Story Crater Bot
1a4160b3de
process.exit() right after console.log() can drop buffered stdout when it's piped (not a TTY) -- exactly kubectl exec's case. Explains the silent empty-output-exit-1 failures. process.exitCode + natural exit lets the event loop drain and flush first.
832 lines
40 KiB
YAML
832 lines
40 KiB
YAML
apiVersion: v1
|
|
kind: ConfigMap
|
|
metadata:
|
|
name: coordinator-src
|
|
namespace: agent-pod
|
|
data:
|
|
coordinator.js: |
|
|
#!/usr/bin/env node
|
|
// coordinator: local CLI that drives a multi-phase, multi-task pipeline of
|
|
// planner/investigator/implementer/judge stages, same state machine as
|
|
// hub.js's old /pipeline handler. The difference is who owns each stage's
|
|
// session: hub.js spawned raw tmux sessions itself and polled sentinel
|
|
// files on a 10s timer to notice completion. This spawns every interactive
|
|
// stage through the patched agent-manager fork's `spawn` subcommand, so it
|
|
// is a real tracked session (tmux pane on agent-manager's private socket +
|
|
// a state.db row) from the moment it exists -- attachable and visible in
|
|
// agent-manager's own TUI the whole time it runs. The coordinator never
|
|
// kills a stage's session; it rests at an idle prompt once its sentinel
|
|
// file lands, and agent-manager's session list becomes the audit trail of
|
|
// everything the pipeline ran. Completion is still signaled by sentinel
|
|
// files under the task's worktree (unchanged convention), but waited on
|
|
// with fs.watch instead of polling.
|
|
const fs = require("node:fs");
|
|
const path = require("node:path");
|
|
const { spawn } = require("node:child_process");
|
|
|
|
const WORK_DIR = process.env.HUB_WORK_DIR || path.join(require("node:os").tmpdir(), "agent-harness-work");
|
|
|
|
// Never rely on a bare `pi`/`agent-manager` on $PATH -- see PI_BIN's own
|
|
// comment below; the same collision risk applies to any CLI name. Always
|
|
// invoke explicit pinned paths.
|
|
const PI_BIN =
|
|
process.env.PI_BIN ||
|
|
path.join(__dirname, "..", ".pi-cli", "node_modules", "@earendil-works", "pi-coding-agent", "dist", "cli.js");
|
|
|
|
const AGENT_MANAGER_BIN = process.env.AGENT_MANAGER_BIN || path.join(__dirname, "..", ".bin", "agent-manager-fork");
|
|
|
|
// Empty means "let pi fall back to ~/.pi/agent/settings.json's default"
|
|
// (currently anthropic/claude-sonnet-4-5, real paid usage). Set both to
|
|
// route every stage -- headless (spawnPi) and interactive (runStage) --
|
|
// at the homelab model instead, e.g. AGENT_PROVIDER=homelab-ornith
|
|
// AGENT_MODEL=ornith:35b.
|
|
const AGENT_PROVIDER = process.env.AGENT_PROVIDER || "";
|
|
const AGENT_MODEL = process.env.AGENT_MODEL || "";
|
|
|
|
// judge (both per-task and phase-judge -- both spawn with stageLabel
|
|
// "judge", see runTaskInteractive/runPipeline) can run a different model
|
|
// than the rest of the chain, e.g. homelab-reasoning instead of
|
|
// homelab-ornith now that verifier/PRM is retired. Falls back to
|
|
// AGENT_PROVIDER/AGENT_MODEL when unset, so a run that doesn't care keeps
|
|
// one uniform model everywhere, same as before this existed.
|
|
const JUDGE_PROVIDER = process.env.JUDGE_PROVIDER || AGENT_PROVIDER;
|
|
const JUDGE_MODEL = process.env.JUDGE_MODEL || AGENT_MODEL;
|
|
|
|
function providerModelFor(stageLabel) {
|
|
return stageLabel === "judge" ? { provider: JUDGE_PROVIDER, model: JUDGE_MODEL } : { provider: AGENT_PROVIDER, model: AGENT_MODEL };
|
|
}
|
|
|
|
// agent-manager's private tmux server and session-naming scheme
|
|
// (internal/tmux/tmux.go: defaultSocket = "agentmgr", sessionName(id) =
|
|
// "am_"+id) -- stable, documented internals of the fork, used here only
|
|
// for read-only introspection (pane capture) and the RETRY nudge, exactly
|
|
// the class of operation hub.js already ran directly against its own
|
|
// sessions rather than asking a model to do it.
|
|
const AM_SOCKET = "agentmgr";
|
|
function amSessionName(id) {
|
|
return `am_${id}`;
|
|
}
|
|
function runAmTmux(args) {
|
|
return runCmd("tmux", ["-L", AM_SOCKET, ...args]);
|
|
}
|
|
|
|
const ROLE_SKILLS = new Set(["planner", "investigator", "info-collector", "implementer", "judge", "resolver"]);
|
|
|
|
const HARD_RULES =
|
|
"Read and follow ~/.pi/agent/skills/karpathy-guidelines/SKILL.md and " +
|
|
"~/.pi/agent/skills/caveman/SKILL.md as hard rules for this entire task, before anything else. ";
|
|
|
|
function parseVerdictLine(text, label) {
|
|
if (!text) return null;
|
|
const re = new RegExp(`${label}:\\s*(\\w+)`, "i");
|
|
const m = text.match(re);
|
|
return m ? m[1].toUpperCase() : null;
|
|
}
|
|
|
|
function runCmd(bin, args, cwd) {
|
|
return new Promise((resolve) => {
|
|
const child = spawn(bin, args, { cwd, stdio: ["ignore", "pipe", "pipe"] });
|
|
let out = "";
|
|
child.stdout.on("data", (c) => (out += c));
|
|
child.stderr.on("data", (c) => (out += c));
|
|
child.on("close", (code) => resolve({ code, out: out.trim() }));
|
|
});
|
|
}
|
|
|
|
function runGit(cwd, args) {
|
|
return runCmd("git", args, cwd);
|
|
}
|
|
|
|
// A stage saying "commit" in its prompt is a request, not a guarantee -- seen
|
|
// in practice: a stage writes a real file and simply never runs `git add`/
|
|
// `git commit`, leaving it untracked and invisible to every later `git diff`.
