bare-agent 0.22.0 → 0.23.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +1 -1
- package/bareagent.context.md +15 -2
- package/package.json +1 -1
- package/src/recurse.d.ts +80 -0
- package/src/recurse.js +166 -6
package/README.md
CHANGED
|
@@ -94,7 +94,7 @@ Every piece works alone — take what you need, ignore the rest. Two axes: **Act
|
|
|
94
94
|
|
|
95
95
|
### Recurse — break a hard task into a tree *(the RLM primitive)*
|
|
96
96
|
|
|
97
|
-
`recurse(task, ctx, opts)` does **decompose → fan-out → verify → synthesize** in one call — Recursive Language Models as a single import, composed *around* the Loop (never a new engine). The default is **model-driven**: the worker is handed a `spawn_child` tool and decides whether to split, bounded by depth + bareguard (no second guard layer). Forced fan-out (`count` / `mode:'fanout'`) and data-driven width (`mode:'partition'`, measured from a corpus) are opt-in. Give workers a stance with `opts.persona` (prepended to every worker, carries down the tree, deliberately kept out of the isolated verifier). The headline guarantee: **aggregation is code, never a model-stated number**, and a dead worker or exhausted guard returns an honest `{ incomplete, missingSlices }` — never a faked pass.
|
|
97
|
+
`recurse(task, ctx, opts)` does **decompose → fan-out → verify → synthesize** in one call — Recursive Language Models as a single import, composed *around* the Loop (never a new engine). The default is **model-driven**: the worker is handed a `spawn_child` tool and decides whether to split, bounded by depth + bareguard (no second guard layer). Forced fan-out (`count` / `mode:'fanout'`) and data-driven width (`mode:'partition'`, measured from a corpus) are opt-in. Give workers a stance with `opts.persona` (prepended to every worker, carries down the tree, deliberately kept out of the isolated verifier), and tell them *where they are* with `opts.context` (a read-only paths/cwd blob threaded to every worker so a sliced child can locate its artifact — facts, not a stance). For a leaf that should self-correct, pass `opts.refineLeaf` (opt-in): a definite leaf becomes a bounded generate→sense→regenerate loop driven by *your* deterministic sensor (test/compile/lint), feeding the gap back with escalating temperature. The headline guarantee: **aggregation is code, never a model-stated number**, and a dead worker or exhausted guard returns an honest `{ incomplete, missingSlices }` — never a faked pass.
|
|
98
98
|
|
|
99
99
|
Over a corpus, context reaches a worker as a **handle routed by question shape** (`opts.retrieval`):
|
|
100
100
|
|
package/bareagent.context.md
CHANGED
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
# bareagent — Integration Guide
|
|
2
2
|
|
|
3
3
|
> For AI assistants and developers wiring bareagent into a project.
|
|
4
|
-
> v0.
|
|
4
|
+
> v0.23.0 | Node.js >= 18 | zero required deps (`bareguard ^0.9.0` optional peer for governance) | Apache 2.0
|
|
5
5
|
>
|
|
6
6
|
> Full human guide with composition examples, design philosophy, and recipes: [Usage Guide](docs/02-features/usage-guide.md)
|
|
7
7
|
|
|
@@ -37,6 +37,8 @@ Eight entry points:
|
|
|
37
37
|
| Decompose a hard task into a verified tree (RLM) | recurse — decompose → fan-out → verify → synthesize in one call (**wire a gate**, cost is open by design) |
|
|
38
38
|
| Count / answer "how many / all" over a corpus, honestly | recurse(task, ctx, `{ corpus, retrieval: 'scan' }`) — scans every slice, CODE-counts |
|
|
39
39
|
| Give recurse workers a persona/role (senior-dev stance) | recurse(task, ctx, `{ persona }`) — prepended to every worker, carries down the tree; not applied to the verifier |
|
|
40
|
+
| Tell recurse workers WHERE they are (paths/cwd) so a slice can find its file | recurse(task, ctx, `{ context }`) — read-only blob on every worker's task message + the Planner + verifier; carries down (facts, not a stance) |
|
|
41
|
+
| Let a recurse LEAF retry its own failure with a deterministic check | recurse(task, ctx, `{ refineLeaf: { sensor } }`) — leaf becomes a bounded generate→sense→regenerate loop; your sensor (test/compile/lint), gap fed back, escalating temperature |
|
|
40
42
|
| Track task state (pending/running/done/failed) | StateMachine |
|
|
41
43
|
| Run agent turns on a schedule (cron, timers) | Scheduler |
|
|
42
44
|
| Require human approval before dangerous actions | Checkpoint |
|
|
@@ -669,7 +671,18 @@ const out = await recurse('Audit auth.js, billing.js, gateway.js for authz bugs'
|
|
|
669
671
|
});
|
|
670
672
|
```
|
|
671
673
|
|
|
672
|
-
**
|
|
674
|
+
**Worker context (`opts.context`, v0.23.0):** a read-only working-context string (paths/cwd) PREPENDED to every worker's TASK message as a `Working context:` block — so a sliced child can **locate its artifact** (the Planner paraphrases the goal into subtasks and drops absolute paths; without this, workers guess `.`/`~`/`/tmp` and get denied). Forwarded to the Planner as `info` (path-aware slices) and shown to the verifier too (neutral FACTS, not a stance — distinct from `persona`, which is a privileged SYSTEM-prompt stance). Carries down the tree. **Security:** it still becomes part of the prompt, so pass caller-trusted run-state only, never untrusted/end-user text (lower-privilege than `persona` — user message, not system — but still an injection surface). Absent ⇒ the task message is unchanged.
