@pmelab/gtd 36.0.0 → 38.0.0
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/.claude-plugin/plugin.json +1 -1
- package/README.md +30 -9
- package/claude/hooks/context.ts +66 -0
- package/claude/hooks/drive.ts +3 -0
- package/claude/hooks/register.tsx +69 -6
- package/dist/chunks/{reserved-BAv71I4i.mjs → reserved-sfMdy7v7.mjs} +83 -22
- package/dist/chunks/{tail-kar1zhOb.mjs → tail-Cgca5Gp-.mjs} +1 -1
- package/dist/chunks/{wire-D3Rc5r22.mjs → wire-CSsa3Wxd.mjs} +117 -2
- package/dist/gtd.bundle.mjs +498 -56
- package/dist/testkit.mjs +115 -3
- package/dist/wire.mjs +2 -2
- package/dist/workflows/cleanup.mjs +2 -2
- package/dist/workflows/feature.mjs +277 -58
- package/dist/workflows/fix.mjs +2 -2
- package/dist/workflows/review.mjs +2 -2
- package/dist/workflows/rollout.mjs +1 -1
- package/dist/workflows.mjs +355 -75
- package/package.json +1 -1
- package/src/cli/Cli.test.ts +10 -3
- package/src/cli/Cli.ts +76 -1
- package/src/cli/Program.test.ts +1 -0
- package/src/cli/Program.ts +28 -0
- package/src/cli/ToolCommands.ts +3 -0
- package/src/edge/Edge.ts +11 -0
- package/src/flows/runtime.ts +11 -0
- package/src/replay/Replay.ts +51 -2
- package/src/testkit/Layers.ts +1 -0
- package/src/testkit/steps.ts +35 -0
- package/src/wire/BeatDocument.ts +12 -2
- package/src/wire/BeatStatus.ts +3 -0
- package/src/wire/ContextLimit.fixture.ts +30 -0
- package/src/wire/ContextLimit.test.ts +96 -0
- package/src/wire/ContextLimit.ts +169 -0
- package/src/wire/index.ts +1 -0
- package/src/wire/types.ts +7 -0
- package/src/workflow/Workflow.ts +9 -1
- package/src/workflows/feature/context-limit.feature +37 -0
- package/src/workflows/feature/coverage.ts +22 -8
- package/src/workflows/feature/criteria.test.ts +17 -1
- package/src/workflows/feature/criteria.ts +26 -3
- package/src/workflows/feature/feature.ts +69 -16
- package/src/workflows/feature/index.ts +1 -0
- package/src/workflows/feature/refactor-absorbs.feature +12 -5
- package/src/workflows/feature/refactor-first.feature +11 -2
- package/src/workflows/feature/refactor-lens-models.feature +4 -0
- package/src/workflows/feature/refactor-rebuild.feature +269 -0
- package/src/workflows/feature/refactor-review.feature +45 -1
- package/src/workflows/feature/refactor.ts +9 -3
- package/src/workflows/feature/seam.test.ts +108 -3
- package/src/workflows/feature/seam.ts +169 -31
- package/src/workflows/feature/slicing-cap.feature +14 -0
- package/src/workflows/feature/slicing.feature +225 -0
- package/src/workflows/feature/slicing.test.ts +49 -0
- package/src/workflows/feature/slicing.ts +105 -0
- package/src/workflows/index.ts +1 -0
- package/src/workflows/shared/fences.ts +5 -0
- package/src/workflows/shared/options.test.ts +11 -2
- package/src/workflows/shared/options.ts +19 -4
- package/src/workflows/shared/steps.ts +2 -0
- package/src/workflows/shared/text.ts +129 -28
- package/src/workflows/shared/vars.ts +1 -0
- package/src/workflows/shared/workflows.steps.ts +13 -0
package/package.json
CHANGED
package/src/cli/Cli.test.ts
CHANGED
|
@@ -1106,6 +1106,7 @@ describe("standaloneKinds / needsOf", () => {
|
|
|
1106
1106
|
"uncheck",
|
|
1107
1107
|
"install",
|
|
1108
1108
|
"loop",
|
|
1109
|
+
"contextHook",
|
|
1109
1110
|
"judgeRun",
|
|
1110
1111
|
"strays",
|
|
1111
1112
|
])
|
|
@@ -1117,6 +1118,7 @@ describe("standaloneKinds / needsOf", () => {
|
|
|
1117
1118
|
expect(needsOf("uncheck")).toBe("fs")
|
|
1118
1119
|
expect(needsOf("install")).toBe("none")
|
|
1119
1120
|
expect(needsOf("loop")).toBe("none")
|
|
1121
|
+
expect(needsOf("contextHook")).toBe("none")
|
|
1120
1122
|
expect(needsOf("judgeRun")).toBe("none")
|
|
1121
1123
|
expect(needsOf("strays")).toBe("none")
|
|
1122
1124
|
for (const kind of [
|
|
@@ -1202,19 +1204,24 @@ describe("renderHelp", () => {
|
|
|
1202
1204
|
|
|
1203
1205
|
it("README.md and COMPATIBILITY.md list exactly the commands", () => {