|
|
// Sweep and commit anything left dirty after every stage, deterministically.
|
|
async function commitPending(cwd, message) {
|
|
await runGit(cwd, ["add", "-A"]);
|
|
const status = await runGit(cwd, ["status", "--porcelain"]);
|
|
if (!status.out) return { committed: false };
|
|
const commit = await runGit(cwd, ["commit", "-m", message]);
|
|
return { committed: commit.code === 0, error: commit.code !== 0 ? commit.out : undefined };
|
|
}
|
|
|
|
// Headless one-shot pi call (`pi -p --mode json <prompt>`), used only for
|
|
// quick diagnostic/mechanical calls that don't need to be a watchable
|
|
// session: resolver's crash/stall diagnosis, and the initial clone. Kept
|
|
// exactly as before -- only the interactive per-role stages (runStage,
|
|
// below) move to agent-manager.
|
|
function spawnPi({ agent, prompt, cwd }) {
|
|
const finalPrompt = ROLE_SKILLS.has(agent) ? `/skill:${agent} ${HARD_RULES}${prompt}` : prompt;
|
|
const args = ["-p", "--mode", "json"];
|
|
if (AGENT_PROVIDER) args.push("--provider", AGENT_PROVIDER);
|
|
if (AGENT_MODEL) args.push("--model", AGENT_MODEL);
|
|
args.push(finalPrompt);
|
|
const child = spawn(PI_BIN, args, { stdio: ["ignore", "pipe", "pipe"], cwd });
|
|
let lastText = "";
|
|
let stderrTail = "";
|
|
let buf = "";
|
|
child.stdout.on("data", (chunk) => {
|
|
buf += chunk;
|
|
let idx;
|
|
while ((idx = buf.indexOf("\n")) !== -1) {
|
|
const line = buf.slice(0, idx);
|
|
buf = buf.slice(idx + 1);
|
|
if (!line.trim()) continue;
|
|
try {
|
|
const event = JSON.parse(line);
|
|
if (event.type === "message_end" && event.message && Array.isArray(event.message.content)) {
|
|
const text = event.message.content
|
|
.filter((c) => c.type === "text")
|
|
.map((c) => c.text)
|
|
.join("\n");
|
|
if (text) lastText = text;
|
|
}
|
|
} catch {
|
|
// non-JSON stdout noise, ignore
|
|
}
|
|
}
|
|
});
|
|
child.stderr.on("data", (chunk) => {
|
|
process.stderr.write(chunk);
|
|
stderrTail = (stderrTail + chunk.toString()).slice(-4000);
|
|
});
|
|
return new Promise((resolve) => {
|
|
child.on("close", (code) => resolve({ code, lastText, stderrTail }));
|
|
});
|
|
}
|
|
|
|
async function askResolver(cwd, task, diagnosticPrompt) {
|
|
const result = await spawnPi({ agent: "resolver", prompt: diagnosticPrompt, cwd });
|
|
return parseVerdictLine(result.lastText, "RESOLUTION");
|
|
}
|
|
|
|
function taskSessionLabel(task) {
|
|
return `task-${task.replace(/[^a-zA-Z0-9]/g, "-")}`;
|
|
}
|
|
|
|
// POSIX single-quote wrap, for embedding a literal `agent-manager spawn`
|
|
// command inside another stage's prompt -- the same escaping tmux.ShellQuote
|
|
// does on the Go side of the fork.
|
|
function shQuote(s) {
|
|
return "'" + String(s).replace(/'/g, `'\\''`) + "'";
|
|
}
|
|
|
|
// The literal shell command an agent runs, as its own last action, to spawn
|
|
// its successor stage itself instead of the coordinator spawning it. Always
|
|
// applies HARD_RULES to the embedded prompt -- the single place that
|
|
// happens for agent-issued spawns, mirroring spawnRole's HARD_RULES + prompt
|
|
// for coordinator-issued ones (see runTaskInteractive) so it's applied
|
|
// exactly once regardless of who calls agent-manager spawn.
|
|
function spawnCommandLine(cwd, task, stageLabel, prompt) {
|
|
const label = `${taskSessionLabel(task)}-${stageLabel}`;
|
|
const parts = [AGENT_MANAGER_BIN, "spawn", "--tool", "pi", "--cwd", cwd, "--name", label, "--prompt", HARD_RULES + prompt];
|
|
const { provider, model } = providerModelFor(stageLabel);
|
|
if (provider) parts.push("--provider", provider);
|
|
if (model) parts.push("--model", model);
|
|
return parts.map(shQuote).join(" ");
|
|
}
|
|
|
|
// Instruction builders for the self-chained happy path: planner spawns
|
|
// investigator spawns implementer spawns judge, each as its own last
|
|
// action. Built bottom-up (judge first) since each level embeds the next
|
|
// level's full spawn command as literal text. Judge never self-spawns --
|
|
// the coordinator reads its verdict and decides PASS/retry/revise itself
|
|
// (see runTaskInteractive), so retries and plan revisions stay a
|
|
// deterministic, hard-capped branch in code, not something an agent
|
|
// decides on its own.
