@dungle-scrubs/harness-cli-normalizer 0.3.1 → 0.4.0
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 +64 -86
- package/dist/cli/args.d.ts +11 -0
- package/dist/cli/args.d.ts.map +1 -1
- package/dist/cli/args.js +23 -5
- package/dist/cli/args.js.map +1 -1
- package/dist/cli/config.d.ts +25 -0
- package/dist/cli/config.d.ts.map +1 -0
- package/dist/cli/config.js +139 -0
- package/dist/cli/config.js.map +1 -0
- package/dist/cli/inspect.d.ts.map +1 -1
- package/dist/cli/inspect.js +35 -1
- package/dist/cli/inspect.js.map +1 -1
- package/dist/cli/render.d.ts.map +1 -1
- package/dist/cli/render.js +10 -1
- package/dist/cli/render.js.map +1 -1
- package/dist/cli/run.d.ts.map +1 -1
- package/dist/cli/run.js +134 -4
- package/dist/cli/run.js.map +1 -1
- package/dist/execution/failure.d.ts +21 -4
- package/dist/execution/failure.d.ts.map +1 -1
- package/dist/execution/failure.js +26 -3
- package/dist/execution/failure.js.map +1 -1
- package/dist/execution/stream-turn.d.ts +4 -0
- package/dist/execution/stream-turn.d.ts.map +1 -1
- package/dist/execution/stream-turn.js +28 -4
- package/dist/execution/stream-turn.js.map +1 -1
- package/dist/interpretation/argv.d.ts +1 -0
- package/dist/interpretation/argv.d.ts.map +1 -1
- package/dist/interpretation/argv.js +13 -19
- package/dist/interpretation/argv.js.map +1 -1
- package/dist/interpretation/hints.d.ts +22 -0
- package/dist/interpretation/hints.d.ts.map +1 -0
- package/dist/interpretation/hints.js +59 -0
- package/dist/interpretation/hints.js.map +1 -0
- package/dist/interpretation/refusal.d.ts +25 -4
- package/dist/interpretation/refusal.d.ts.map +1 -1
- package/dist/interpretation/refusal.js +16 -0
- package/dist/interpretation/refusal.js.map +1 -1
- package/dist/interpretation/resolve-options.d.ts +42 -0
- package/dist/interpretation/resolve-options.d.ts.map +1 -0
- package/dist/interpretation/resolve-options.js +140 -0
- package/dist/interpretation/resolve-options.js.map +1 -0
- package/dist/interpretation/support.d.ts +32 -0
- package/dist/interpretation/support.d.ts.map +1 -0
- package/dist/interpretation/support.js +107 -0
- package/dist/interpretation/support.js.map +1 -0
- package/dist/interpretation/tool-selection.d.ts +29 -0
- package/dist/interpretation/tool-selection.d.ts.map +1 -0
- package/dist/interpretation/tool-selection.js +122 -0
- package/dist/interpretation/tool-selection.js.map +1 -0
- package/dist/interpretation/turn-options.d.ts.map +1 -1
- package/dist/interpretation/turn-options.js +4 -0
- package/dist/interpretation/turn-options.js.map +1 -1
- package/dist/knowledge/claude-code.d.ts.map +1 -1
- package/dist/knowledge/claude-code.js +34 -1
- package/dist/knowledge/claude-code.js.map +1 -1
- package/dist/knowledge/codex.d.ts.map +1 -1
- package/dist/knowledge/codex.js +19 -0
- package/dist/knowledge/codex.js.map +1 -1
- package/dist/knowledge/descriptor.d.ts +32 -0
- package/dist/knowledge/descriptor.d.ts.map +1 -1
- package/dist/knowledge/muse.d.ts.map +1 -1
- package/dist/knowledge/muse.js +19 -0
- package/dist/knowledge/muse.js.map +1 -1
- package/dist/knowledge/pi.d.ts.map +1 -1
- package/dist/knowledge/pi.js +21 -0
