task tool + fork-gate: put the "pi-task, not fork" rule where the decision is made

The rule was correct and written down twice (global AGENTS.md, this skill)
and was still violated by agents that had just read it: on 2026-09-17 all
five fork briefs in one session carried "do not", one returned confident
verbatim quotes that did not exist, and four had disjoint write boundaries
fork cannot enforce (re-run as pi-task, they passed). Three mechanisms,
none of them wording:

  1. fork is a TOOL — its self-recommending description ("implementation,
     testing, review…") is in the model's face every turn; pi-task was a CLI
     to be remembered and reached through bash with a hand-written JSON.
  2. The skill is gone after the first compaction; the tool list never is.
     The asymmetry widens in exactly the long sessions where fork is worst.
  3. Friction: one string vs a spec file + bash + reading result.json.

extensions/task.ts registers pi-task as the `task` tool. Flat parameters
build the spec; the decision rule sits in the description and in
promptGuidelines (appended to the system prompt, so compaction cannot remove
it). Before spending a model run it rejects the two spec errors that make a
boundary violation certain — write_allowed not an exact subset of roots, and
a writable root nested in a watched-only root (the parent's porcelain would
change every time; pi-task keys deltas by root string) — and it serialises
sibling tasks whose roots overlap (parallel siblings saw each other's writes
as violations, 2026-09-17). Returns the CLI's own parent-facing report; a
FAIL verdict is a result, only a CLI refusal is an error.

extensions/fork-gate.ts is a tool_call hook that BLOCKS a fork whose brief
contains a prohibition, a write boundary, or a clause-initial file-changing
imperative, and returns as the reason the exact task(...) to make instead.
Wording, not intent — the message says so and how to rephrase a genuinely
read-only brief. PI_FORK_GATE=off logs instead; /ext disables.

Evidence: test/fork-gate.test.mjs is two-sided (15 must-block incl. the real
shapes, 10 must-pass incl. "Write a summary…", "Report which files were
modified…", "Give me an update…"); the classifier redirects 5/5 of the real
briefs from the motivating session. test/task.test.mjs pins root overlap and
the pre-launch validation. Live in `pi -p`: the fork was intercepted before
any child spawned (no /tmp/pi-fork-* dir) and the model received the
redirect; task returned PASS with an evidence pointer and audit dir, a
budget-overrun returned a FAIL result (isError=false), and a nested-root spec
was rejected with no audit dir created.