|
|
function judgeInstructions(task, baseBranch, attempt) {
|
|
return (
|
|
`Use the judge skill to review the diff against ${baseBranch}...HEAD. Write your verdict to ` +
|
|
`.task-result-${task} as a single "VERDICT: PASS" or "VERDICT: FAIL" line plus one line of ` +
|
|
`rationale, then run: touch .stage-done-${task}-judge-${attempt}`
|
|
);
|
|
}
|
|
|
|
function implementerInstructions(cwd, task, baseBranch, attempt, feedbackHint) {
|
|
const judgeSpawn = spawnCommandLine(cwd, task, "judge", judgeInstructions(task, baseBranch, attempt));
|
|
return (
|
|
`Use the implementer skill to implement what the current PLAN.md specifies (commit as you go). ${feedbackHint} ` +
|
|
`When your implementation is committed, spawn the judge stage yourself by running exactly this command: ${judgeSpawn}\n\n` +
|
|
`Then run: touch .stage-done-${task}-implementer-${attempt}`
|
|
);
|
|
}
|
|
|
|
function investigatorInstructions(cwd, task, baseBranch) {
|
|
const implSpawn = spawnCommandLine(cwd, task, "implementer", implementerInstructions(cwd, task, baseBranch, 1, ""));
|
|
return (
|
|
`Use the investigator skill to confirm PLAN.md against real sources, append findings, commit. ` +
|
|
`When finished, spawn the implementer stage yourself by running exactly this command: ${implSpawn}\n\n` +
|
|
`Then run: touch .stage-done-${task}-investigator`
|
|
);
|
|
}
|
|
|
|
function plannerInstructions(cwd, task, baseBranch, specHint) {
|
|
const invSpawn = spawnCommandLine(cwd, task, "investigator", investigatorInstructions(cwd, task, baseBranch));
|
|
return (
|
|
`Use the planner skill to draft PLAN.md for task ${task}, reading its spec (${specHint}). Commit PLAN.md. ` +
|
|
`When finished, spawn the investigator stage yourself by running exactly this command: ${invSpawn}\n\n` +
|
|
`Then run: touch .stage-done-${task}-planner`
|
|
);
|
|
}
|
|
|
|
async function runConcurrent(items, limit, worker) {
|
|
const results = new Array(items.length);
|
|
let i = 0;
|
|
async function next() {
|
|
while (i < items.length) {
|
|
const idx = i++;
|
|
results[idx] = await worker(items[idx], idx);
|
|
}
|
|
}
|
|
await Promise.all(Array.from({ length: Math.min(limit, items.length) }, next));
|
|
return results;
|
|
}
|
|
|
|
const STAGE_TIMEOUT_MS = 10 * 60 * 1000;
|
|
const NUDGE_TIMEOUT_MS = 5 * 60 * 1000;
|
|
|
|
// Resolves as soon as filePath appears (fs.watch on its directory), or
|
|
// after limitMs with no sign of it -- the "notify" half of the migration:
|
|
// event-driven completion instead of hub.js's old 10s poll.
|
|
function waitForSentinel(filePath, limitMs) {
|
|
return new Promise((resolve) => {
|
|
if (fs.existsSync(filePath)) return resolve(true);
|
|
const dir = path.dirname(filePath);
|
|
let settled = false;
|
|
let watcher;
|
|
const finish = (result) => {
|
|
if (settled) return;
|
|
settled = true;
|
|
clearTimeout(timer);
|
|
if (watcher) {
|
|
try {
|
|
watcher.close();
|
|
} catch {
|
|
// already closed
|
|
}
|
|
}
|
|
resolve(result);
|
|
};
|
|
try {
|
|
watcher = fs.watch(dir, () => {
|
|
if (fs.existsSync(filePath)) finish(true);
|
|
});
|
|
} catch {
|
|
// dir missing at watch time is a real bug elsewhere (cwd should
|
|
// already exist); surface it as a timeout rather than hang forever.
|
|
return finish(false);
|
|
}
|
|
// Closes the race between the existsSync check above and the watcher
|
|
// actually being attached.
|
|
if (fs.existsSync(filePath)) return finish(true);
|
|
const timer = setTimeout(() => finish(false), limitMs);
|
|
});
|
|
}
|
|
|
|
// Runs one role as its own fresh interactive pi session, spawned as a real
|
|
// agent-manager-tracked session (see AGENT_MANAGER_BIN above) instead of a
|
|
// raw tmux pane. Planner, investigator, implementer, judge, and
|
|
// phase-judge are separate agents with separate context, coordinating only
|
|
// through what's on disk in the task's worktree -- unchanged. The session
|
|
// is never killed here: it rests at an idle prompt once its sentinel file
|
|
// lands, staying attachable in agent-manager for as long as the user wants.
|
|
//
|
|
// Used for stages the coordinator itself spawns: the very first stage of a
|
|
// task (planner), implementer retries and planner-revise (both are the
|
|
// coordinator's own branch decision after reading judge's verdict -- see
|
|
// runTaskInteractive), and phase-judge. Stages self-spawned by their
|
|
// predecessor (see spawnCommandLine and the instruction builders above) go
|
|
// through waitForStage below instead, since the coordinator never issued
|
|
// the spawn itself.