|
|
675
|
+
|
|
676
|
+
**Leaf self-correction (`opts.refineLeaf`, v0.23.0, opt-in):** turn a **definite leaf** (a node offered no `spawn_child` — `simple` tier or at `maxDepth`) into a bounded generate→sense→regenerate loop instead of a single pass: `{ sensor, maxIterations?, temperatures? }`. `sensor(result, { task, context, contract }) → Verdict` is YOUR **deterministic** close (test/compile/lint — not a model judge); on a non-pass its `critique` (the gap, not the transcript) is fed FRESH into the next attempt and the **retry temperature ESCALATES** (default `[0.2, 0.7, 1.0]` — load-bearing: a weak model at a flat temperature regenerates identical wrong code and ignores even crisp feedback). Each attempt is gate-checked + metered; a HaltError mid-loop → clean `{ incomplete }`; honest non-recovery → `receipts.refineLeaf.passed === false` (never a faked pass); `receipts.tokens` sums all attempts. The error-keyed `recall` stays YOUR tool (`opts.tools`), keyed off the fed-back critique — bareagent stays litectx-agnostic. Carries down (engages at the leaves). Absent ⇒ a leaf is a single pass.
|
|
677
|
+
|
|
678
|
+
```javascript
|
|
679
|
+
const out = await recurse('Fix the failing function in calc.js', ctx, {
|
|
680
|
+
context: `project root: ${process.cwd()}\nresolve relative paths against it`,
|
|
681
|
+
refineLeaf: { sensor: (code) => runTestsAndGrade(code) }, // your deterministic test/compile close
|
|
682
|
+
});
|
|
683
|
+
```
|
|
684
|
+
|
|
685
|
+
**What a delegated child inherits (important — the setpoint is the TOP node's job):** when a worker delegates with `spawn_child`, the child runs a **fresh `recurse`** that inherits `tools`, `synthesize`, `maxDepth`, `persona`, `context`, and `refineLeaf` — but the parent's **`contract`/`evaluate` are stripped** (and the forced `count`/`mode` + the corpus `retrieval` knobs). A slice is not graded against the *whole*-task definition-of-done (that verdict is the top node's, and a slice satisfying the whole DoD is the wrong question); only the top `recurse` verifies the synthesized result. The non-overridable `critical → force-verify` safety floor still fires per node (it keys on the task text, not the contract). So: set `contract`/`evaluate` once at the top; they do not — and should not — re-run per intermediate node.
|
|
673
686
|
|
|
674
687
|
## Wiring with Evaluator + refine (output-side verification)
|
|
675
688
|
|
package/package.json
CHANGED
package/src/recurse.d.ts
CHANGED
|
@@ -64,11 +64,51 @@ export type RecurseOptions = {
|
|
|
64
64
|
* and can override it (and any safety framing) for every worker in the tree. Caller-trusted input only.
|
|
65
65
|
*/
|
|
66
66
|
persona?: string | undefined;
|
|
67
|
+
/**
|
|
68
|
+
* - (BA-9 / relayfact F19) An optional caller-supplied READ-ONLY working-context
|
|
69
|
+
* blob (e.g. "project root: /abs/path\nfiles are relative to it") prepended to EVERY worker's task message as
|
|
70
|
+
* a `Working context:` block, so a sliced child can LOCATE its artifact — the concrete context (absolute
|
|
71
|
+
* paths / cwd) the Planner strips when it paraphrases the parent goal into child subtasks. CARRIES DOWN the
|
|
72
|
+
* tree (preserved by `forChild`, like `persona`) and, when forced fan-out plans, is forwarded as the Planner's
|
|
73
|
+
* `info` so the slices themselves are path-aware. Also shown to the verifier (neutral FACTS, not a stance, so
|
|
74
|
+
* no anti-sycophancy concern — and an agentic critic needs the path to exercise the artifact). Distinct from
|
|
75
|
+
* `persona`: persona is a privileged SYSTEM-prompt stance; context is run-state facts on the USER message.
|
|
76
|
+
* Absent/blank ⇒ byte-identical to pre-BA-9 (backward-compatible). Validated live (`poc/ba9-context-thread.mjs`:
|
|
77
|
+
* a weak model went 0/3 → 3/3 at locating an unguessable file once the root was threaded).
|
|
78
|
+
* **SECURITY:** this becomes part of every worker's prompt (and the verifier's) — lower-privilege than
|
|
79
|
+
* `persona` (the USER message, not the SYSTEM prompt) but still a prompt-injection surface. Intended for
|
|
80
|
+
* TRUSTED run-state (paths/cwd); do NOT pass untrusted / end-user-controlled text here.
|
|
81
|
+
*/
|
|
82
|
+
context?: string | undefined;
|
|
67
83
|
/**
|
|
68
84
|
* - Handle tools offered to EVERY worker (RC-5 pull-default: litectx
|
|
69
85
|
* `recall`/`get`, wired at build step 7). Workers query on demand; never the whole corpus.