|
|
1204
1206
|
const commands = new Set(
|
|
1205
|
-
[...renderHelp().matchAll(/^ {2}([a-z]+(?: [a-z]+)?)(?= )/gm)].map(
|
|
1207
|
+
[...renderHelp().matchAll(/^ {2}([a-z]+(?:-[a-z]+)?(?: [a-z]+)?)(?= )/gm)].map(
|
|
1208
|
+
([, name]) => name,
|
|
1209
|
+
),
|
|
1206
1210
|
)
|
|
1207
1211
|
expect(commands).toContain("judge answer")
|
|
1212
|
+
expect(commands).toContain("context-hook")
|
|
1208
1213
|
const listed = (file: string, list: RegExp): Set<string> => {
|
|
1209
1214
|
const doc = readFileSync(resolve(import.meta.dirname, file), "utf8")
|
|
1210
1215
|
const match = doc.match(list)
|
|
1211
1216
|
expect(match, file).not.toBeNull()
|
|
1212
|
-
return new Set(
|
|
1217
|
+
return new Set(
|
|
1218
|
+
[...match![1]!.matchAll(/`([a-z]+(?:-[a-z]+)?(?: [a-z]+)?)`/g)].map(([, name]) => name!),
|
|
1219
|
+
)
|
|
1213
1220
|
}
|
|
1214
1221
|
expect(
|
|
1215
1222
|
listed("../../README.md", /The `gtd` bin's commands — ([\s\S]*?) — are\s+listed/),
|
|
1216
1223
|
).toEqual(commands)
|
|
1217
|
-
expect(listed("../../COMPATIBILITY.md", /every command \(([\s\S]*?)\)
|
|
1224
|
+
expect(listed("../../COMPATIBILITY.md", /every command \(([\s\S]*?)\),\s+option/)).toEqual(
|
|
1218
1225
|
commands,
|
|
1219
1226
|
)
|
|
1220
1227
|
})
|
package/src/cli/Cli.ts
CHANGED
|
@@ -35,7 +35,7 @@ import {
|
|
|
35
35
|
type JsonMode,
|
|
36
36
|
} from "./Program.js"
|
|
37
37
|
import type { ToolCommands } from "./ToolCommands.js"
|
|
38
|
-
import { EXIT_RUNTIME_ERROR, EXIT_USAGE_ERROR } from "../wire/index.js"
|
|
38
|
+
import { EXIT_RUNTIME_ERROR, EXIT_USAGE_ERROR, parseContextWindow } from "../wire/index.js"
|
|
39
39
|
|
|
40
40
|
export type { Command, CommandRequirements } from "./Program.js"
|
|
41
41
|
export { needsOf, standaloneKinds } from "./Program.js"
|
|
@@ -90,6 +90,13 @@ const nonNegativeNumber = (raw: string, flag: string): Either.Either<number, str
|
|
|
90
90
|
return Either.right(n)
|
|
91
91
|
}
|
|
92
92
|
|
|
93
|
+
const share = (raw: string, flag: string): Either.Either<number, string> => {
|
|
94
|
+
const n = Number(raw)
|
|
95
|
+
return raw.trim() !== "" && n > 0 && n <= 1
|
|
96
|
+
? Either.right(n)
|
|
97
|
+
: Either.left(`gtd: ${flag} must be a share of the context window in (0, 1] — got "${raw}"`)
|
|
98
|
+
}
|
|
99
|
+
|
|
93
100
|
const FLAGS: readonly FlagRow[] = [
|
|
94
101
|
{
|
|
95
102
|
name: "--json",
|
|
@@ -245,6 +252,41 @@ const FLAGS: readonly FlagRow[] = [
|
|
|
245
252
|
"GTD_JUDGE_MODEL, then judge.json, then sonnet",
|
|
246
253
|
],
|
|
247
254
|
},
|
|
255
|
+
{
|
|
256
|
+
name: "--soft",
|
|
257
|
+
arity: 1,
|
|
258
|
+
repeatable: false,
|
|
259
|
+
scope: (kind) => kind === "contextHook",
|
|
260
|
+
decode: ([raw]) => share(raw ?? "", "--soft"),
|
|
261
|
+
scopeError: "gtd: --soft is only valid for `gtd context-hook`",
|
|
262
|
+
valueHint: "<share>",
|
|
263
|
+
help: ["(gtd context-hook only) the soft limit; default 0.6"],
|
|
264
|
+
},
|
|
265
|
+
{
|
|
266
|
+
name: "--hard",
|
|
267
|
+
arity: 1,
|
|
268
|
+
repeatable: false,
|
|
269
|
+
scope: (kind) => kind === "contextHook",
|
|
270
|
+
decode: ([raw]) => share(raw ?? "", "--hard"),
|
|
271
|
+
scopeError: "gtd: --hard is only valid for `gtd context-hook`",
|
|
272
|
+
valueHint: "<share>",
|
|
273
|
+
help: ["(gtd context-hook only) the hard limit; default 0.7"],
|
|
274
|
+
},
|
|
275
|
+
{
|
|
276
|
+
name: "--window",
|
|
277
|
+
arity: 1,
|
|
278
|
+
repeatable: false,
|
|
279
|
+
scope: (kind) => kind === "contextHook",
|
|
280
|
+
decode: ([raw]) => {
|
|
281
|
+
const window = parseContextWindow(raw)
|
|
282
|
+
return window !== undefined
|
|
283
|
+
? Either.right(window)
|
|
284
|
+
: Either.left(`gtd: --window must be a positive integer — got "${raw ?? ""}"`)
|
|
285
|
+
},
|
|
286
|
+
scopeError: "gtd: --window is only valid for `gtd context-hook`",
|
|
287
|
+