- package/dist/knowledge/pi.js.map +1 -1
- package/dist/knowledge/profile.d.ts +13 -0
- package/dist/knowledge/profile.d.ts.map +1 -0
- package/dist/knowledge/profile.js +31 -0
- package/dist/knowledge/profile.js.map +1 -0
- package/package.json +2 -23
- package/src/cli/args.ts +31 -1
- package/src/cli/config.ts +166 -0
- package/src/cli/inspect.ts +34 -1
- package/src/cli/render.ts +10 -1
- package/src/cli/run.ts +161 -5
- package/src/execution/failure.ts +44 -5
- package/src/execution/stream-turn.ts +33 -3
- package/src/interpretation/argv.ts +14 -19
- package/src/interpretation/hints.ts +99 -0
- package/src/interpretation/refusal.ts +32 -3
- package/src/interpretation/resolve-options.ts +190 -0
- package/src/interpretation/support.ts +119 -0
- package/src/interpretation/tool-selection.ts +161 -0
- package/src/interpretation/turn-options.ts +4 -0
- package/src/knowledge/claude-code.ts +34 -1
- package/src/knowledge/codex.ts +19 -0
- package/src/knowledge/descriptor.ts +32 -0
- package/src/knowledge/muse.ts +19 -0
- package/src/knowledge/pi.ts +21 -0
- package/src/knowledge/profile.ts +33 -0
- package/dist/index.d.ts +0 -8
- package/dist/index.d.ts.map +0 -1
- package/dist/index.js +0 -8
- package/dist/index.js.map +0 -1
- package/src/index.ts +0 -7
|
@@ -17,19 +17,32 @@ export const REFUSAL_ISSUES = deepFreeze([
|
|
|
17
17
|
"unknown-model",
|
|
18
18
|
"invalid-env",
|
|
19
19
|
"invalid-tool-grant",
|
|
20
|
+
"unknown-tool-name",
|
|
21
|
+
"mutually-exclusive-options",
|
|
20
22
|
"prompt-flag-injection",
|
|
21
23
|
"no-autonomy-mode",
|
|
22
24
|
"no-session-mode",
|
|
23
25
|
] as const);
|
|
24
26
|
export type RefusalIssue = (typeof REFUSAL_ISSUES)[number];
|
|
25
27
|
|
|
28
|
+
/** The option a refusal names: turn-option spec keys (descriptor tables)
|
|
29
|
+
* plus the tool-list dimensions and autonomy, which render via dedicated
|
|
30
|
+
* descriptor fields rather than a spec table. Kept closed so consumers can
|
|
31
|
+
* branch on it without a default arm. */
|
|
32
|
+
export type RefusalOption =
|
|
33
|
+
| TurnOptionKey
|
|
34
|
+
| "tools"
|
|
35
|
+
| "excludeTools"
|
|
36
|
+
| "autonomy"
|
|
37
|
+
| `discovery.${string}`;
|
|
38
|
+
|
|
26
39
|
/** One helper builds the message from the structured fields so message and
|
|
27
40
|
* fields cannot drift. Every message names an alternative, not only a
|
|
28
41
|
* negation, so an agent can pivot without reading the descriptor. */
|
|
29
42
|
export const buildRefusalMessage = (
|
|
30
43
|
issue: RefusalIssue,
|
|
31
44
|
harness: HarnessName,
|
|
32
|
-
option?:
|
|
45
|
+
option?: RefusalOption,
|
|
33
46
|
facet?: DiscoveryFacet,
|
|
34
47
|
supported: readonly string[] = [],
|
|
35
48
|
detail?: string,
|
|
@@ -66,6 +79,10 @@ export const buildRefusalMessage = (
|
|
|
66
79
|
return `invalid env key or value for ${harness}${detailSuffix}; ${supportedStr} - keys must match ^[A-Za-z_][A-Za-z0-9_]*$ and contain no NUL`;
|
|
67
80
|
case "invalid-tool-grant":
|
|
68
81
|