|
|
70
86
|
*/
|
|
71
87
|
tools?: import("../types").ToolDef[] | undefined;
|
|
88
|
+
/**
|
|
89
|
+
* (Opt-in, BA-8 / relayfact F17) Turn a DEFINITE LEAF (a node that is offered no `spawn_child` — `simple`
|
|
90
|
+
* tier or at `maxDepth`) into a bounded generate→sense→regenerate loop instead of a single pass, so a failed
|
|
91
|
+
* slice can self-correct. `sensor` is a DETERMINISTIC close (test/compile/lint — NOT a model judge, R-S8) that
|
|
92
|
+
* returns a `Verdict`; on a non-pass its `critique` (the GAP, not the transcript) is fed FRESH into the next
|
|
93
|
+
* attempt (D6/A1 anti-anchoring) and the **retry temperature ESCALATES** (`temperatures`, default
|
|
94
|
+
* `[0.2,0.7,1.0]`) — the live-validated requirement that lets a weak model escape a repeat-the-same-mistake rut
|
|
95
|
+
* (`poc/ba8-leaf-refine.mjs`: 0/5 → 2-3/5; flat temp recovers 0/5). `maxIterations` defaults to
|
|
96
|
+
* `temperatures.length`; the REAL bound is bareguard (each attempt is gate-checked + metered). CARRIES DOWN the
|
|
97
|
+
* tree (preserved by `forChild`), so it engages at the leaves of a Family-A decomposition. Recovery is PARTIAL
|
|
98
|
+
* (a stubborn blind spot may persist) — `receipts.refineLeaf.passed` reports honestly. Does NOT apply to a node
|
|
99
|
+
* that delegates (its children + the tree verify own quality), nor to the scan/fanout/partition dispatch paths.
|
|
100
|
+
* Absent ⇒ a leaf is a single pass (byte-identical to pre-BA-8). An error-keyed `recall` is the CALLER's tool
|
|
101
|
+
* (`opts.tools`) keyed off the fed-back critique — bareagent stays litectx-agnostic.
|
|
102
|
+
*/
|
|
103
|
+
refineLeaf?: {
|
|
104
|
+
sensor: (result: any, ctx: {
|
|
105
|
+
task: string;
|
|
106
|
+
context: string | undefined;
|
|
107
|
+
contract: string | null;
|
|
108
|
+
}) => (Verdict | Promise<Verdict>);
|
|
109
|
+
maxIterations?: number;
|
|
110
|
+
temperatures?: number[];
|
|
111
|
+
} | undefined;
|
|
72
112
|
/**
|
|
73
113
|
* - Definition of done (A3). When present, the verifier grades against THIS,
|
|
74
114
|
* not the loose task, and verification always runs.
|
|
@@ -177,6 +217,16 @@ export type RecurseNode = {
|
|
|
177
217
|
* - The worker Loop's `metrics.tokens`.
|
|
178
218
|
*/
|
|
179
219
|
tokens: object | null;
|
|
220
|
+
/**
|
|
221
|
+
* - (BA-8) when this leaf
|
|
222
|
+
* ran as a bounded refine loop: how many attempts it took and whether the deterministic sensor finally passed
|
|
223
|
+
* (false = honest non-recovery, not a faked success).
|
|
224
|
+
*/
|
|
225
|
+
refineLeaf?: {
|
|
226
|
+
iterations: number;
|
|
227
|
+
passed: boolean;
|
|
228
|
+
temperatures: number[];
|
|
229
|
+
} | undefined;
|
|
180
230
|
model: string | null;
|
|
181
231
|
/**
|
|
182
232
|
* - (§10 step 7) the retrieval mode this node ran (`scan`/`search`/`exact`),
|
|
@@ -283,8 +333,35 @@ export type Slice = {
|
|
|
283
333
|
* **SECURITY:** this is a PRIVILEGED system-prompt seam — treat `persona` like a system prompt. Do NOT pass
|
|
284
334
|
* untrusted / end-user-controlled text here; a hostile persona is prepended ahead of the decomposition policy
|
|
285
335
|
* and can override it (and any safety framing) for every worker in the tree. Caller-trusted input only.
|
|
336
|
+
* @property {string} [context] - (BA-9 / relayfact F19) An optional caller-supplied READ-ONLY working-context
|
|
337
|
+
* blob (e.g. "project root: /abs/path\nfiles are relative to it") prepended to EVERY worker's task message as
|
|
338
|
+
* a `Working context:` block, so a sliced child can LOCATE its artifact — the concrete context (absolute
|
|
339
|
+
* paths / cwd) the Planner strips when it paraphrases the parent goal into child subtasks. CARRIES DOWN the
|
|
340
|
+
* tree (preserved by `forChild`, like `persona`) and, when forced fan-out plans, is forwarded as the Planner's
|
|
341
|
+
* `info` so the slices themselves are path-aware. Also shown to the verifier (neutral FACTS, not a stance, so
|
|
342
|
+
* no anti-sycophancy concern — and an agentic critic needs the path to exercise the artifact). Distinct from
|
|
343
|
+
* `persona`: persona is a privileged SYSTEM-prompt stance; context is run-state facts on the USER message.
|
|
344
|
+
* Absent/blank ⇒ byte-identical to pre-BA-9 (backward-compatible). Validated live (`poc/ba9-context-thread.mjs`:
|
|
345
|
+
* a weak model went 0/3 → 3/3 at locating an unguessable file once the root was threaded).
|
|
346
|
+
* **SECURITY:** this becomes part of every worker's prompt (and the verifier's) — lower-privilege than
|
|
347
|
+
* `persona` (the USER message, not the SYSTEM prompt) but still a prompt-injection surface. Intended for
|
|
348
|
+
* TRUSTED run-state (paths/cwd); do NOT pass untrusted / end-user-controlled text here.
|
|
286
349
|
* @property {ToolDef[]} [tools] - Handle tools offered to EVERY worker (RC-5 pull-default: litectx
|
|
287
350
|
* `recall`/`get`, wired at build step 7). Workers query on demand; never the whole corpus.