valueHint: "<tokens>",
|
|
288
|
+
help: ["(gtd context-hook only) the model's context window in", "tokens; default 200000"],
|
|
289
|
+
},
|
|
248
290
|
{
|
|
249
291
|
name: "--slice",
|
|
250
292
|
arity: 1,
|
|
@@ -593,6 +635,22 @@ const COMMAND_ROWS: readonly CommandRow[] = [
|
|
|
593
635
|
"~/.config/gtd/loop.json); see docs/driver.md",
|
|
594
636
|
],
|
|
595
637
|
},
|
|
638
|
+
{
|
|
639
|
+
token: "context-hook",
|
|
640
|
+
kind: "contextHook",
|
|
641
|
+
arity: "none",
|
|
642
|
+
details: [
|
|
643
|
+
"A Claude Code PostToolUse and PreCompact hook that stops an",
|
|
644
|
+
"agent turn at a share of its context window. Reads the",
|
|
645
|
+
"hook's input on stdin and the context the transcript's",
|
|
646
|
+
"latest assistant message used. Past --soft it blocks with a",
|
|
647
|
+
"request to end the turn with a handoff note; past --hard,",
|
|
648
|
+
"or at PreCompact, it stops the turn with a stopReason",
|
|
649
|
+
'starting "gtd: context limit". Prints nothing otherwise,',
|
|
650
|
+
"and for a subagent's tool call. Always exits 0 on valid",
|
|
651
|
+
"input; see docs/driver.md",
|
|
652
|
+
],
|
|
653
|
+
},
|
|
596
654
|
{
|
|
597
655
|
token: "summary",
|
|
598
656
|
kind: "summary",
|
|
@@ -1182,6 +1240,9 @@ export const parseArgv = (argv: readonly string[]): CliPlan => {
|
|
|
1182
1240
|
readonly "--provider"?: "fixed" | "jev" | "llm"
|
|
1183
1241
|
readonly "--answers"?: string
|
|
1184
1242
|
readonly "--slice"?: string
|
|
1243
|
+
readonly "--soft"?: number
|
|
1244
|
+
readonly "--hard"?: number
|
|
1245
|
+
readonly "--window"?: number
|
|
1185
1246
|
}
|
|
1186
1247
|
|
|
1187
1248
|
// `present.has("--json")` alone can't distinguish bare `--json` (no value
|
|
@@ -1285,6 +1346,20 @@ export const parseArgv = (argv: readonly string[]): CliPlan => {
|
|
|
1285
1346
|
}
|
|
1286
1347
|
}
|
|
1287
1348
|
|
|
1349
|
+
if (kind === "contextHook") {
|
|
1350
|
+
return {
|
|
1351
|
+
kind: "command",
|
|
1352
|
+
command: {
|
|
1353
|
+
kind: "contextHook",
|
|
1354
|
+
...(bag["--soft"] !== undefined ? { soft: bag["--soft"] } : {}),
|
|
1355
|
+
...(bag["--hard"] !== undefined ? { hard: bag["--hard"] } : {}),
|
|
1356
|
+
...(bag["--window"] !== undefined ? { window: bag["--window"] } : {}),
|
|
1357
|
+
},
|
|
1358
|
+
json,
|
|
1359
|
+
verbose,
|
|
1360
|
+
}
|
|
1361
|
+
}
|
|
1362
|
+
|
|
1288
1363
|
if (kind === "summary") {
|
|
1289
1364
|
return {
|
|
1290
1365
|
kind: "command",
|
package/src/cli/Program.test.ts
CHANGED
|
@@ -1686,6 +1686,7 @@ describe("runCommand — refuses in a repository with no commits", () => {
|
|
|
1686
1686
|
uncheck: { kind: "uncheck", file: ".gtd/REVIEW.md" },
|
|
1687
1687
|
install: { kind: "install" },
|
|
1688
1688
|
loop: { kind: "loop" },
|
|
1689
|
+
contextHook: { kind: "contextHook" },
|
|
1689
1690
|
summary: { kind: "summary" },
|
|
1690
1691
|
base: { kind: "base" },
|
|
1691
1692
|
judge: { kind: "judge" },
|
package/src/cli/Program.ts
CHANGED
|
@@ -162,6 +162,12 @@ export type Command =
|
|
|
162
162
|
| { readonly kind: "uncheck"; readonly file: string }
|
|
163
163
|
| { readonly kind: "install" }
|
|
164
164
|
| { readonly kind: "loop" }
|
|
165
|
+
| {
|
|
166
|
+
readonly kind: "contextHook"
|
|
167
|
+
readonly soft?: number
|
|
168
|
+
readonly hard?: number
|
|
169
|
+
readonly window?: number
|
|
170
|
+
}
|
|
165
171
|
| { readonly kind: "summary"; readonly slice?: string }
|
|
166
172
|
| { readonly kind: "base" }
|
|
167
173
|
| { readonly kind: "judge" }
|
|
@@ -434,6 +440,24 @@ const readStdin = (): Effect.Effect<string, Error> =>
|
|
|
434
440
|
catch: (e) => (e instanceof Error ? e : new Error(String(e))),
|
|
435
441
|
})
|
|
436
442
|
|
|
443
|
+
/** Stdout is written once: a hook reads one JSON document, or none. */
|
|
444
|
+
const runContextHookCommand = (
|
|
445
|
+
command: Extract<Command, { kind: "contextHook" }>,