|
|
async function spawnAndWaitStage(cwd, task, stageLabel, stagePrompt, sentinelFile) {
|
|
const label = `${taskSessionLabel(task)}-${stageLabel}`;
|
|
fs.rmSync(sentinelFile, { force: true });
|
|
|
|
const spawnArgs = ["spawn", "--tool", "pi", "--cwd", cwd, "--name", label, "--prompt", stagePrompt];
|
|
const { provider, model } = providerModelFor(stageLabel);
|
|
if (provider) spawnArgs.push("--provider", provider);
|
|
if (model) spawnArgs.push("--model", model);
|
|
const spawned = await runCmd(AGENT_MANAGER_BIN, spawnArgs);
|
|
if (spawned.code !== 0) return { ok: false, crashed: true, error: spawned.out, sessionName: label };
|
|
const sessionId = spawned.out;
|
|
const target = amSessionName(sessionId);
|
|
|
|
let ok = await waitForSentinel(sentinelFile, STAGE_TIMEOUT_MS);
|
|
if (!ok) {
|
|
const pane = await runAmTmux(["capture-pane", "-t", target, "-p", "-S", "-200"]);
|
|
const resolution = await askResolver(
|
|
cwd,
|
|
task,
|
|
`Task ${task}'s "${stageLabel}" stage hasn't finished after 10 minutes. Its pane tail:\n${pane.out.slice(-3000)}\n\n` +
|
|
`Decide: is it still making real progress and worth nudging to wrap up, or stuck and worth abandoning?`
|
|
);
|
|
if (resolution === "RETRY") {
|
|
await runAmTmux([
|
|
"send-keys",
|
|
"-t",
|
|
target,
|
|
`Please wrap up the "${stageLabel}" stage now and touch ${path.basename(sentinelFile)} when done.`,
|
|
"Enter",
|
|
]);
|
|
ok = await waitForSentinel(sentinelFile, NUDGE_TIMEOUT_MS);
|
|
}
|
|
}
|
|
return { ok, sessionName: label, sessionId };
|
|
}
|
|
|
|
// Waits for a stage that spawned itself (its predecessor's prompt embedded
|
|
// the agent-manager spawn command -- see spawnCommandLine). No session id
|
|
// is available here for pane capture on a timeout, since the coordinator
|
|
// never issued the spawn; the resolver gets a diagnostic prompt pointing at
|
|
// the session name to attach to directly instead, and RETRY just gives it
|
|
// the nudge window rather than an actual send-keys nudge.
|
|
async function waitForStage(cwd, task, stageLabel, sentinelFile) {
|
|
const label = `${taskSessionLabel(task)}-${stageLabel}`;
|
|
let ok = await waitForSentinel(sentinelFile, STAGE_TIMEOUT_MS);
|
|
if (!ok) {
|
|
const resolution = await askResolver(
|
|
cwd,
|
|
task,
|
|
`Task ${task}'s "${stageLabel}" stage hasn't finished after 10 minutes. It was self-spawned by its ` +
|
|
`predecessor agent -- attach to its agent-manager session ("${label}") directly to inspect progress; no ` +
|
|
`pane tail is available here. Decide: is it still making real progress and worth waiting longer, or stuck ` +
|
|
`and worth abandoning?`
|
|
);
|
|
if (resolution === "RETRY") {
|
|
ok = await waitForSentinel(sentinelFile, NUDGE_TIMEOUT_MS);
|
|
}
|
|
}
|
|
return { ok, sessionName: label };
|
|
}
|
|
|
|
const MAX_IMPLEMENT_ATTEMPTS = 5;
|
|
const MAX_PLAN_REVISIONS = 3;
|
|
|
|
// Runs one task in its own git worktree (see runPhase): planner drafts
|
|
// PLAN.md, investigator confirms it, then implementer and judge go back and
|
|
// forth -- judge's FAIL rationale lands in .task-result-<task>, which the
|
|
// next implementer attempt is told to read and address. After
|
|
// MAX_IMPLEMENT_ATTEMPTS straight fails, planner is brought back in to
|
|
// judge whether the *plan* itself is wrong, not just the implementation; if
|
|
// so it revises PLAN.md and the implementer gets a fresh attempt budget
|
|
// against the new plan. MAX_PLAN_REVISIONS caps this from looping forever
|
|
// on a task that's genuinely stuck.
|
|
async function runTaskInteractive(cwd, baseBranch, task, pipelineSession, judgeOnly) {
|
|
const resultFile = path.join(cwd, `.task-result-${task}`);
|
|
fs.rmSync(resultFile, { force: true });
|
|
|
|
const specHint = `the file under tasks/ starting with "${task}-"`;
|
|
|
|
// spawnRole: coordinator issues the spawn itself (planner's first call,
|
|
// implementer retries, planner-revise, the judgeOnly quick check).
|
|
// waitRole: the stage was self-spawned by its predecessor (see
|
|
// spawnCommandLine/instruction builders above) -- the coordinator only
|
|
// waits. Both still run the commitPending backstop sweep afterward and
|
|
// update pipelineSession's progress display, unchanged from before.