|
|
351
|
+
* @property {{sensor: (result: any, ctx: {task: string, context: string|undefined, contract: string|null}) => (Verdict|Promise<Verdict>), maxIterations?: number, temperatures?: number[]}} [refineLeaf]
|
|
352
|
+
* (Opt-in, BA-8 / relayfact F17) Turn a DEFINITE LEAF (a node that is offered no `spawn_child` — `simple`
|
|
353
|
+
* tier or at `maxDepth`) into a bounded generate→sense→regenerate loop instead of a single pass, so a failed
|
|
354
|
+
* slice can self-correct. `sensor` is a DETERMINISTIC close (test/compile/lint — NOT a model judge, R-S8) that
|
|
355
|
+
* returns a `Verdict`; on a non-pass its `critique` (the GAP, not the transcript) is fed FRESH into the next
|
|
356
|
+
* attempt (D6/A1 anti-anchoring) and the **retry temperature ESCALATES** (`temperatures`, default
|
|
357
|
+
* `[0.2,0.7,1.0]`) — the live-validated requirement that lets a weak model escape a repeat-the-same-mistake rut
|
|
358
|
+
* (`poc/ba8-leaf-refine.mjs`: 0/5 → 2-3/5; flat temp recovers 0/5). `maxIterations` defaults to
|
|
359
|
+
* `temperatures.length`; the REAL bound is bareguard (each attempt is gate-checked + metered). CARRIES DOWN the
|
|
360
|
+
* tree (preserved by `forChild`), so it engages at the leaves of a Family-A decomposition. Recovery is PARTIAL
|
|
361
|
+
* (a stubborn blind spot may persist) — `receipts.refineLeaf.passed` reports honestly. Does NOT apply to a node
|
|
362
|
+
* that delegates (its children + the tree verify own quality), nor to the scan/fanout/partition dispatch paths.
|
|
363
|
+
* Absent ⇒ a leaf is a single pass (byte-identical to pre-BA-8). An error-keyed `recall` is the CALLER's tool
|
|
364
|
+
* (`opts.tools`) keyed off the fed-back critique — bareagent stays litectx-agnostic.
|
|
288
365
|
* @property {string} [contract] - Definition of done (A3). When present, the verifier grades against THIS,
|
|
289
366
|
* not the loose task, and verification always runs.
|
|
290
367
|
* @property {(result: any, ctx: {contract: string|null, task: string}) => (Verdict|Promise<Verdict>)} [evaluate]
|
|
@@ -343,6 +420,9 @@ export type Slice = {
|
|
|
343
420
|
* @property {boolean} incomplete
|
|
344
421
|
* @property {boolean} halted
|
|
345
422
|
* @property {object|null} tokens - The worker Loop's `metrics.tokens`.
|
|
423
|
+
* @property {{iterations: number, passed: boolean, temperatures: number[]}} [refineLeaf] - (BA-8) when this leaf
|
|
424
|
+
* ran as a bounded refine loop: how many attempts it took and whether the deterministic sensor finally passed
|
|
425
|
+
* (false = honest non-recovery, not a faked success).
|
|
346
426
|
* @property {string|null} model
|
|
347
427
|
* @property {string|null} [retrieval] - (§10 step 7) the retrieval mode this node ran (`scan`/`search`/`exact`),
|
|
348
428
|
* or null/absent for a plain reasoning node.
|
package/src/recurse.js
CHANGED
|
@@ -28,6 +28,7 @@ const { Evaluator } = require('./evaluator');
|
|
|
28
28
|
const { Planner } = require('./planner');
|
|
29
29
|
const { runPlan } = require('./run-plan');
|
|
30
30
|
const { assessComplexity, isCritical } = require('./complexity');
|
|
31
|
+
const { refine } = require('./refine');
|
|
31
32
|
const { HaltError } = require('./errors');
|
|
32
33
|
const { DECOMPOSITION_POLICY, capabilityScrub } = require('./recurse-prompts');
|
|
33
34
|
const { synthesize } = require('./recurse-synthesize');
|
|
@@ -51,6 +52,11 @@ const DEFAULT_FANOUT_CONCURRENCY = 4;
|
|
|
51
52
|
// many parallel scan-workers a measured corpus needs. A calibratable knob (the §9.1 algorithm; corpus-specific),
|
|
52
53
|
// not a discovered constant — overridable via `opts.workerBudget`. 100 items ≈ a worker doing ~25 scan windows.
|
|
53
54
|
const DEFAULT_WORKER_BUDGET = 100;
|
|
55
|
+
// BA-8 leaf-refine: temperature ESCALATES per retry. The live POC (poc/ba8-leaf-refine.mjs) found that at a flat
|
|
56
|
+
// low temperature a weak model regenerates byte-identical wrong code and IGNORES even crisp deterministic
|
|
57
|
+
// feedback (0/5 recovery); recovery only appears once retries are given room to vary (0/5 → 2-3/5). So escalation
|
|
58
|
+
// is a DESIGN REQUIREMENT of the seam, not a tuning nicety. Overridable via `opts.refineLeaf.temperatures`.
|
|
59
|
+
const DEFAULT_REFINE_TEMPS = [0.2, 0.7, 1.0];
|
|
54
60
|
|
|
55
61
|
/**
|
|
56
62
|
* Split an array into EXACTLY `n` contiguous, near-equal chunks (the data-partition for the §11 width path).