return `tool grant for ${harness} contains an empty entry or a comma; a blank tool flag value grants nothing detectable, and a comma inside one name silently splits the grant; ${supportedStr} - provide comma-free, non-empty tool names as separate entries`;
|
|
82
|
+
case "unknown-tool-name":
|
|
83
|
+
return `${harness} cannot compute a tool complement around an unknown name${detailSuffix}; ${supportedStr} - exclude only curated names, or pass the unknown name through an include list instead`;
|
|
84
|
+
case "mutually-exclusive-options":
|
|
85
|
+
return `${harness} cannot combine${optionPart}${detailSuffix}; ${supportedStr} - pass exactly one of them`;
|
|
69
86
|
case "prompt-flag-injection":
|
|
70
87
|
return `positional prompt may not start with '-'; it would be parsed as a flag by ${harness}${detailSuffix}; ${supportedStr} - remove leading '-' or prefix with a space`;
|
|
71
88
|
case "no-autonomy-mode":
|
|
@@ -83,15 +100,25 @@ export const buildRefusalMessage = (
|
|
|
83
100
|
export class ArgvRefusalError extends Error {
|
|
84
101
|
readonly issue: RefusalIssue;
|
|
85
102
|
readonly harness: HarnessName;
|
|
86
|
-
readonly option?:
|
|
103
|
+
readonly option?: RefusalOption;
|
|
87
104
|
readonly facet?: DiscoveryFacet;
|
|
88
105
|
readonly supported: readonly string[];
|
|
106
|
+
/** D7: which harnesses DO express the refused option, native spellings
|
|
107
|
+
* included. Derived by the raise site from descriptors - absent when the
|
|
108
|
+
* refusing layer has no descriptor set in scope. */
|
|
109
|
+
readonly supportedBy?: ReadonlyArray<{ harness: string; spelling: string }>;
|
|
110
|
+
/** D8: nearest-alternative suggestion for the CURRENT harness - keeps a
|
|
111
|
+
* scanning agent on its chosen harness instead of switching. Curatorial
|
|
112
|
+
* data set at the raise site; absent when no hint exists. */
|
|
113
|
+
readonly hint?: string;
|
|
89
114
|
constructor(args: {
|
|
90
115
|
readonly issue: RefusalIssue;
|
|
91
116
|
readonly harness: HarnessName;
|
|
92
|
-
readonly option?:
|
|
117
|
+
readonly option?: RefusalOption;
|
|
93
118
|
readonly facet?: DiscoveryFacet;
|
|
94
119
|
readonly supported?: readonly string[];
|
|
120
|
+
readonly supportedBy?: ReadonlyArray<{ harness: string; spelling: string }>;
|
|
121
|
+
readonly hint?: string;
|
|
95
122
|
readonly detail?: string;
|
|
96
123
|
readonly message?: string;
|
|
97
124
|
}) {
|
|
@@ -112,5 +139,7 @@ export class ArgvRefusalError extends Error {
|
|
|
112
139
|
this.option = args.option;
|
|
113
140
|
this.facet = args.facet;
|
|
114
141
|
this.supported = args.supported ?? [];
|
|
142
|
+
this.supportedBy = args.supportedBy;
|
|
143
|
+
this.hint = args.hint;
|
|
115
144
|
}
|
|
116
145
|
}
|
|
@@ -0,0 +1,190 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Option resolution: the precedence chain args > user config > built-in
|
|
3
|
+
* profile, plus provenance. Pure - the config FILE is read by the CLI layer
|
|
4
|
+
* and passed in as data; this layer only decides what wins.
|
|
5
|
+
*
|
|
6
|
+
* Launch-only: callers apply resolved options on launch, never resume (a
|
|
7
|
+
* resumed session keeps its session's settings - same rule the codex
|
|
8
|
+
* sandbox default already follows).