|
|
const spawnRole = async (stageLabel, prompt, sentinel) => {
|
|
pipelineSession.activeTasks[task] = { stage: stageLabel, startedAt: new Date().toISOString() };
|
|
logProgress(pipelineSession);
|
|
const result = await spawnAndWaitStage(cwd, task, stageLabel, HARD_RULES + prompt, sentinel);
|
|
await commitPending(cwd, `task: ${task} (${stageLabel})`);
|
|
return result;
|
|
};
|
|
|
|
const waitRole = async (stageLabel, sentinel) => {
|
|
pipelineSession.activeTasks[task] = { stage: stageLabel, startedAt: new Date().toISOString() };
|
|
logProgress(pipelineSession);
|
|
const result = await waitForStage(cwd, task, stageLabel, sentinel);
|
|
await commitPending(cwd, `task: ${task} (${stageLabel})`);
|
|
return result;
|
|
};
|
|
|
|
const abandon = (stageLabel, result, attempt) => {
|
|
delete pipelineSession.activeTasks[task];
|
|
logProgress(pipelineSession);
|
|
return {
|
|
task,
|
|
status: result.crashed ? "spawn-crashed" : "timed-out",
|
|
error: result.error,
|
|
stoppedAt: stageLabel,
|
|
...(attempt !== undefined ? { attempt } : {}),
|
|
};
|
|
};
|
|
|
|
if (judgeOnly) {
|
|
const quick = await spawnRole(
|
|
"judge",
|
|
`Task ${task} may already be implemented on this branch -- check ` +
|
|
`\`git log --oneline --grep '${task}'\` and the current code directly against its spec ` +
|
|
`(${specHint})'s acceptance criteria (no PLAN.md exists for this task yet). Write your ` +
|
|
`verdict to .task-result-${task} as a single "VERDICT: PASS" or "VERDICT: FAIL" line plus ` +
|
|
`one line of rationale, then run: touch .stage-done-${task}-judge-0`,
|
|
path.join(cwd, `.stage-done-${task}-judge-0`)
|
|
);
|
|
if (!quick.ok) return abandon("judge", quick, 0);
|
|
|
|
const quickText = fs.existsSync(resultFile) ? fs.readFileSync(resultFile, "utf8") : "";
|
|
if (parseVerdictLine(quickText, "VERDICT") === "PASS") {
|
|
delete pipelineSession.activeTasks[task];
|
|
logProgress(pipelineSession);
|
|
return { task, status: "done", judgeRationale: quickText, judgeOnlyPass: true };
|
|
}
|
|
}
|
|
|
|
// Happy path is self-chained: planner spawns investigator spawns
|
|
// implementer (attempt 1) spawns judge, each as its own last action (see
|
|
// the instruction builders above). The coordinator spawns only planner
|
|
// here, then just waits for each stage's sentinel in turn -- same
|
|
// per-stage timeout/resolver/commit backstop as before, just not the one
|
|
// issuing the spawn for the self-chained stages.
|
|
let result = await spawnRole("planner", plannerInstructions(cwd, task, baseBranch, specHint), path.join(cwd, `.stage-done-${task}-planner`));
|
|
if (!result.ok) return abandon("planner", result);
|
|
|
|
result = await waitRole("investigator", path.join(cwd, `.stage-done-${task}-investigator`));
|
|
if (!result.ok) return abandon("investigator", result);
|
|
|
|
let planRevisions = 0;
|
|
let implementAttempt = 0;
|
|
let verdict = null;
|
|
let resultText = "";
|
|
let justRevisedPlan = false;
|
|
// Attempt 1's implementer was already self-spawned by investigator above
|
|
// (its exact prompt was pre-baked into investigatorInstructions with
|
|
// attempt=1 and an empty feedbackHint, matching what attempt 1 always
|
|
// computes below) -- the coordinator only waits for it. Every retry
|
|
// after that is the coordinator's own branch decision on judge's
|
|
// verdict, so it spawns implementer directly instead.
|
|
let implementerSelfChained = true;
|
|
|
|
while (true) {
|
|
implementAttempt++;
|
|
const feedbackHint = fs.existsSync(resultFile)
|
|
? justRevisedPlan
|
|
? `.task-result-${task} holds the judge's feedback against the OLD plan, which prompted a plan revision -- ` +
|
|
`PLAN.md has since changed. Read the current PLAN.md as the source of truth, not the old feedback verbatim.`
|
|
: `A previous judge review exists at .task-result-${task} -- read it and address every issue it raises.`
|
|
: "";
|
|
justRevisedPlan = false;
|
|
|
|
if (implementerSelfChained) {
|
|
result = await waitRole("implementer", path.join(cwd, `.stage-done-${task}-implementer-${implementAttempt}`));
|
|
} else {
|
|
result = await spawnRole(
|
|
"implementer",
|
|
implementerInstructions(cwd, task, baseBranch, implementAttempt, feedbackHint),
|
|
path.join(cwd, `.stage-done-${task}-implementer-${implementAttempt}`)
|
|
);
|
|
}
|
|
implementerSelfChained = false;
|
|
if (!result.ok) return abandon("implementer", result, implementAttempt);
|
|
|
|
// Judge is always self-spawned by implementer, coordinator- or
|
|
// self-chained alike -- implementerInstructions embeds the same spawn
|
|
// command either way.
|
|
result = await waitRole("judge", path.join(cwd, `.stage-done-${task}-judge-${implementAttempt}`));
|
|
if (!result.ok) return abandon("judge", result, implementAttempt);
|
|
|
|
resultText = fs.existsSync(resultFile) ? fs.readFileSync(resultFile, "utf8") : "";
|
|
verdict = parseVerdictLine(resultText, "VERDICT");
|
|
if (verdict === "PASS") break;
|
|
|
|
if (implementAttempt >= MAX_IMPLEMENT_ATTEMPTS) {
|
|
if (planRevisions >= MAX_PLAN_REVISIONS) break;
|
|
planRevisions++;
|
|
result = await spawnRole(
|
|
"planner-revise",
|
|
`Implementer failed judge review ${MAX_IMPLEMENT_ATTEMPTS} times in a row for task ${task}. Read PLAN.md, ` +
|
|
`the judge's feedback in .task-result-${task}, and the current diff against ${baseBranch}...HEAD. Decide ` +
|
|
`whether the plan's approach itself is wrong, not just the implementation -- if so, revise PLAN.md and ` +
|
|
`commit. If you change the approach, also use the investigator skill to confirm the new approach against ` +
|
|
`real sources before committing. If the plan is sound, note why in PLAN.md and leave it as-is. Then run: ` +
|
|
`touch .stage-done-${task}-planner-revise-${planRevisions}`,
|
|
path.join(cwd, `.stage-done-${task}-planner-revise-${planRevisions}`)
|
|
);
|
|
if (!result.ok) return abandon("planner-revise", result, planRevisions);
|
|
implementAttempt = 0;
|
|
justRevisedPlan = true;
|
|
}
|
|
}
|
|
|
|
delete pipelineSession.activeTasks[task];
|
|
logProgress(pipelineSession);
|
|
|
|
if (verdict !== "PASS" && planRevisions >= MAX_PLAN_REVISIONS) {
|
|
return { task, status: "unresolved", judgeRationale: resultText, implementAttempts: implementAttempt, planRevisions };
|
|
}
|
|
return { task, status: verdict === "PASS" ? "done" : "done-with-concerns", judgeRationale: resultText };
|
|
}
|
|
|
|
// How many tasks' independent chains can be mid-flight at once within a
|
|
// phase. This is the real buffer against a single-instance backend: a
|
|
// lock around agent-manager spawn wouldn't work (spawn is fire-and-forget
|
|
// -- it creates the tmux session and returns immediately, the actual LLM
|
|
// call happens later, asynchronously, inside that detached session, so a
|
|
// lock held only for the spawn call releases before the call it's
|
|
// supposed to guard even starts). Ollama itself queues concurrent
|
|
// requests to one loaded model rather than erroring, but each one waits
|
|
// out the others -- PHASE_CONCURRENCY=1 makes that explicit: tasks run
|
|
// strictly one after another, one chain ever active against the backend
|
|
// at a time. Bump it back up once there's more than one model instance.