|
|
@@ -82,6 +88,23 @@ function workerPersonaPrefix(persona) {
|
|
|
82
88
|
return p ? p + '\n\n' : '';
|
|
83
89
|
}
|
|
84
90
|
|
|
91
|
+
/**
|
|
92
|
+
* BA-9 (relayfact F19): prepend the caller's read-only working-context blob to a worker's task message, so a
|
|
93
|
+
* sliced child can LOCATE its artifact (absolute paths / cwd) — the concrete context the Planner otherwise
|
|
94
|
+
* strips when it paraphrases the parent goal into child subtasks. Distinct from `persona`: persona is a STANCE
|
|
95
|
+
* on the SYSTEM prompt (a privileged seam); context is neutral run-state FACTS on the USER message (the form
|
|
96
|
+
* the live POC `poc/ba9-context-thread.mjs` validated: no-context 0/3 → context 3/3 on a weak model). Absent/
|
|
97
|
+
* blank ⇒ the task message is byte-identical to pre-BA-9 (backward-compatible). Carries down the tree via
|
|
98
|
+
* `forChild` (a child of a worker rooted at `/proj` is still rooted at `/proj`), like `persona`.
|
|
99
|
+
* @param {string} task
|
|
100
|
+
* @param {unknown} context
|
|
101
|
+
* @returns {string}
|
|
102
|
+
*/
|
|
103
|
+
function withContext(task, context) {
|
|
104
|
+
const c = typeof context === 'string' ? context.trim() : '';
|
|
105
|
+
return c ? `Working context (read-only):\n${c}\n\n${task}` : task;
|
|
106
|
+
}
|
|
107
|
+
|
|
85
108
|
/**
|
|
86
109
|
* The opts a delegated child inherits. Strips the parent's TOP-LEVEL SETPOINT — `contract`/`evaluate` grade
|
|
87
110
|
* the WHOLE task's final answer; a child grading its own slice against the whole definition-of-done is wasted
|
|
@@ -91,9 +114,10 @@ function workerPersonaPrefix(persona) {
|
|
|
91
114
|
* answered over the parent's corpus; a child has its own subtask and must not re-scan the parent's full corpus
|
|
92
115
|
* (that would fan a whole-corpus count out under every child). The `critical → force-verify` SAFETY FLOOR is
|
|
93
116
|
* unaffected — it keys on the task text via `isCritical`, not the contract, so a critical child still
|
|
94
|
-
* self-verifies. Handle tools (`opts.tools`), `synthesize`, `maxDepth`, and **`
|
|
95
|
-
* persona is a DURABLE worker stance (a child of a "senior security engineer" is still one)
|
|
96
|
-
*
|
|
117
|
+
* self-verifies. Handle tools (`opts.tools`), `synthesize`, `maxDepth`, **`persona`**, and **`context`** (BA-9)
|
|
118
|
+
* carry down — the persona is a DURABLE worker stance (a child of a "senior security engineer" is still one)
|
|
119
|
+
* and the context is durable run-state (a child rooted at `/proj` is still rooted at `/proj`), unlike the
|
|
120
|
+
* top-only `contract`/`evaluate` setpoint. They ride through the `...opts` spread (not in the strip list).
|
|
97
121
|
* @param {RecurseOptions} opts
|
|
98
122
|
* @returns {RecurseOptions}
|
|
99
123
|
*/
|
|
@@ -173,8 +197,35 @@ function auditSafeCtx(ctx, overrides = {}) {
|
|
|
173
197
|
* **SECURITY:** this is a PRIVILEGED system-prompt seam — treat `persona` like a system prompt. Do NOT pass
|
|
174
198
|
* untrusted / end-user-controlled text here; a hostile persona is prepended ahead of the decomposition policy
|
|
175
199
|
* and can override it (and any safety framing) for every worker in the tree. Caller-trusted input only.
|
|
200
|
+
* @property {string} [context] - (BA-9 / relayfact F19) An optional caller-supplied READ-ONLY working-context
|
|
201
|
+
* blob (e.g. "project root: /abs/path\nfiles are relative to it") prepended to EVERY worker's task message as
|
|
202
|
+
* a `Working context:` block, so a sliced child can LOCATE its artifact — the concrete context (absolute
|
|
203
|
+
* paths / cwd) the Planner strips when it paraphrases the parent goal into child subtasks. CARRIES DOWN the
|
|
204
|
+
* tree (preserved by `forChild`, like `persona`) and, when forced fan-out plans, is forwarded as the Planner's
|
|
205
|
+
* `info` so the slices themselves are path-aware. Also shown to the verifier (neutral FACTS, not a stance, so
|
|
206
|
+
* no anti-sycophancy concern — and an agentic critic needs the path to exercise the artifact). Distinct from
|
|
207
|
+
* `persona`: persona is a privileged SYSTEM-prompt stance; context is run-state facts on the USER message.
|
|
208
|
+
* Absent/blank ⇒ byte-identical to pre-BA-9 (backward-compatible). Validated live (`poc/ba9-context-thread.mjs`:
|
|
209
|
+
* a weak model went 0/3 → 3/3 at locating an unguessable file once the root was threaded).
|
|
210
|
+
* **SECURITY:** this becomes part of every worker's prompt (and the verifier's) — lower-privilege than
|
|
211
|
+
* `persona` (the USER message, not the SYSTEM prompt) but still a prompt-injection surface. Intended for
|
|
212
|
+
* TRUSTED run-state (paths/cwd); do NOT pass untrusted / end-user-controlled text here.
|
|
176
213
|
* @property {ToolDef[]} [tools] - Handle tools offered to EVERY worker (RC-5 pull-default: litectx
|
|
177
214
|
* `recall`/`get`, wired at build step 7). Workers query on demand; never the whole corpus.