skill/SKILL.md: Part 1 now opens with "decide the rung before the brief"
(the table, the three-question pre-flight, the roots contract, overlap,
what isolation does not fix); the ladder section and quick reference no
longer say pi-task "will never appear in your tool list". package.json gains
"type": "module" and a test script.
This commit is contained in:
2026-09-19 16:41:23 +02:00
parent 2610545c83
commit 25c1265681
8 changed files with 699 additions and 22 deletions
+139
View File
@@ -0,0 +1,139 @@
/**
* fork-gate — put the "pi-task, not fork" rule where the decision is made.
*
* WHY. `fork` is a registered tool: its description is in the model's face on
* every turn, and it recommends itself for "implementation". `pi-task` is a
* CLI with no tool description, reached through bash after remembering it
* exists. The prose rule that says which to use (global AGENTS.md, the
* pi-extensions skill) was correct and lost anyway — the skill is gone after
* the first compaction, the tool description never is. Measured on this fleet
* (2026-09-01/06 mbp-m1-2020, 2026-09-17 tor-ms22): forks with prohibitions in
* the brief ignored them, invented quotes, answered in the user's voice; the
* same briefs as pi-task runs passed their envelope. Root cause is by design:
* a fork child receives the WHOLE parent branch with the brief as the last
* message (pi-fork src/index.ts), so in a long session the narrative outweighs
* the instruction.
*
* WHAT. A `tool_call` hook on `fork` that BLOCKS when the brief carries the
* three things a fork cannot be trusted with — a prohibition, a write
* boundary, or an imperative to change files — and returns, as the block
* reason, the exact `task(...)` call to make instead. Fork stays available for
* what it is good at: read-only exploration that needs THIS conversation, and
* N parallel opinions on one question. A false positive costs one turn and
* teaches the rule in-context, which is the point; the message says how to
* rephrase a genuinely read-only brief.
*
* The classifier matches WORDING, not intent, and says so.
*
* OFF SWITCH. `PI_FORK_GATE=off` in the environment disables blocking for a
* session (the match is still logged to stderr); `/ext` (ext-toggle) disables
* the extension entirely. Neither is needed in normal operation.
*
* Companion: `task.ts` registers the `task` tool this gate points at. The
* gate does not require it — the reason text also gives the CLI form.
*/
import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
// ── Classifier (pure; exported so test/fork-gate.test.mjs can drive it) ──────
export interface GateRule {
/** Class of hazard, used in the block reason. */
kind: "prohibition" | "boundary" | "mutation";
pattern: RegExp;
/** One line: why a fork cannot be trusted with this. */
why: string;
}
export const GATE_RULES: GateRule[] = [
{
kind: "prohibition",
pattern:
/\b(do not|don'?t|must not|mustn'?t|never|not allowed|forbidden|prohibited|refrain from|under no circumstances|avoid (touching|editing|modifying|changing|writing))\b/i,
why: "a fork inherits your entire branch; the parent narrative outweighs a prohibition placed at the end of it",
},
{
kind: "boundary",
pattern:
/\b(only (touch|edit|modify|change|write|create|alter)|(touch|edit|modify|change|write) only|nothing else|no other (files?|dirs?|directories|paths?)|outside (of )?(this|these|that|the|its|your) (dir|directory|directories|file|files|folder|path|root|scope|repo|tree)|write[- ]?boundar(y|ies)|read[- ]only|stay (within|inside)|confined? to|limited to (the |these |those )?(\S+ )?(files?|dir|directory|directories|paths?|tree|repo))\b/i,
why: "fork has no boundary diff — a write outside the named set is invisible; pi-task diffs roots[] before and after",
},
{
// Clause-initial imperatives that change files. Deliberately excludes
// "write" and "create" (too often "write a summary"); the boundary and
// prohibition rules catch write tasks that phrase themselves carefully.
kind: "mutation",
pattern:
/(^|[.!?:;\n]\s*|\b(then|and|also|please|now)\s+)(commit|push|implement|refactor|rewrite|rename|delete|edit|modify|update|fix|apply|migrate|convert|replace|install|patch)\b/im,
why: "work that changes files needs a checkable PASS/FAIL and an audit trail; fork returns prose and deletes its own temp dir on exit",
},
];
export interface GateVerdict {
block: boolean;
kind?: GateRule["kind"];
matched?: string;
why?: string;
}
/** Decide whether a fork brief must be redirected to pi-task. */
export function classifyForkBrief(brief: unknown): GateVerdict {
if (typeof brief !== "string" || brief.length === 0) return { block: false };
for (const rule of GATE_RULES) {
const m = rule.pattern.exec(brief);
if (m) {
// Report the whole match trimmed, so the agent sees which words fired.
return { block: true, kind: rule.kind, matched: m[0].trim(), why: rule.why };
}
}
return { block: false };
}
/** The block reason the model reads. Exported so the test can pin its content. */
export function blockReason(v: GateVerdict): string {
return [
`fork BLOCKED by fork-gate — the brief contains a ${v.kind}: "${v.matched}".`,
`Why: ${v.why}.`,
"",
"Use the `task` tool instead (pi-task, L0–L2: isolated child that sees ONLY your spec,",
"immutable spec, PASS/FAIL envelope, write-boundary diff, audit trail):",
"",
' task(id="short-slug", goal="<verbatim task>", deliverable="<exact shape wanted>",',
' effort="fast|balanced|deep", read_only=false,',
' roots=["/abs/repo/docs", "/abs/repo/src"], # WATCHED, each diffed on its own',
' write_allowed=["/abs/repo/docs"], # CHANGEABLE: exact subset of roots; never nested in another root',
' facts=["<verified fact>"], files=["/abs/path/to/read"])',
"",
" CLI form if the tool is absent: /opt/pi-toolkit/bin/pi-task run <spec.json> (`pi-task schema` lists fields)",
" Tasks whose roots overlap run one after another, never in parallel (the tool serialises them).",
"",
"If this really is READ-ONLY exploration that needs this conversation's context, or N parallel",
"opinions on one question, drop the prohibition/mutation wording and call fork again — the gate",
"matches wording, not intent. If the brief needs the prohibition, it needs pi-task.",
].join("\n");
}
// ── Extension ────────────────────────────────────────────────────────────────
export default function (pi: ExtensionAPI) {
const mode = (process.env.PI_FORK_GATE ?? "block").toLowerCase();
pi.on("tool_call", (event, ctx) => {
if (event.toolName !== "fork") return;
const input = event.input as { task?: unknown } | undefined;
const verdict = classifyForkBrief(input?.task);
if (!verdict.block) return;
const reason = blockReason(verdict);
if (mode === "off") {
process.stderr.write(
`[fork-gate] PI_FORK_GATE=off — would have blocked fork (${verdict.kind}: "${verdict.matched}")\n`,
);
return;
}
if (ctx.hasUI) {
ctx.ui.notify(`fork-gate: blocked a fork (${verdict.kind}: "${verdict.matched}") → use task`, "warning");
}
return { block: true, reason };
});
}