|
|
const PHASE_CONCURRENCY = Number(process.env.PHASE_CONCURRENCY) || 3;
|
|
|
|
// Runs one phase (a batch of tasks with no declared dependency on each
|
|
// other) with up to PHASE_CONCURRENCY tasks in flight at once. Each task
|
|
// gets its own git worktree off workBranch -- concurrent pi sessions writing
|
|
// into one shared working tree would corrupt the index; worktrees share the
|
|
// same object database but give each task an isolated checkout. After a
|
|
// task's session ends, its branch is merged back into workBranch and pushed,
|
|
// one merge at a time (git ref updates aren't safe to run concurrently even
|
|
// though the worktrees themselves are isolated).
|
|
async function runPhase(pipelineId, cwd, workBranch, phaseTasks, pipelineSession) {
|
|
const entries = phaseTasks.map((t) => (typeof t === "string" ? { id: t, judgeOnly: false } : t));
|
|
|
|
const worktrees = {};
|
|
for (const entry of entries) {
|
|
const task = entry.id;
|
|
const wtDir = path.join(WORK_DIR, pipelineId, `wt-${task.replace(/[^a-zA-Z0-9]/g, "-")}`);
|
|
const taskBranch = `task/${task}`;
|
|
const add = await runGit(cwd, ["worktree", "add", "-b", taskBranch, wtDir, workBranch]);
|
|
if (add.code !== 0) {
|
|
pipelineSession.taskResults.push({ task, status: "worktree-crashed", error: add.out });
|
|
continue;
|
|
}
|
|
worktrees[task] = { wtDir, taskBranch };
|
|
}
|
|
|
|
const runnable = entries.filter((e) => worktrees[e.id]);
|
|
await runConcurrent(runnable, PHASE_CONCURRENCY, async (entry) => {
|
|
const { wtDir } = worktrees[entry.id];
|
|
const result = await runTaskInteractive(wtDir, workBranch, entry.id, pipelineSession, entry.judgeOnly);
|
|
pipelineSession.taskResults.push(result);
|
|
return result;
|
|
});
|
|
|
|
for (const entry of runnable) {
|
|
const task = entry.id;
|
|
const { wtDir, taskBranch } = worktrees[task];
|
|
const result = pipelineSession.taskResults.find((r) => r.task === task);
|
|
|
|
const merge = await runGit(cwd, ["merge", "--no-ff", taskBranch, "-m", `merge: ${task}`]);
|
|
if (merge.code !== 0) {
|
|
await runGit(cwd, ["merge", "--abort"]);
|
|
if (result) {
|
|
result.status = "merge-conflict";
|
|
result.mergeError = merge.out;
|
|
}
|
|
} else {
|
|
const push = await runGit(cwd, ["push", "-u", "origin", workBranch]);
|
|
if (result) {
|
|
result.pushed = push.code === 0;
|
|
if (!result.pushed) result.pushError = push.out;
|
|
}
|
|
}
|
|
|
|
await runGit(cwd, ["worktree", "remove", wtDir, "--force"]);
|
|
await runGit(cwd, ["branch", "-D", taskBranch]);
|
|
logProgress(pipelineSession);
|
|
}
|
|
}
|
|
|
|
// Discovers phases/tasks from the repo's own tasks/INDEX.md instead of
|
|
// requiring the caller to pass --tasks. Matches this convention's board
|
|
// shape (see e.g. Poimen/agent-rust's tasks/INDEX.md): a numbered phase
|
|
// heading ("## 1 — Foundations · T0.x"), followed by a markdown table
|
|
// whose rows link to each task's own spec file ("| [T0.1](T0.1-....md) |
|
|
// ... |"). Headings that aren't a numbered phase (prose sections like
|
|
// "## Ordering — declared, never derived", "## Progress") are skipped --
|
|
// only "## <digits> — ..." starts a new phase. Returns null if
|
|
// tasks/INDEX.md doesn't exist; an empty array if it exists but no phase
|
|
// yielded any task rows.