|
|
9
|
+
*/
|
|
10
|
+
import type { HarnessDescriptor } from "../knowledge/descriptor.js";
|
|
11
|
+
import { DEFAULT_TURN_PROFILE, type ProfileKey } from "../knowledge/profile.js";
|
|
12
|
+
import type { TurnOptions } from "./argv.js";
|
|
13
|
+
|
|
14
|
+
export type ProvenanceTier = "arg" | "project-config" | "user-config" | "profile" | "harness";
|
|
15
|
+
|
|
16
|
+
export interface ProvenanceEntry {
|
|
17
|
+
readonly key: string;
|
|
18
|
+
readonly value: unknown;
|
|
19
|
+
readonly tier: ProvenanceTier;
|
|
20
|
+
}
|
|
21
|
+
|
|
22
|
+
/** D5: the project floor caps arg grants; exceeding it refuses naming both
|
|
23
|
+
* sets. Structured-first: fields, not prose parsing. */
|
|
24
|
+
export class FloorExceededError extends Error {
|
|
25
|
+
readonly harness: string;
|
|
26
|
+
readonly excess: readonly string[];
|
|
27
|
+
readonly floor: readonly string[];
|
|
28
|
+
constructor(harness: string, excess: readonly string[], floor: readonly string[]) {
|
|
29
|
+
super(
|
|
30
|
+
`tool grant exceeds the project floor: ${JSON.stringify(excess)} not in floor ${JSON.stringify(floor)} - request a grant within the floor or raise the floor in the project config`,
|
|
31
|
+
);
|
|
32
|
+
this.name = "FloorExceededError";
|
|
33
|
+
this.harness = harness;
|
|
34
|
+
this.excess = excess;
|
|
35
|
+
this.floor = floor;
|
|
36
|
+
}
|
|
37
|
+
}
|
|
38
|
+
|
|
39
|
+
export interface ResolvedOptions {
|
|
40
|
+
readonly options: TurnOptions;
|
|
41
|
+
readonly provenance: readonly ProvenanceEntry[];
|
|
42
|
+
/** Profile dimensions this harness cannot express (skip-and-report,
|
|
43
|
+
* never refuse): rendered as divergence, not failure. */
|
|
44
|
+
readonly unrenderable: readonly string[];
|
|
45
|
+
}
|
|
46
|
+
|
|
47
|
+
/** Expressibility per profile dimension. Dimensions whose "on" state is
|
|
48
|
+
* the harness's own default (discovery all-on) or whose "off" state emits
|
|
49
|
+
* nothing (autonomy false) are expressible EVERYWHERE - the profile value
|
|
50
|
+
* resolves to "emit nothing," which every harness can do. Divergence is
|
|
51
|
+
* reserved for dimensions that would emit a flag the harness lacks. */
|
|
52
|
+
const EXPRESSIBLE: Readonly<Record<ProfileKey, (h: HarnessDescriptor) => boolean>> = {
|
|
53
|
+
effort: (h) => h.turnOptions.effort !== undefined,
|
|
54
|
+
sandbox: (h) => h.turnOptions.sandbox !== undefined,
|
|
55
|
+
discovery: () => true,
|
|
56
|
+
autonomy: () => true,
|
|
57
|
+
};
|
|
58
|
+
|
|
59
|
+
export interface ConfigTiers {
|
|
60
|
+
/** ~/.config/hcn/config.json (XDG) - machine-wide defaults. */
|
|
61
|
+
readonly user?: Readonly<Partial<TurnOptions>>;
|
|
62
|
+
/** <git-root>/.hcn/config.json - auto-discovered (ratified A), the ALL-
|
|
63
|
+
* OFF tier; its `tools` key is both the default grant and the FLOOR: an
|
|
64
|
+
* arg grant exceeding it refuses, naming both sets (D5). */
|
|
65
|
+
readonly project?: Readonly<Partial<TurnOptions>>;
|
|
66
|
+
}
|
|
67
|
+
|
|
68
|
+
/** Merge semantics (gap 1, resolved): config keys are scalars and lists in
|
|
69
|
+
* schema v1 - there is nothing to deep-merge INTO - so precedence is whole-
|
|
70
|
+
* key replacement: arg > project > user > profile. A future nested key
|
|
71
|
+
* (per-harness sections) ships with schema v2 and its own merge rule. */
|
|
72
|
+
const effectiveConfig = (tiers: ConfigTiers): Readonly<Partial<TurnOptions>> => ({
|
|
73
|
+
...(tiers.user ?? {}),
|
|
74
|
+
...(tiers.project ?? {}),
|
|
75
|
+
});
|
|
76
|
+
|
|
77
|
+
/** Resolve the effective options for a LAUNCH. `args` is what the caller
|
|
78
|
+
* passed explicitly (highest tier); `userConfig` the parsed config file;
|
|
79
|
+
* the profile is the floor. Keys absent everywhere defer to the harness
|
|
80
|
+