|
|
446
|
+
out: ArtifactOut,
|
|
447
|
+
): Effect.Effect<void, Error, ToolCommands> =>
|
|
448
|
+
Effect.gen(function* () {
|
|
449
|
+
const tools = yield* ToolCommands
|
|
450
|
+
const input = yield* readStdin()
|
|
451
|
+
out.write(
|
|
452
|
+
yield* tools.contextHook({
|
|
453
|
+
input,
|
|
454
|
+
...(command.soft !== undefined ? { soft: command.soft } : {}),
|
|
455
|
+
...(command.hard !== undefined ? { hard: command.hard } : {}),
|
|
456
|
+
...(command.window !== undefined ? { window: command.window } : {}),
|
|
457
|
+
}),
|
|
458
|
+
)
|
|
459
|
+
})
|
|
460
|
+
|
|
437
461
|
/** Stdout is written once, after the answerer resolves — a rejection leaves it empty. */
|
|
438
462
|
const runJudgeRunCommand = (
|
|
439
463
|
command: Extract<Command, { kind: "judgeRun" }>,
|
|
@@ -1189,6 +1213,7 @@ export const needsOf = (kind: Command["kind"]): Needs => {
|
|
|
1189
1213
|
case "lsp":
|
|
1190
1214
|
case "install":
|
|
1191
1215
|
case "loop":
|
|
1216
|
+
case "contextHook":
|
|
1192
1217
|
case "judgeRun":
|
|
1193
1218
|
case "strays":
|
|
1194
1219
|
return "none"
|
|
@@ -1207,6 +1232,7 @@ export const standaloneKinds = (): readonly Command["kind"][] => [
|
|
|
1207
1232
|
"uncheck",
|
|
1208
1233
|
"install",
|
|
1209
1234
|
"loop",
|
|
1235
|
+
"contextHook",
|
|
1210
1236
|
"judgeRun",
|
|
1211
1237
|
"strays",
|
|
1212
1238
|
]
|
|
@@ -1278,6 +1304,8 @@ const dispatchVoidCommand = (
|
|
|
1278
1304
|
return runInstallCommand(out)
|
|
1279
1305
|
case "loop":
|
|
1280
1306
|
return runLoopCommand()
|
|
1307
|
+
case "contextHook":
|
|
1308
|
+
return runContextHookCommand(command, out)
|
|
1281
1309
|
case "summary":
|
|
1282
1310
|
return runSummaryCommand(out, command.slice)
|
|
1283
1311
|
case "base":
|
package/src/cli/ToolCommands.ts
CHANGED
|
@@ -1,6 +1,7 @@
|
|
|
1
1
|
import { Context } from "effect"
|
|
2
2
|
import type { Effect } from "effect"
|
|
3
3
|
import type { UiSettings } from "../flows/index.js"
|
|
4
|
+
import type { ContextHookRequest } from "../wire/index.js"
|
|
4
5
|
|
|
5
6
|
/** How to run THIS gtd — a driver (the loop, the LSP) never reads a different version's state than the one it ships with. */
|
|
6
7
|
export type SelfGtd = readonly [string, ...string[]]
|
|
@@ -38,6 +39,8 @@ export class ToolCommands extends Context.Tag("ToolCommands")<
|
|
|
38
39
|
) => Effect.Effect<void, Error>
|
|
39
40
|
readonly lsp: (gtd: SelfGtd) => Effect.Effect<void, Error>
|
|
40
41
|
readonly loop: (gtd: SelfGtd) => Effect.Effect<void, Error>
|
|
42
|
+
/** The hook's stdout: a JSON document, or empty to let the turn go on. */
|
|
43
|
+
readonly contextHook: (request: ContextHookRequest) => Effect.Effect<string, Error>
|
|
41
44
|
readonly judge: (request: JudgeRequest) => Promise<string>
|
|
42
45
|
}
|
|
43
46
|
>() {}
|
package/src/edge/Edge.ts
CHANGED
|
@@ -54,6 +54,7 @@ import {
|
|
|
54
54
|
type ModelCost,
|
|
55
55
|
type SliceEntry,
|
|
56
56
|
type StepAccess,
|
|
57
|
+
type StepContextLimit,
|
|
57
58
|
} from "../wire/index.js"
|
|
58
59
|
|
|
59
60
|
import { GTD_VERSION } from "../install/index.js"
|
|
@@ -722,6 +723,13 @@ type RequestOf<K extends ReachedStep["request"]["kind"]> = Extract<
|
|
|
722
723
|
{ kind: K }
|
|
723
724
|
>
|
|
724
725
|
|
|
726
|
+
const stepContext = (
|
|
727
|
+
context: { readonly soft?: number | undefined; readonly hard?: number | undefined } | undefined,
|
|
728
|
+
): StepContextLimit | undefined =>
|
|
729
|
+
context === undefined
|
|
730
|
+
? undefined
|
|
731
|
+
: { ...optional("soft", context.soft), ...optional("hard", context.hard) }
|
|
732
|
+
|
|
725
733
|
const promptDef = (
|
|
726
734
|
common: StepCommon,
|
|
727
735
|
request: RequestOf<"agent">,
|
|
@@ -735,6 +743,7 @@ const promptDef = (
|
|
|
735
743
|
...optional("system", request.options.system),
|
|
736
744
|
