "instructions: ensured the standing context-map rule is present");
-      }
       await omniroute.config?.(cfg);
       await pool.config?.(cfg);
     } catch (error) {

package/docs/BENCHMARKS.md

@@ -88,10 +88,9 @@ never priced at 0 and never priced at a guess. If you name such a model
 explicitly with `--model` / `--baseline-model`, the run exits non-zero and tells
 you to refresh or to pass `--allow-unknown`.
 
-This is the reason the results table below is empty. The pinned snapshot covers
-the bundled free preset only; the default baseline model is not in it, so the
-default offline run cannot produce an honest ratio. That is the harness working
-as designed.
+This is why the results table below is empty. The pinned snapshot covers the
+bundled free preset only; the default baseline model is not in it, so the
+default offline run cannot produce an honest ratio.
 
 ## Results
 
@@ -116,11 +115,10 @@ harness prints `unknown`.
 `--out docs/BENCHMARKS.md` and exits 2.
 
 That refusal is the point. This file ships inside the npm tarball (`files`
-includes `docs`). A generated table dropped on top of it would replace the
-explanation with a dated artifact that reads as a published measurement of the
-package, and it would outlive the machine and the day that produced it. So the
-hand-written explainer is protected, and the generated output — which carries a
-footer pointing back here — goes to its own gitignored filename.
+includes `docs`) and is the hand-written explainer. A generated table is a
+measurement of one machine on one day, not a claim about the package, so the
+explainer is protected and the generated output — which carries a footer
+pointing back here — goes to its own gitignored filename.
 
 ```bash
 npm run benchmark -- --out docs/BENCHMARKS.generated.md

package/docs/CONFIGURATION.md

@@ -43,10 +43,10 @@ writes the context map and has no business in a browser.
 opencode folds an agent's `tools` into its permission list **at config-load
 time**, while decoding the config — before any plugin runs. A plugin's `config`
 hook mutates the already-decoded config, so the fold never re-runs and anything
-it puts in `agent.tools` is inert. (Verified on opencode 1.18.34/1.18.35 with
-`opencode debug agent explore`: a grant written in your own config shows two
-`playwright_*` entries — `deny` then `allow`, last one winning — while the same
-grant injected by the plugin shows only the `deny`.) LaCode's other injected
+it puts in `agent.tools` is inert. On opencode 1.18.34/1.18.35, a grant
+written in your own config shows two `playwright_*` entries — `deny` then
+`allow`, last one winning — while the same grant injected by the plugin shows
+only the `deny`. LaCode's other injected
 fields are not affected; `permission`, `prompt`, `model`, `description`,
 `temperature`, `mode` and `hidden` are all read after the plugin hook runs.
 
@@ -78,8 +78,8 @@ process.
 **18 agent keys**: the 5 bases `build`, `plan`, `explore`, `implement`,
 `context`, plus one hidden
 `<base>-1`..`-4` pool variant for each of the 3 pooled bases, plus the hidden
-`manager` agent — the one writer of `.opencode/context/manager.md`, reachable
-only from the `/manager-directive` command.
+`instructions` agent — the one writer of the project's root `AGENTS.md`,
+reachable only from the `/agent-instructions` command.
 
 `build` and `plan` exist in opencode as built-ins, so on a stock config those
 two keys are already taken and merge-without-clobber keeps the built-in — the
@@ -92,13 +92,13 @@ duplicate of `plan` apart from a pinned model, and `plan` is injected instead.
 description — a package cannot ship someone else's provider, quota or benchmark
 results — though its escalation ladder still resolves, because the two deep
 tiers it escalates to *are* injected. And `explore`, `general` and `implement`
-are only `{ "disable": true }` stubs in the author's own config: turning off
-opencode's built-in agents is a decision about one person's configuration, not
-a default a package should impose on a stranger.
+are only `{ "disable": true }` stubs: turning off opencode's built-in agents
+is a decision about one person's configuration, not a default a package should
+impose on a stranger.
 
-One **field** is deliberately not injected either: `tools`, for the
-load-time-folding reason explained above. Nothing else the plugin writes into
-an agent is subject to it.
+One **field** is deliberately not injected: `tools`, for the load-time-folding
+reason explained above. Nothing else the plugin writes into an agent is subject
+to it.
 
 **Installation merges, it does not clobber** (US-16). An agent, command or
 instruction is added only when that key is absent from your config. A