* and are reported with tier "harness" only when something (config or
|
|
81
|
+
* profile) attempted them - a truly untouched dimension is nobody's
|
|
82
|
+
* business and appears in provenance as tier "harness" with no value. */
|
|
83
|
+
/** A discovery value emits nothing when every facet is true (on). */
|
|
84
|
+
const emitsNothing = (value: unknown): boolean =>
|
|
85
|
+
typeof value === "object" &&
|
|
86
|
+
value !== null &&
|
|
87
|
+
Object.values(value as Record<string, unknown>).every((v) => v === true);
|
|
88
|
+
|
|
89
|
+
export const resolveEffectiveOptions = (
|
|
90
|
+
h: HarnessDescriptor,
|
|
91
|
+
args: TurnOptions,
|
|
92
|
+
tiers: ConfigTiers = {},
|
|
93
|
+
): ResolvedOptions => {
|
|
94
|
+
const provenance: ProvenanceEntry[] = [];
|
|
95
|
+
const unrenderable: string[] = [];
|
|
96
|
+
const config = effectiveConfig(tiers);
|
|
97
|
+
const sourceTier = (key: string): ProvenanceTier | undefined => {
|
|
98
|
+
if (tiers.project?.[key as keyof TurnOptions] !== undefined) return "project-config";
|
|
99
|
+
if (tiers.user?.[key as keyof TurnOptions] !== undefined) return "user-config";
|
|
100
|
+
return undefined;
|
|
101
|
+
};
|
|
102
|
+
|
|
103
|
+
// D5 toolset expansion: a bare --tools name matching a configured
|
|
104
|
+
// toolset resolves to its list BEFORE the floor check, so a named set
|
|
105
|
+
// within the floor passes and one exceeding it refuses naming the set's
|
|
106
|
+
// members. Project toolsets win name collisions over user toolsets.
|
|
107
|
+
const toolsets = {
|
|
108
|
+
...((tiers.user as { toolsets?: Record<string, string[]> } | undefined)?.toolsets ?? {}),
|
|
109
|
+
...((tiers.project as { toolsets?: Record<string, string[]> } | undefined)?.toolsets ?? {}),
|
|
110
|
+
};
|
|
111
|
+
let effectiveArgs = args;
|
|
112
|
+
if (
|
|
113
|
+
args.tools !== undefined &&
|
|
114
|
+
args.tools.length === 1 &&
|
|
115
|
+
toolsets[args.tools[0] as string] !== undefined
|
|
116
|
+
) {
|
|
117
|
+
effectiveArgs = { ...args, tools: toolsets[args.tools[0] as string] };
|
|
118
|
+
provenance.push({
|
|
119
|
+
key: "tools",
|
|
120
|
+
value: effectiveArgs.tools,
|
|
121
|
+
tier: "arg",
|
|
122
|
+
});
|
|
123
|
+
}
|
|
124
|
+
const resolved: Record<string, unknown> = { ...effectiveArgs };
|
|
125
|
+
|
|
126
|
+
// D5 floor: a project toolset floor caps any arg grant; exceeding it is
|
|
127
|
+
// a structured refusal naming both sets - never a silent clamp.
|
|
128
|
+
const floor = tiers.project?.tools;
|
|
129
|
+
if (floor !== undefined && effectiveArgs.tools !== undefined) {
|
|
130
|
+
const floorSet = new Set(floor);
|
|
131
|
+
const excess = effectiveArgs.tools.filter((t) => !floorSet.has(t));
|
|
132
|
+
if (excess.length > 0) {
|
|
133
|
+
throw new FloorExceededError(h.name, excess, [...floor]);
|
|
134
|
+
}
|
|
135
|
+
}
|
|
136
|
+
|
|
137
|
+
// Profile is the floor: apply only where nothing above it set the key.
|
|
138
|
+
for (const [key, value] of Object.entries(DEFAULT_TURN_PROFILE)) {
|
|
139
|
+
const argsSet = effectiveArgs[key as keyof TurnOptions] !== undefined;
|
|
140
|
+
const tier = sourceTier(key);
|
|
141
|
+
if (argsSet) {
|
|
142
|
+
provenance.push({ key, value: effectiveArgs[key as keyof TurnOptions], tier: "arg" });
|
|
143
|
+
continue;
|
|
144
|
+
}
|
|
145
|
+
if (tier !== undefined) {
|
|
146
|
+
provenance.push({ key, value: config[key as keyof TurnOptions], tier });
|
|
147
|
+
resolved[key] = config[key as keyof TurnOptions];
|
|
148
|
+
continue;
|
|
149
|
+
}
|
|
150
|
+
const expressible = EXPRESSIBLE[key as ProfileKey]?.(h) ?? false;
|
|
151
|
+
if (!expressible) {
|
|
152
|
+
// Skip-and-report: a profile default this harness cannot express is
|
|
153
|
+
// reported divergence, never a refusal and never silence.