...optional("allowEmpty", request.options.allowEmpty),
|
|
737
745
|
...optional("keep", request.options.keep),
|
|
746
|
+
...optional("context", stepContext(request.options.context)),
|
|
738
747
|
...optional("skills", skills),
|
|
739
748
|
...optional("access", access),
|
|
740
749
|
})
|
|
@@ -850,6 +859,7 @@ export interface RestHints {
|
|
|
850
859
|
readonly mode?: string
|
|
851
860
|
readonly skills?: readonly string[]
|
|
852
861
|
readonly access?: StepAccess
|
|
862
|
+
readonly context?: StepContextLimit
|
|
853
863
|
}
|
|
854
864
|
|
|
855
865
|
const hintsOf = (def: StepDef): RestHints => ({
|
|
@@ -861,6 +871,7 @@ const hintsOf = (def: StepDef): RestHints => ({
|
|
|
861
871
|
...optional("mode", def.mode),
|
|
862
872
|
...optional("skills", def.skills),
|
|
863
873
|
...optional("access", def.access),
|
|
874
|
+
...optional("context", def.context),
|
|
864
875
|
})
|
|
865
876
|
|
|
866
877
|
const normalizeStatus = (raw: string): ChangeStatus => (raw === "A" ? "A" : raw === "D" ? "D" : "M")
|
package/src/flows/runtime.ts
CHANGED
|
@@ -16,9 +16,18 @@ export interface SteeringOptions {
|
|
|
16
16
|
readonly base?: string | undefined
|
|
17
17
|
}
|
|
18
18
|
|
|
19
|
+
/** Shares of the model's context window at which a driver stops the turn — `gtd next --json`'s `context`. */
|
|
20
|
+
export interface ContextLimitOptions {
|
|
21
|
+
/** Ask the agent to wrap up with a handoff note. Driver default 0.6. */
|
|
22
|
+
readonly soft?: number | undefined
|
|
23
|
+
/** Stop the turn outright. Driver default 0.7. */
|
|
24
|
+
readonly hard?: number | undefined
|
|
25
|
+
}
|
|
26
|
+
|
|
19
27
|
export interface AgentOptions extends SteeringOptions {
|
|
20
28
|
readonly model?: string | undefined
|
|
21
29
|
readonly system?: string | undefined
|
|
30
|
+
readonly context?: ContextLimitOptions | undefined
|
|
22
31
|
/**
|
|
23
32
|
* A turn that changes nothing completes the step. Without this an empty turn
|
|
24
33
|
* is an attempt: recorded, but the process stays at the step (a stall).
|
|
@@ -115,6 +124,8 @@ export interface ScopeOptions {
|
|
|
115
124
|
readonly model?: string | undefined
|
|
116
125
|
/** The system prompt every agent step inside runs with, unless it sets its own. */
|
|
117
126
|
readonly system?: string | undefined
|
|
127
|
+
/** The context limit every agent step inside runs with; a nested scope's or a step's own replaces it key by key. */
|
|
128
|
+
readonly context?: ContextLimitOptions | undefined
|
|
118
129
|
/** The skill names every agent step inside declares — `gtd next --json`'s `skills`. A nested scope inherits it unless it sets its own. */
|
|
119
130
|
readonly skills?: readonly string[] | undefined
|
|
120
131
|
/** The file access every agent step inside runs with — `gtd next --json`'s `access`. A nested scope inherits it; its own replaces it wholesale. */
|
package/src/replay/Replay.ts
CHANGED
|
@@ -8,6 +8,7 @@ import {
|
|
|
8
8
|
type ScopeAccess,
|
|
9
9
|
type AccessDef,
|
|
10
10
|
type ScopeOptions,
|
|
11
|
+
type ContextLimitOptions,
|
|
11
12
|
type StepRequest,
|
|
12
13
|
type Flow,
|
|
13
14
|
type Mark,
|
|
@@ -21,7 +22,7 @@ import {
|
|
|
21
22
|
unansweredQuestions,
|
|
22
23
|
} from "../steering/index.js"
|
|
23
24
|
import { accessShapeFault, foldAccess, sameAccess } from "./Access.js"
|
|
24
|
-
import { globMatches } from "../wire/index.js"
|
|
25
|
+
import { contextLimitFault, contextShapeFault, globMatches } from "../wire/index.js"
|
|
25
26
|
import { refNameFault } from "./RefName.js"
|
|
26
27
|
import { IMPLICIT_SLICE, SLICE_NAME, type SliceSpan } from "./Slices.js"
|
|
27
28
|
import { diffTrees, isEmptyDiff, type TreeView } from "./Tree.js"
|
|
@@ -207,6 +208,28 @@ const budgeted = (
|
|
|
207
208
|
return { evidence: bounded, truncated }
|
|
208
209
|
}
|
|
209
210
|
|
|
211
|
+
// The innermost of `chain`, named by its full path.