|
|
function parseTaskBoard(cwd) {
|
|
const indexPath = path.join(cwd, "tasks", "INDEX.md");
|
|
if (!fs.existsSync(indexPath)) return null;
|
|
|
|
const phaseHeaderRe = /^##\s+\d+\s+—/;
|
|
const taskRowRe = /^\|\s*\[([A-Za-z0-9.]+)\]\(/;
|
|
|
|
const phases = [];
|
|
let current = null;
|
|
for (const line of fs.readFileSync(indexPath, "utf8").split("\n")) {
|
|
if (phaseHeaderRe.test(line)) {
|
|
current = [];
|
|
phases.push(current);
|
|
continue;
|
|
}
|
|
const m = line.match(taskRowRe);
|
|
if (m && current) current.push(m[1]);
|
|
}
|
|
return phases.filter((phase) => phase.length > 0);
|
|
}
|
|
|
|
function phaseLabelFor(phaseTasks, index) {
|
|
const first = phaseTasks[0];
|
|
const id = typeof first === "string" ? first : first.id;
|
|
const dot = id.indexOf(".");
|
|
return dot === -1 ? `phase-${index}` : id.slice(0, dot);
|
|
}
|
|
|
|
function logProgress(pipelineSession) {
|
|
console.log(`[pipeline ${pipelineSession.id}] ${JSON.stringify(pipelineSession)}`);
|
|
}
|
|
|
|
// tasks: array of phases, each phase an array of task ids with no declared
|
|
// dependency on each other (e.g. [["T0.1","T0.2"], ["T1.1","T1.2","T1.3"]]).
|
|
// A flat array of ids is also accepted and treated as one single phase.
|
|
// Phases run strictly sequentially; tasks within a phase run concurrently,
|
|
// each in its own worktree -- see runPhase. Each phase gets its own branch
|
|
// (agent-run/<id>/<phaseLabel>, e.g. .../T1); once every task in that
|
|
// phase lands "done" or "done-with-concerns" AND the phase judge passes
|
|
// the integration review, the phase branch is squash-merged into
|
|
// baseBranch and pushed, then the next phase branches off that updated
|
|
// base. Any failure halts the pipeline before merging.
|
|
async function runPipeline({ pipelineId, repo, baseBranch, tasks, branchName }) {
|
|
const cwd = path.join(WORK_DIR, pipelineId);
|
|
fs.mkdirSync(cwd, { recursive: true });
|
|
|
|
const pipelineSession = {
|
|
id: pipelineId,
|
|
status: "running",
|
|
taskResults: [],
|
|
activeTasks: {},
|
|
totalTasks: 0,
|
|
startedAt: new Date().toISOString(),
|
|
};
|
|
|
|
const finish = (status) => {
|
|
pipelineSession.status = status;
|
|
pipelineSession.endedAt = new Date().toISOString();
|
|
logProgress(pipelineSession);
|
|
return pipelineSession;
|
|
};
|
|
|
|
// Deterministic, not routed through an LLM -- clone is 100% mechanical
|
|
// (same reasoning as commitPending/the squash-merge sequence below), and
|
|
// was the one place left that broke that pattern: a headless spawnPi
|
|
// call here meant a crash gave zero diagnostic output, just a silent
|
|
// exit code with nothing to debug from.
|
|
const clone = await runGit(cwd, ["clone", "--branch", baseBranch, repo, "."]);
|
|
if (clone.code !== 0) {
|
|
pipelineSession.gitError = clone.out;
|
|
return finish("clone-crashed");
|
|
}
|
|
if (!fs.existsSync(path.join(cwd, ".git"))) return finish("clone-missing");
|
|
|
|
// No --tasks given: discover the phase/task board from the repo's own
|
|
// tasks/INDEX.md (see parseTaskBoard) instead of requiring the caller to
|
|
// already know it. Each discovered task runs judgeOnly first -- a cheap
|
|
// "is this already done on this branch" check against the board's own
|
|
// possibly-stale checkmarks, falling through to the full planner/
|
|
// investigator/implementer/judge flow only when it isn't.
|
|
let phases = tasks ? (Array.isArray(tasks[0]) ? tasks : [tasks]) : parseTaskBoard(cwd);
|
|
if (!phases || phases.length === 0) {
|
|
pipelineSession.gitError = "no --tasks given and tasks/INDEX.md not found or empty";
|
|
return finish("no-tasks-found");
|
|
}
|
|
if (!tasks) {
|
|
phases = phases.map((phase) => phase.map((id) => ({ id, judgeOnly: true })));
|
|
}
|
|
pipelineSession.totalTasks = phases.flat().length;
|
|
|
|
for (let i = 0; i < phases.length; i++) {
|
|
const phaseTasks = phases[i];
|
|
const phaseLabel = phaseLabelFor(phaseTasks, i);
|
|
const phaseBranch = branchName ? `${branchName}/${phaseLabel}` : `agent-run/${pipelineId}/${phaseLabel}`;
|
|
|
|
const branchResult = await runGit(cwd, ["checkout", "-b", phaseBranch]);
|
|
if (branchResult.code !== 0) {
|
|
pipelineSession.gitError = branchResult.out;
|
|
return finish("branch-crashed");
|
|
}
|
|
|
|
if (i === 0) {
|
|
const gitignoreAdditions = [
|
|
"",
|
|
"# agent-harness: build artifacts and vendored archives never belong in source control",
|
|
"*.tar.gz",
|
|
"*.tgz",
|
|
"*.crate",
|
|
"*.zip",
|
|
"*.bin",
|
|
"*.whl",
|
|
"vendor/",
|
|
"node_modules/",
|
|
"",
|
|
"# agent-harness: task/phase completion sentinel files, harness bookkeeping only",
|
|
".task-result-*",
|
|
".phase-result-*",
|
|
".stage-done-*",
|
|
].join("\n");
|
|
fs.appendFileSync(path.join(cwd, ".gitignore"), gitignoreAdditions + "\n");
|
|
await runGit(cwd, ["add", ".gitignore"]);