|
|
215
|
+
* @property {{sensor: (result: any, ctx: {task: string, context: string|undefined, contract: string|null}) => (Verdict|Promise<Verdict>), maxIterations?: number, temperatures?: number[]}} [refineLeaf]
|
|
216
|
+
* (Opt-in, BA-8 / relayfact F17) Turn a DEFINITE LEAF (a node that is offered no `spawn_child` — `simple`
|
|
217
|
+
* tier or at `maxDepth`) into a bounded generate→sense→regenerate loop instead of a single pass, so a failed
|
|
218
|
+
* slice can self-correct. `sensor` is a DETERMINISTIC close (test/compile/lint — NOT a model judge, R-S8) that
|
|
219
|
+
* returns a `Verdict`; on a non-pass its `critique` (the GAP, not the transcript) is fed FRESH into the next
|
|
220
|
+
* attempt (D6/A1 anti-anchoring) and the **retry temperature ESCALATES** (`temperatures`, default
|
|
221
|
+
* `[0.2,0.7,1.0]`) — the live-validated requirement that lets a weak model escape a repeat-the-same-mistake rut
|
|
222
|
+
* (`poc/ba8-leaf-refine.mjs`: 0/5 → 2-3/5; flat temp recovers 0/5). `maxIterations` defaults to
|
|
223
|
+
* `temperatures.length`; the REAL bound is bareguard (each attempt is gate-checked + metered). CARRIES DOWN the
|
|
224
|
+
* tree (preserved by `forChild`), so it engages at the leaves of a Family-A decomposition. Recovery is PARTIAL
|
|
225
|
+
* (a stubborn blind spot may persist) — `receipts.refineLeaf.passed` reports honestly. Does NOT apply to a node
|
|
226
|
+
* that delegates (its children + the tree verify own quality), nor to the scan/fanout/partition dispatch paths.
|
|
227
|
+
* Absent ⇒ a leaf is a single pass (byte-identical to pre-BA-8). An error-keyed `recall` is the CALLER's tool
|
|
228
|
+
* (`opts.tools`) keyed off the fed-back critique — bareagent stays litectx-agnostic.
|
|
178
229
|
* @property {string} [contract] - Definition of done (A3). When present, the verifier grades against THIS,
|
|
179
230
|
* not the loose task, and verification always runs.
|
|
180
231
|
* @property {(result: any, ctx: {contract: string|null, task: string}) => (Verdict|Promise<Verdict>)} [evaluate]
|
|
@@ -234,6 +285,9 @@ function auditSafeCtx(ctx, overrides = {}) {
|
|
|
234
285
|
* @property {boolean} incomplete
|
|
235
286
|
* @property {boolean} halted
|
|
236
287
|
* @property {object|null} tokens - The worker Loop's `metrics.tokens`.
|
|
288
|
+
* @property {{iterations: number, passed: boolean, temperatures: number[]}} [refineLeaf] - (BA-8) when this leaf
|
|
289
|
+
* ran as a bounded refine loop: how many attempts it took and whether the deterministic sensor finally passed
|
|
290
|
+
* (false = honest non-recovery, not a faked success).
|
|
237
291
|
* @property {string|null} model
|
|
238
292
|
* @property {string|null} [retrieval] - (§10 step 7) the retrieval mode this node ran (`scan`/`search`/`exact`),
|
|
239
293
|
* or null/absent for a plain reasoning node.
|
|
@@ -401,6 +455,15 @@ async function recurse(task, ctx = {}, opts = {}) {
|
|
|
401
455
|
if (Array.isArray(opts.corpus)) retrievalTools.push(buildExactTool(normalizeCorpus(opts.corpus)));
|
|
402
456
|
}
|
|
403
457
|
const handleTools = [...(Array.isArray(opts.tools) ? opts.tools : []), ...retrievalTools];
|
|
458
|
+
|
|
459
|
+
// BA-8 (opt-in): a DEFINITE leaf (no spawn offered — `simple` tier or at `maxDepth`) with a caller sensor runs
|
|
460
|
+
// as a bounded refine-with-escalation loop instead of a single pass, so a failed slice self-corrects. Gating on
|
|
461
|
+
// `!canSpawn` keeps it predictable (a node that may delegate is an orchestrator, not a leaf) and means the seam
|
|
462
|
+
// engages exactly at the leaves of a Family-A tree (it carries down via forChild). A no-op when unset.
|
|
463
|
+
if (!canSpawn && opts.refineLeaf && typeof opts.refineLeaf.sensor === 'function') {
|
|
464
|
+
return recurseRefineLeaf(task, ctx, opts, { provider, system, handleTools, depth, critical, node, sensor: opts.refineLeaf.sensor });
|
|
465
|
+
}
|
|
466
|
+
|
|
404
467
|
// NB-3: collect each child's declared RESULT value (copy-on-return: the value, never its transcript) so the
|
|
405
468
|
// reducer can aggregate them. Step-3's seam handed the receipts only, so a code-reduce could not see what to
|
|
406
469
|
// combine — this closes that gap and is what Family B (step 5) will reduce over `runPlan` results[].
|
|
@@ -421,7 +484,8 @@ async function recurse(task, ctx = {}, opts = {}) {
|
|
|
421
484
|
// Fresh message array = a true fresh window (RC-2 copy-on-return, IN side): the worker sees ONLY its task,
|
|
422
485
|
// never a parent transcript. `ctx.depth` is threaded so bareguard's policy can enforce the depth cap (§6).
|
|
423
486
|
const out = await loop.run(
|
|
424
|
-
|
|
487
|
+
// BA-9: prepend the caller's read-only working-context (paths/cwd) so this worker can locate its artifact.