|
|
212
|
+
const scopeLabel = (chain: readonly ScopeOptions[]): string => {
|
|
213
|
+
const path = chain.flatMap((o) => (o.name === undefined ? [] : [o.name])).join(".")
|
|
214
|
+
if (chain.at(-1)?.name !== undefined) return `scope "${path}"`
|
|
215
|
+
return path === "" ? "an unnamed scope" : `an unnamed scope inside "${path}"`
|
|
216
|
+
}
|
|
217
|
+
|
|
218
|
+
// Key by key, innermost last: a scope or step that sets only `hard` keeps the outer `soft`.
|
|
219
|
+
const inheritedContext = (
|
|
220
|
+
layers: readonly (ContextLimitOptions | undefined)[],
|
|
221
|
+
): ContextLimitOptions | undefined =>
|
|
222
|
+
layers.reduce<ContextLimitOptions | undefined>(
|
|
223
|
+
(merged, layer) =>
|
|
224
|
+
layer === undefined
|
|
225
|
+
? merged
|
|
226
|
+
: {
|
|
227
|
+
...merged,
|
|
228
|
+
...Object.fromEntries(Object.entries(layer).filter(([, value]) => value !== undefined)),
|
|
229
|
+
},
|
|
230
|
+
undefined,
|
|
231
|
+
)
|
|
232
|
+
|
|
210
233
|
interface Identity {
|
|
211
234
|
readonly model: string | undefined
|
|
212
235
|
readonly system: string | undefined
|
|
@@ -243,7 +266,7 @@ const STEERING_OPTIONS = ["file", "mode", "label", "base"]
|
|
|
243
266
|
// The option keys each step accepts. A gtd.config.ts is evaluated without a
|
|
244
267
|
// type check, so a misspelt key would otherwise be silently ignored.
|
|
245
268
|
const KNOWN_OPTIONS: Readonly<Record<StepKind, ReadonlySet<string>>> = {
|
|
246
|
-
agent: new Set([...STEERING_OPTIONS, "model", "system", "allowEmpty", "keep"]),
|
|
269
|
+
agent: new Set([...STEERING_OPTIONS, "model", "system", "context", "allowEmpty", "keep"]),
|
|
247
270
|
human: new Set([...STEERING_OPTIONS, "message", "acceptClean", "approve"]),
|
|
248
271
|
run: new Set(STEERING_OPTIONS),
|
|
249
272
|
judge: new Set([...STEERING_OPTIONS, "message"]),
|
|
@@ -426,11 +449,13 @@ export const replay = async (input: ReplayInput): Promise<ReplayOutcome> => {
|
|
|
426
449
|
if (s.model !== undefined) persona.model = s.model
|
|
427
450
|
if (s.system !== undefined) persona.system = s.system
|
|
428
451
|
}
|
|
452
|
+
const context = inheritedContext([...scopes.map((s) => s.context), request.options.context])
|
|
429
453
|
return {
|
|
430
454
|
...request,
|
|
431
455
|
options: {
|
|
432
456
|
...persona,
|
|
433
457
|
...request.options,
|
|
458
|
+
...(context === undefined ? {} : { context }),
|
|
434
459
|
},
|
|
435
460
|
}
|
|
436
461
|
}
|
|
@@ -575,6 +600,28 @@ export const replay = async (input: ReplayInput): Promise<ReplayOutcome> => {
|
|
|
575
600
|
return landPending(step, input.pending)
|
|
576
601
|
}
|
|
577
602
|
|
|
603
|
+
// Each layer on its own first, so a bad scope or workflow option is named as
|
|
604
|
+
// such rather than blamed on the step it reached; then the merged limit.
|
|
605
|
+
const scopeContextFault = (): string | undefined => {
|
|
606
|
+
for (const [i, s] of scopes.entries()) {
|
|
607
|
+
const fault = s.context === undefined ? undefined : contextShapeFault(s.context)
|
|
608
|
+
if (fault !== undefined)
|
|
609
|
+
return `the context of ${scopeLabel(scopes.slice(0, i + 1))}: ${fault}`
|
|
610
|
+
}
|
|
611
|
+
return undefined
|
|
612
|
+
}
|
|
613
|
+
|
|
614
|
+
const resolvedContextFault = (
|
|
615
|
+
request: Extract<StepRequest, { kind: "agent" }>,
|
|
616
|
+
): string | undefined => {
|
|
617
|
+
const own = request.options.context
|
|
618
|
+
const layered = scopeContextFault() ?? (own === undefined ? undefined : contextShapeFault(own))
|
|
619
|
+
if (layered !== undefined) return layered
|
|
620
|
+
const resolved = resolve(request)
|
|
621
|
+
const context = resolved.kind === "agent" ? resolved.options.context : undefined
|
|
622
|
+
return context === undefined ? undefined : contextLimitFault(context)
|
|
623
|
+
}
|
|
624
|
+
|
|
578
625
|
// gtd.config.ts is loaded without a type check.