|
|
154
|
+
unrenderable.push(key);
|
|
155
|
+
provenance.push({ key, value, tier: "harness" });
|
|
156
|
+
continue;
|
|
157
|
+
}
|
|
158
|
+
// Dimensions whose value reduces to "emit nothing" (autonomy false,
|
|
159
|
+
// discovery all-on) stay ABSENT from the resolved options - the
|
|
160
|
+
// harness's default already satisfies the profile, and emitting
|
|
161
|
+
// explicit on-flags would change resume grammar and add breakage
|
|
162
|
+
// surface for no semantic gain. Provenance still records the tier.
|
|
163
|
+
if ((key === "autonomy" && value === false) || (key === "discovery" && emitsNothing(value))) {
|
|
164
|
+
provenance.push({ key, value, tier: "profile" });
|
|
165
|
+
continue;
|
|
166
|
+
}
|
|
167
|
+
resolved[key] = value;
|
|
168
|
+
provenance.push({ key, value, tier: "profile" });
|
|
169
|
+
}
|
|
170
|
+
|
|
171
|
+
// Config keys outside the profile surface pass through at their own tier
|
|
172
|
+
// (validated later by the same renderers as args).
|
|
173
|
+
for (const [key, value] of Object.entries(config)) {
|
|
174
|
+
if (key === "toolsets") continue; // expanded into args above, never a turn option
|
|
175
|
+
if (key in DEFAULT_TURN_PROFILE) continue;
|
|
176
|
+
if (effectiveArgs[key as keyof TurnOptions] !== undefined) {
|
|
177
|
+
provenance.push({ key, value: effectiveArgs[key as keyof TurnOptions], tier: "arg" });
|
|
178
|
+
continue;
|
|
179
|
+
}
|
|
180
|
+
resolved[key] = value;
|
|
181
|
+
const tier = sourceTier(key) ?? "user-config";
|
|
182
|
+
provenance.push({ key, value, tier });
|
|
183
|
+
}
|
|
184
|
+
|
|
185
|
+
return {
|
|
186
|
+
options: resolved as unknown as TurnOptions,
|
|
187
|
+
provenance,
|
|
188
|
+
unrenderable,
|
|
189
|
+
};
|
|
190
|
+
};
|
|
@@ -0,0 +1,119 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Cross-harness support derivation: pure functions that answer "which
|
|
3
|
+
* harnesses express this option, and under what native spelling" from the
|
|
4
|
+
* descriptor set. Phase 3 (D7): refusal diagnostics derive support lists at
|
|
5
|
+
* runtime instead of hardcoding them, so a descriptor edit can never leave
|
|
6
|
+
* a refusal message stale. The autonomy refusal in argv.ts was the
|
|
7
|
+
* counter-pattern - a hardcoded flag array that drifts on descriptor
|
|
8
|
+
* change.
|
|
9
|
+
*
|
|
10
|
+
* Interpretation-layer pure function; takes the descriptor set as an
|
|
11
|
+
* argument rather than importing the defaults, so override sets work.