|
|
488
|
+
[{ role: 'user', content: withContext(task, opts.context) }],
|
|
425
489
|
tools,
|
|
426
490
|
// auditSafeCtx: the run ctx reaches the gate as `_ctx`; strip the key-bearing provider (F16/BA-1). The
|
|
427
491
|
// worker Loop already has `provider` as a constructor option, so stripping it from the run ctx is invisible
|
|
@@ -504,6 +568,96 @@ async function recurse(task, ctx = {}, opts = {}) {
|
|
|
504
568
|
}
|
|
505
569
|
}
|
|
506
570
|
|
|
571
|
+
/**
|
|
572
|
+
* BA-8 leaf-refine — run a DEFINITE leaf as a bounded generate→sense→regenerate loop (relayfact F17). Reuses the
|
|
573
|
+
* existing `refine.js` primitive (the Outcomes iterate→grade→revise port): each attempt is a FRESH leaf Loop
|
|
574
|
+
* (fresh window = fresh-feedback, D6/A1) seeded with the working-context'd task + (on a retry) the prior GAP, run
|
|
575
|
+
* at an ESCALATING temperature — the live-validated requirement that lets a weak model escape a repeat-the-same-
|
|
576
|
+
* mistake rut (a flat temperature recovered 0/5 in `poc/ba8-leaf-refine.mjs`). The `sensor` is the caller's
|
|
577
|
+
* DETERMINISTIC close (test/compile/lint, not a model judge). Governance is bareguard's: every attempt is gate-
|
|
578
|
+
* checked (`ctx.policy`) and metered (`onLlmResult`); a HaltError mid-loop is a clean `{incomplete}`. Honest
|
|
579
|
+
* non-recovery is reported (`receipts.refineLeaf.passed=false`), never a faked pass. An optional rubric `verify`
|
|
580
|
+
* still runs on top when a `contract`/`evaluate`/critical applies (the sensor gates retries; the rubric grades).
|
|
581
|
+
* @param {string} task
|
|
582
|
+
* @param {RecurseCtx} ctx
|
|
583
|
+
* @param {RecurseOptions} opts
|
|
584
|
+
* @param {{provider: Provider, system: string, handleTools: ToolDef[], depth: number, critical: boolean, node: RecurseNode, sensor: Function}} state
|
|
585
|
+
* @returns {Promise<RecurseResult>}
|
|
586
|
+
*/
|
|
587
|
+
async function recurseRefineLeaf(task, ctx, opts, state) {
|
|
588
|
+
const { provider, system, handleTools, depth, critical, node, sensor } = state;
|
|
589
|
+
node.model = provider.model || null;
|
|
590
|
+
const cfg = /** @type {{maxIterations?: number, temperatures?: number[]}} */ (opts.refineLeaf || {});
|
|
591
|
+
const temps = Array.isArray(cfg.temperatures) && cfg.temperatures.length ? cfg.temperatures : DEFAULT_REFINE_TEMPS;
|
|
592
|
+
const maxIterations = Number.isInteger(cfg.maxIterations) && /** @type {number} */ (cfg.maxIterations) > 0
|
|
593
|
+
? /** @type {number} */ (cfg.maxIterations) : temps.length;
|
|
594
|
+
|
|
595
|
+
// A refine leaf runs N Loops, so its receipts.tokens SUMS every attempt's spend (not just the last) — the
|
|
596
|
+
// honest cost of the node. The 4-tier tokens object (`{input,output,cacheCreation,cacheRead}`, loop.js) is flat
|
|
597
|
+
// numeric, so we accrue field-wise (robust to extra/renamed numeric fields). The gate already sees each attempt
|
|
598
|
+
// via onLlmResult independently; this is the receipts mirror. Stays null until an attempt produces metrics.
|
|
599
|
+
/** @type {Record<string, number>|null} */
|
|
600
|
+
let tokensSum = null;
|
|
601
|
+
const accrueTokens = (/** @type {any} */ t) => {
|
|
602
|
+
if (!t || typeof t !== 'object') return;
|
|
603
|
+
tokensSum = tokensSum || {};
|
|
604
|
+
for (const [k, v] of Object.entries(t)) if (typeof v === 'number') tokensSum[k] = (tokensSum[k] || 0) + v;
|
|
605
|
+
};
|
|
606
|
+
// One attempt = a fresh leaf Loop (no spawn tool: a retry is a direct correction, not a re-decomposition) at the
|
|
607
|
+
// iteration's temperature, with the GAP fed forward as fresh feedback. A governance halt → throw so refine stops.