|
|
579
626
|
const requestFault = (request: Exclude<StepRequest, { kind: "restart" }>): string | undefined => {
|
|
580
627
|
if (request.kind === "run" && typeof request.body !== "string") {
|
|
@@ -582,6 +629,8 @@ export const replay = async (input: ReplayInput): Promise<ReplayOutcome> => {
|
|
|
582
629
|
}
|
|
583
630
|
if (request.kind !== "agent") return undefined
|
|
584
631
|
const name = scoped(request.name)
|
|
632
|
+
const contextFault = resolvedContextFault(request)
|
|
633
|
+
if (contextFault !== undefined) return `gtd: step "${name}": ${contextFault}`
|
|
585
634
|
if (request.options.keep === true) {
|
|
586
635
|
// A clean allowEmpty keep would land an ordinary commit with no Gtd-Kept: a silent skip.
|
|
587
636
|
if (request.options.allowEmpty === true) {
|
package/src/testkit/Layers.ts
CHANGED
|
@@ -206,6 +206,7 @@ const missingToolCommands: ToolCommands["Type"] = {
|
|
|
206
206
|
ui: missingTool("ui"),
|
|
207
207
|
lsp: missingTool("lsp"),
|
|
208
208
|
loop: missingTool("loop"),
|
|
209
|
+
contextHook: missingTool("context-hook"),
|
|
209
210
|
judge: () =>
|
|
210
211
|
Promise.reject(
|
|
211
212
|
new Error("gtd judge run: ToolCommands has no test double wired into testLayers()"),
|
package/src/testkit/steps.ts
CHANGED
|
@@ -26,6 +26,8 @@ export class TestkitWorld extends QuickPickleWorld {
|
|
|
26
26
|
/** Per step, what its scripted actor does, in the order the scenario declared it. */
|
|
27
27
|
actions = new Map<string, Actor[]>()
|
|
28
28
|
verdicts = new Map<string, string>()
|
|
29
|
+
/** Per judge step, a verdict for each text its questions may contain; the first match answers. */
|
|
30
|
+
conditionalVerdicts = new Map<string, [text: string, verdict: string][]>()
|
|
29
31
|
private created: TestRepo | undefined = undefined
|
|
30
32
|
|
|
31
33
|
// The repository is created on first use and keeps the config and
|
|
@@ -79,6 +81,19 @@ export class TestkitWorld extends QuickPickleWorld {
|
|
|
79
81
|
}
|
|
80
82
|
actors[step] = () => verdict
|
|
81
83
|
}
|
|
84
|
+
for (const [step, choices] of this.conditionalVerdicts) {
|
|
85
|
+
if (this.actions.has(step) || this.verdicts.has(step)) {
|
|
86
|
+
throw new Error(`step "${step}" has a conditional verdict and another script`)
|
|
87
|
+
}
|
|
88
|
+
// Matched against the pending questions, which `gtd judge` prints and the beat does not.
|
|
89
|
+
actors[step] = async (_, repo) => {
|
|
90
|
+
const { stdout } = await repo.gtd("judge")
|
|
91
|
+
const hit = choices.find(([text]) => stdout.includes(text))
|
|
92
|
+
if (hit === undefined)
|
|
93
|
+
throw new Error(`judge step "${step}" matched no verdict:\n${stdout}`)
|
|
94
|
+
return hit[1]
|
|
95
|
+
}
|
|
96
|
+
}
|
|
82
97
|
await this.repo().drive(actors)
|
|
83
98
|
}
|
|
84
99
|
}
|
|
@@ -137,6 +152,16 @@ Given(
|
|
|
137
152
|
},
|
|
138
153
|
)
|
|
139
154
|
|
|
155
|
+
// Lets one step write differently per visit, told apart by what it is prompted with.
|
|
156
|
+
Given(
|
|
157
|
+
"step {string}, when prompted with {string}, writes {string} with:",
|
|
158
|
+
(world: TestkitWorld, step: string, text: string, path: string, content: string) => {
|
|
159
|
+
world.act(step, (beat, repo) => {
|
|
160
|
+
if (beat.content.includes(text)) repo.write(path, withNewline(content))
|
|
161
|
+
})
|
|
162
|
+
},
|
|
163
|
+
)
|
|
164
|
+
|
|
140
165
|
// A standing assertion like any queued action: it applies on every later visit
|
|
141
166
|
// to the step until `from now on` clears the list.