|
|
await runGit(cwd, ["commit", "-m", "chore: broaden .gitignore for agent-run artifacts"]);
|
|
}
|
|
|
|
await runPhase(pipelineId, cwd, phaseBranch, phaseTasks, pipelineSession);
|
|
|
|
const phaseTaskIds = new Set(phaseTasks.map((t) => (typeof t === "string" ? t : t.id)));
|
|
const phaseResults = pipelineSession.taskResults.filter((r) => phaseTaskIds.has(r.task));
|
|
const phaseClean =
|
|
phaseResults.length === phaseTaskIds.size && phaseResults.every((r) => r.status === "done" || r.status === "done-with-concerns");
|
|
|
|
if (!phaseClean) {
|
|
pipelineSession.haltedAt = phaseLabel;
|
|
return finish("halted-phase-failed");
|
|
}
|
|
|
|
const phaseResultFile = path.join(cwd, `.phase-result-${phaseLabel}`);
|
|
fs.rmSync(phaseResultFile, { force: true });
|
|
const phaseSentinel = path.join(cwd, `.stage-done-phase-${phaseLabel}-judge`);
|
|
const phaseJudge = await spawnAndWaitStage(
|
|
cwd,
|
|
`phase-${phaseLabel}`,
|
|
"judge",
|
|
HARD_RULES +
|
|
`Use the judge skill to review the full phase diff for phase ${phaseLabel} against ` +
|
|
`${baseBranch}...HEAD (covers every task in this phase: ${[...phaseTaskIds].join(", ")}). Every ` +
|
|
`individual task already passed its own judge review -- your job here is different: confirm the ` +
|
|
`tasks integrate correctly as one coherent narrative, and that real integration tests (not just ` +
|
|
`each task's isolated unit checks) exist and actually exercise the phase's intended use case end ` +
|
|
`to end. Write your verdict to .phase-result-${phaseLabel} as a single "VERDICT: PASS" or ` +
|
|
`"VERDICT: FAIL" line plus rationale, then run: touch .stage-done-phase-${phaseLabel}-judge`,
|
|
phaseSentinel
|
|
);
|
|
await commitPending(cwd, `phase: ${phaseLabel} integration review`);
|
|
|
|
if (!phaseJudge.ok) {
|
|
pipelineSession.haltedAt = phaseLabel;
|
|
pipelineSession.gitError = phaseJudge.error;
|
|
return finish("phase-judge-crashed");
|
|
}
|
|
|
|
const phaseJudgeText = fs.existsSync(phaseResultFile) ? fs.readFileSync(phaseResultFile, "utf8") : "";
|
|
if (parseVerdictLine(phaseJudgeText, "VERDICT") !== "PASS") {
|
|
pipelineSession.haltedAt = phaseLabel;
|
|
pipelineSession.phaseJudgeRationale = phaseJudgeText;
|
|
return finish("halted-phase-judge-failed");
|
|
}
|
|
|
|
const checkoutBase = await runGit(cwd, ["checkout", baseBranch]);
|
|
if (checkoutBase.code !== 0) {
|
|
pipelineSession.gitError = checkoutBase.out;
|
|
return finish("squash-crashed");
|
|
}
|
|
const squash = await runGit(cwd, ["merge", "--squash", phaseBranch]);
|
|
if (squash.code !== 0) {
|
|
await runGit(cwd, ["merge", "--abort"]);
|
|
pipelineSession.gitError = squash.out;
|
|
return finish("squash-crashed");
|
|
}
|
|
const commit = await runGit(cwd, ["commit", "-m", `feat: ${phaseLabel} (${[...phaseTaskIds].join(", ")})`]);
|
|
if (commit.code !== 0) {
|
|
pipelineSession.gitError = commit.out;
|
|
return finish("squash-crashed");
|
|
}
|
|
const push = await runGit(cwd, ["push", "origin", baseBranch]);
|
|
if (push.code !== 0) {
|
|
pipelineSession.gitError = push.out;
|
|
return finish("squash-push-crashed");
|
|
}
|
|
logProgress(pipelineSession);
|
|
}
|
|
|
|
return finish("completed");
|
|
}
|
|
|
|
function parseArgs(argv) {
|
|
const opts = { base: "main" };
|
|
for (let i = 0; i < argv.length; i++) {
|
|
const a = argv[i];
|
|
if (a === "--repo") opts.repo = argv[++i];
|
|
else if (a === "--base") opts.base = argv[++i];
|
|
else if (a === "--tasks") opts.tasks = argv[++i];
|
|
else if (a === "--branch") opts.branch = argv[++i];
|
|
}
|
|
return opts;
|
|
}
|
|
|
|
async function main() {
|
|
const opts = parseArgs(process.argv.slice(2));
|
|
if (!opts.repo) {
|
|
console.error(
|
|
"usage: coordinator.js --repo <url> [--tasks T0.1,T0.2;T1.1,T1.2,...] [--base main] [--branch <name>]\n" +
|
|
" --tasks omitted: discovers phases from the cloned repo's own tasks/INDEX.md"
|
|
);
|
|
// process.exitCode + natural exit, not process.exit() -- stdout piped
|
|
// through kubectl exec (not a TTY) can drop buffered console.log/
|
|
// console.error output if the process exits before it flushes. Setting
|
|
// exitCode and letting the event loop drain naturally is the
|
|
// documented-safe way to exit with a specific code without racing it.
|
|
process.exitCode = 1;
|
|
return;
|
|
}
|
|
const phases = opts.tasks ? opts.tasks.split(";").map((phase) => phase.split(",")) : null;
|
|
const pipelineId = require("node:crypto").randomUUID();
|
|
const result = await runPipeline({ pipelineId, repo: opts.repo, baseBranch: opts.base, tasks: phases, branchName: opts.branch });
|
|
process.exitCode = result.status === "completed" ? 0 : 1;
|
|
}
|
|
|
|
if (require.main === module) {
|
|
main();
|
|
}
|
|
|
|
module.exports = { runPipeline, spawnCommandLine, spawnAndWaitStage, waitForStage, parseTaskBoard };
|