|
|
608
|
+
const attempt = async ({ iteration, critique }) => {
|
|
609
|
+
const temperature = temps[Math.min(iteration, temps.length - 1)];
|
|
610
|
+
const loop = new Loop({
|
|
611
|
+
provider, system,
|
|
612
|
+
policy: ctx.policy || undefined,
|
|
613
|
+
onLlmResult: ctx.onLlmResult || undefined,
|
|
614
|
+
stream: ctx.stream || undefined,
|
|
615
|
+
throwOnError: false,
|
|
616
|
+
});
|
|
617
|
+
const base = withContext(task, opts.context);
|
|
618
|
+
const userText = critique
|
|
619
|
+
? `${base}\n\nYour previous attempt FAILED these checks:\n${critique}\n\nReturn a corrected result that passes ALL of them.`
|
|
620
|
+
: base;
|
|
621
|
+
const out = await loop.run([{ role: 'user', content: userText }], handleTools, { ctx: auditSafeCtx(ctx, { depth }), temperature });
|
|
622
|
+
accrueTokens(out.metrics ? out.metrics.tokens : null);
|
|
623
|
+
if (typeof out.error === 'string' && out.error.startsWith('halt:')) throw new HaltError('refine-leaf attempt halted', { rule: out.error.slice('halt:'.length) });
|
|
624
|
+
if (out.error) throw new Error(out.error); // a non-halt worker fault → honest incomplete
|
|
625
|
+
return out.text;
|
|
626
|
+
};
|
|
627
|
+
|
|
628
|
+
try {
|
|
629
|
+
const outcome = await refine({
|
|
630
|
+
attempt,
|
|
631
|
+
evaluate: (result, c) => sensor(result, { task, context: opts.context, contract: c.contract }),
|
|
632
|
+
contract: typeof opts.contract === 'string' ? opts.contract : undefined,
|
|
633
|
+
maxIterations,
|
|
634
|
+
});
|
|
635
|
+
node.tokens = tokensSum;
|
|
636
|
+
node.refineLeaf = { iterations: outcome.iterations, passed: !!(outcome.verdict && outcome.verdict.pass), temperatures: temps.slice(0, outcome.iterations) };
|
|
637
|
+
const result = outcome.result;
|
|
638
|
+
|
|
639
|
+
// Optional rubric layer on top of the deterministic sensor (RC-7): forced for critical, or a contract/override.
|
|
640
|
+
const wantVerify = critical || typeof opts.contract === 'string' || typeof opts.evaluate === 'function';
|
|
641
|
+
if (wantVerify) {
|
|
642
|
+
const verdict = await verify(task, result, ctx, opts);
|
|
643
|
+
node.verdict = verdict;
|
|
644
|
+
return { result, verdict, receipts: node };
|
|
645
|
+
}
|
|
646
|
+
// No rubric layer ⇒ the sensor's final verdict IS the node verdict (a non-pass is surfaced, not hidden).
|
|
647
|
+
node.verdict = outcome.verdict || null;
|
|
648
|
+
return { result, verdict: outcome.verdict || null, receipts: node };
|
|
649
|
+
} catch (err) {
|
|
650
|
+
node.tokens = tokensSum; // record whatever attempts DID spend, on both the halt and fault paths
|
|
651
|
+
if (err instanceof HaltError) {
|
|
652
|
+
node.halted = true;
|
|
653
|
+
node.incomplete = true;
|
|
654
|
+
return { incomplete: true, best: null, receipts: node };
|
|
655
|
+
}
|
|
656
|
+
node.incomplete = true;
|
|
657
|
+
return { incomplete: true, best: null, receipts: node };
|
|
658
|
+
}
|
|
659
|
+
}
|
|
660
|
+
|
|
507
661
|
/**
|
|
508
662
|
* SCAN (§10 step 7 / §9.2.1) — the default retrieval mode for a "how many / all" task over a corpus. A
|
|
509
663
|
* deterministic ORCHESTRATION, not a worker model call: every slice is processed, an isolated Loop LLM-judges
|
|
@@ -747,7 +901,10 @@ async function recurseFanout(task, ctx, opts, state) {
|
|
|
747
901
|
const planner = new Planner({ provider, onLlmResult: /** @type {any} */ (ctx.onLlmResult) || undefined });
|
|
748
902
|
let steps;
|
|
749
903
|
try {
|
|
750
|
-
|
|
904
|
+
// BA-9: forward the working-context as the Planner's `info` so the slices it writes are path-aware (a
|
|
905
|
+
// child still also receives `opts.context` directly via `forChild` — this just improves the split).
|
|
906
|
+
const planContext = typeof opts.context === 'string' && opts.context.trim() ? { count, info: opts.context } : { count };
|
|
907
|
+
steps = await planner.plan(task, planContext);
|
|
751
908
|
} catch (err) {
|
|
752
909
|
if (err instanceof HaltError) throw err;
|
|
753
910
|
node.incomplete = true;
|
|
@@ -928,6 +1085,9 @@ function buildSpawnTool(ctx, opts, depth, maxDepth, node, childResults) {
|
|
|
928
1085
|
*/
|
|
929
1086
|
function verify(task, result, ctx, opts) {
|
|
930
1087
|
const contract = typeof opts.contract === 'string' ? opts.contract : null;
|
|
1088
|
+
// BA-9: the verifier sees the working-context too — neutral facts (not a stance, so no anti-sycophancy risk),
|
|
1089
|
+
// and an agentic critic needs the path to exercise the artifact. A caller `evaluate` gets the RAW task (it owns
|
|
1090
|
+
// its own context); only the default isolated grader is contextualized.
|
|
931
1091
|
if (typeof opts.evaluate === 'function') {
|
|
932
1092
|
return Promise.resolve(opts.evaluate(result, { contract, task }));
|
|
933
1093
|
}
|
|
@@ -937,7 +1097,7 @@ function verify(task, result, ctx, opts) {
|
|
|
937
1097
|
? 'Judge whether the result satisfies the definition of done. Be strict and adversarial; cite the specific gap on any shortfall.'
|
|
938
1098
|
: 'Judge whether the result fully and correctly answers the goal. Be strict and adversarial; cite the specific gap on any shortfall.';
|
|
939
1099
|
return evaluator.evaluate(
|
|
940
|
-
task,
|
|
1100
|
+
withContext(task, opts.context),
|
|
941
1101
|
result,
|
|
942
1102
|
{ rubric, contract: contract || undefined },
|
|
943
1103
|
{ onLlmResult: /** @type {any} */ (ctx.onLlmResult), policy: ctx.policy },
|