|
|
142
167
|
Given(
|
|
@@ -169,6 +194,16 @@ Given("judge step {string} answers:", (world: TestkitWorld, step: string, verdic
|
|
|
169
194
|
world.verdicts.set(step, String(verdict))
|
|
170
195
|
})
|
|
171
196
|
|
|
197
|
+
Given(
|
|
198
|
+
"judge step {string}, when its questions contain {string}, answers:",
|
|
199
|
+
(world: TestkitWorld, step: string, text: string, verdict: string) => {
|
|
200
|
+
world.conditionalVerdicts.set(step, [
|
|
201
|
+
...(world.conditionalVerdicts.get(step) ?? []),
|
|
202
|
+
[text, String(verdict)],
|
|
203
|
+
])
|
|
204
|
+
},
|
|
205
|
+
)
|
|
206
|
+
|
|
172
207
|
When("gtd starts workflow {string}", async (world: TestkitWorld, workflow: string) => {
|
|
173
208
|
await world.repo().start(workflow)
|
|
174
209
|
})
|
package/src/wire/BeatDocument.ts
CHANGED
|
@@ -1,6 +1,14 @@
|
|
|
1
1
|
import type { BeatKind, Demand, DemandSession } from "./Demand.js"
|
|
2
2
|
import type { BeatStatus, NextMatch, StatusChange } from "./BeatStatus.js"
|
|
3
|
-
import type {
|
|
3
|
+
import type {
|
|
4
|
+
Actor,
|
|
5
|
+
ApprovalEntry,
|
|
6
|
+
ModelCost,
|
|
7
|
+
StateMode,
|
|
8
|
+
StateName,
|
|
9
|
+
StepAccess,
|
|
10
|
+
StepContextLimit,
|
|
11
|
+
} from "./types.js"
|
|
4
12
|
import { WIRE_SCHEMA } from "./constants.js"
|
|
5
13
|
|
|
6
14
|
/** `gtd next --json`'s `next` key — `null` when landing would commit nothing. */
|
|
@@ -56,6 +64,7 @@ export interface BeatDocument {
|
|
|
56
64
|
readonly session: DemandSession | undefined
|
|
57
65
|
readonly model: string | undefined
|
|
58
66
|
readonly system: string | undefined
|
|
67
|
+
readonly context: StepContextLimit | undefined
|
|
59
68
|
readonly validate: string | undefined
|
|
60
69
|
readonly log: string
|
|
61
70
|
readonly state: StateName
|
|
@@ -76,7 +85,7 @@ export interface BeatDocument {
|
|
|
76
85
|
}
|
|
77
86
|
|
|
78
87
|
/**
|
|
79
|
-
* Flatten a `Demand` plus its `BeatStatus` into the single
|
|
88
|
+
* Flatten a `Demand` plus its `BeatStatus` into the single 26-key document
|
|
80
89
|
* `gtd next --json` emits — the ONLY place the two are joined. `cost`/
|
|
81
90
|
* `costByModel` are omitted together, exactly when no cost was recorded
|
|
82
91
|
* (`cost <= 0`).
|
|
@@ -93,6 +102,7 @@ export const beatDocument = (demand: Demand, status: BeatStatus): BeatDocument =
|
|
|
93
102
|
session,
|
|
94
103
|
model: status.model,
|
|
95
104
|
system: status.system,
|
|
105
|
+
context: status.context,
|
|
96
106
|
validate,
|
|
97
107
|
log: status.log,
|
|
98
108
|
state: status.state,
|
package/src/wire/BeatStatus.ts
CHANGED
|
@@ -5,6 +5,7 @@ import type {
|
|
|
5
5
|
RenderedDemandSource,
|
|
6
6
|
StateMode,
|
|
7
7
|
StateName,
|
|
8
|
+
StepContextLimit,
|
|
8
9
|
} from "./types.js"
|
|
9
10
|
|
|
10
11
|
/** One pending change — `gtd next --json`'s `changes` entries. */
|
|
@@ -32,6 +33,7 @@ export interface BeatStatus {
|
|
|
32
33
|
readonly initial: boolean
|
|
33
34
|
readonly model: string | undefined
|
|
34
35
|
readonly system: string | undefined
|
|
36
|
+
readonly context: StepContextLimit | undefined
|
|
35
37
|
readonly log: string
|
|
36
38
|
readonly state: StateName
|
|
37
39
|
readonly actor: Actor
|
|
@@ -74,6 +76,7 @@ export const statusOf = (input: {
|
|
|
74
76
|
initial,
|
|
75
77
|
model: rendered.model,
|
|
76
78
|
system: rendered.system !== undefined && rendered.system !== "" ? rendered.system : undefined,
|
|
79
|
+
context: rendered.context,
|
|
77
80
|
log,
|
|
78
81
|
state: rendered.state,
|
|
79
82
|
actor: rendered.actor,
|
|
@@ -0,0 +1,30 @@
|
|
|
1
|
+
/** A Claude Code transcript whose latest assistant message used `tokens` of context, as the hook reads it from `transcript_path`. */
|
|
2
|
+
export const transcriptAt = (tokens: number): string =>
|
|
3
|
+
[
|
|
4
|
+
{ type: "user", message: { role: "user", content: "build it" } },
|
|
5
|
+
{
|
|
6
|
+
type: "assistant",
|
|
7
|
+
message: {
|
|
8
|
+
id: "msg_1",
|
|
9
|
+
role: "assistant",
|
|
10
|
+
usage: { input_tokens: 3, cache_read_input_tokens: 10, cache_creation_input_tokens: 7 },
|
|
11
|
+
},
|
|
12
|
+
},
|
|
13
|
+
{ type: "user", message: { role: "user", content: [{ type: "tool_result" }] } },
|
|
14
|
+
{
|
|
15
|
+
type: "assistant",
|
|
16
|
+
message: {
|
|
17
|
+
id: "msg_2",
|
|
18
|
+
role: "assistant",
|
|
19
|
+
usage: {
|
|
20
|
+
input_tokens: 1,
|
|
21
|
+
cache_read_input_tokens: tokens - 1001,
|
|
22
|
+
cache_creation_input_tokens: 1000,
|
|
23
|
+
output_tokens: 500,
|
|
24
|
+
},
|
|
25
|
+
},
|
|
26
|
+
},
|
|
27
|
+
{ type: "attachment", attachment: { type: "hook_success" } },
|
|
28
|
+
]
|
|
29
|
+
.map((entry) => JSON.stringify(entry))
|
|
30
|
+
.join("\n") + "\n"
|