|
|
12
|
+
*/
|
|
13
|
+
|
|
14
|
+
import type { HarnessDescriptor } from "../knowledge/descriptor.js";
|
|
15
|
+
import type { DescriptorSet } from "../knowledge/overrides.js";
|
|
16
|
+
import type { RefusalOption } from "./refusal.js";
|
|
17
|
+
|
|
18
|
+
export interface SupportEntry {
|
|
19
|
+
readonly harness: string;
|
|
20
|
+
/** The native spelling a caller would use directly on that harness. */
|
|
21
|
+
readonly spelling: string;
|
|
22
|
+
}
|
|
23
|
+
|
|
24
|
+
const spellingOf = (h: HarnessDescriptor, option: RefusalOption): string | null => {
|
|
25
|
+
switch (option) {
|
|
26
|
+
case "tools":
|
|
27
|
+
return h.tools.includeFlag;
|
|
28
|
+
case "excludeTools":
|
|
29
|
+
return h.tools.excludeFlag;
|
|
30
|
+
case "autonomy":
|
|
31
|
+
return h.autonomy?.flag ?? null;
|
|
32
|
+
case "effort":
|
|
33
|
+
case "sandbox":
|
|
34
|
+
case "provider":
|
|
35
|
+
case "write":
|
|
36
|
+
case "shell":
|
|
37
|
+
case "maxSteps":
|
|
38
|
+
case "discovery": {
|
|
39
|
+
const spec = h.turnOptions[option];
|
|
40
|
+
if (spec === undefined) return null;
|
|
41
|
+
const rawRender =
|
|
42
|
+
spec.kind === "discovery"
|
|
43
|
+
? (
|
|
44
|
+
Object.values(spec.facets)[0] as
|
|
45
|
+
| { render?: { flag?: string; flags?: string[] } }
|
|
46
|
+
| undefined
|
|
47
|
+
)?.render
|
|
48
|
+
: (spec as { render?: { flag?: string; flags?: string[] } }).render;
|
|
49
|
+
if (rawRender === undefined) return null;
|
|
50
|
+
// flag-value/config-kv carry `flag`; flag-list carries `flags` - the
|
|
51
|
+
// first entry is the spelling a caller would type.
|
|
52
|
+
return rawRender.flag ?? rawRender.flags?.[0] ?? null;
|
|
53
|
+
}
|
|
54
|
+
default:
|
|
55
|
+
return null;
|
|
56
|
+
}
|
|
57
|
+
};
|
|
58
|
+
|
|
59
|
+
/** Every harness in the set that can express `option`, with its native
|
|
60
|
+
* spelling. Ordered by the set's insertion order (defaults: claude, codex,
|
|
61
|
+
* pi, muse) so output is stable across calls. */
|
|
62
|
+
export const supportedBy = (set: DescriptorSet, option: RefusalOption): readonly SupportEntry[] => {
|
|
63
|
+
const out: SupportEntry[] = [];
|
|
64
|
+
for (const h of Object.values(set)) {
|
|
65
|
+
if (h === undefined) continue;
|
|
66
|
+
const spelling = spellingOf(h, option);
|
|
67
|
+
if (spelling !== null) out.push({ harness: h.name, spelling });
|
|
68
|
+
}
|
|
69
|
+
return out;
|
|
70
|
+
};
|
|
71
|
+
|
|
72
|
+
/** Reverse lookup for native-spelling recognition (D7 part B): given a raw
|
|
73
|
+
* flag token a caller typed, find the option it belongs to and which
|
|
74
|
+
* harnesses spell it that way. Returns null for tokens no descriptor
|
|
75
|
+
* knows - those keep the plain unknown-flag error. */
|
|
76
|
+
export const recognizeNativeSpelling = (
|
|
77
|
+
set: DescriptorSet,
|
|
78
|
+
flag: string,
|
|
79
|
+
): { readonly option: RefusalOption; readonly entries: readonly SupportEntry[] } | null => {
|
|
80
|
+
const normalized = flag.toLowerCase();
|
|
81
|
+
const candidates: RefusalOption[] = [
|
|
82
|
+
"tools",
|
|
83
|
+
"excludeTools",
|
|
84
|
+
"autonomy",
|
|
85
|
+
"effort",
|
|
86
|
+
"sandbox",
|
|
87
|
+
"provider",
|
|
88
|
+
"write",
|
|
89
|
+
"shell",
|
|
90
|
+
"maxSteps",
|
|
91
|
+
];
|
|
92
|
+
for (const option of candidates) {
|
|
93
|
+
const entries = supportedBy(set, option).filter((e) => e.spelling.toLowerCase() === normalized);
|
|
94
|
+
if (entries.length > 0) return { option, entries };
|
|
95
|
+
}
|
|
96
|
+
// Discovery facets: single-dash spellings (pi -nt/-nc/-ne/-ns) and
|
|
97
|
+
// claude's --setting-sources. Facet spellings map to the facet name so
|
|
98
|
+
// the redirect can name the normalized spelling.
|
|
99
|
+
const FACET_KEYS = ["tools", "instructionFiles", "extensions", "skills"] as const;
|
|
100
|
+
for (const facet of FACET_KEYS) {
|
|
101
|
+
const entries: SupportEntry[] = [];
|
|
102
|
+
for (const h of Object.values(set)) {
|
|
103
|
+
if (h === undefined) continue;
|
|
104
|
+
const spec = h.turnOptions.discovery;
|
|
105
|
+
if (spec === undefined || spec.kind !== "discovery") continue;
|
|
106
|
+
const facetSpec = spec.facets[facet];
|
|
107
|
+
if (facetSpec === undefined) continue;
|
|
108
|
+
const render = facetSpec.render as { flag?: string; flags?: string[] };
|
|
109
|
+
const spelling = render.flag ?? render.flags?.[0];
|
|
110
|
+
if (spelling !== undefined && spelling.toLowerCase() === normalized) {
|
|
111
|
+
entries.push({ harness: h.name, spelling });
|
|
112
|
+
}
|
|
113
|
+
}
|
|
114
|
+
if (entries.length > 0) {
|
|
115
|
+
return { option: `discovery.${facet}` as RefusalOption, entries };
|
|
116
|
+
}
|
|
117
|
+
}
|
|
118
|
+
return null;
|
|
119
|
+
};
|
|
@@ -0,0 +1,161 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Tool-selection rendering: turns normalized include/exclude lists into the
|
|
3
|
+
* per-harness argv tokens, per decisions D1-D3 and the Phase 0 evidence.
|
|
4
|
+
*
|
|
5
|
+
* Semantics:
|
|
6
|
+
* - include: exact allowlist over the descriptor's curated names. pi's flag
|
|
7
|
+
* is strict over built-ins, rendered directly. claude's include flag only
|
|
8
|
+
* pre-approves without restricting the visible set, so an exact allowlist
|
|
9
|
+
* on claude renders as a deny-complement (all known minus included).
|
|
10
|
+
* - exclude: the complement of the named tools over all descriptor-known
|
|
11
|
+
* names (D2). pi's native exclude subtracts from its default set (4
|
|
12
|
+
* tools), so a D2 exclude renders as a computed include list there.
|
|
13
|
+
* - mutual exclusion (D1): both flags in one call refuse.
|
|
14
|
+
* - extensible rule (D3): curated names validate and map; unknown
|
|
15
|
+
* clean-selector names pass through and are reported as unmapped so
|
|
16
|
+
* provenance can surface them.
|
|
17
|
+
*/
|
|
18
|
+
import type { HarnessDescriptor } from "../knowledge/descriptor.js";
|
|
19
|
+
import { defaultDescriptors } from "../knowledge/overrides.js";
|
|
20
|
+
import { ArgvRefusalError } from "./refusal.js";
|
|
21
|
+
import { supportedBy } from "./support.js";
|
|
22
|
+
|
|
23
|
+
export const TOOL_SELECTOR = /^[A-Za-z0-9][A-Za-z0-9._:/@-]{0,127}$/;
|
|
24
|
+
|
|
25
|
+
export interface ToolSelection {
|
|
26
|
+
readonly include?: readonly string[];
|
|
27
|
+
readonly exclude?: readonly string[];
|
|
28
|
+
}
|
|
29
|
+
|
|
30
|
+
export interface RenderedToolSelection {
|
|
31
|
+
readonly tokens: readonly string[];
|
|
32
|
+
readonly unmapped: readonly string[];
|
|
33
|
+
}
|
|
34
|
+
|
|
35
|
+
const validateNames = (h: HarnessDescriptor, names: readonly string[]): string[] => {
|
|
36
|
+
for (const name of names) {
|
|
37
|
+
if (name.trim() === "" || name.includes(",")) {
|
|
38
|
+
throw new ArgvRefusalError({
|
|
39
|
+
issue: "invalid-tool-grant",
|
|
40
|
+
harness: h.name,
|
|
41
|
+
supported: ["non-empty, comma-free tool names"],
|
|
42
|
+
detail: `tools=${JSON.stringify(names)}`,
|
|
43
|
+
});
|
|
44
|
+
}
|
|
45
|
+
if (!TOOL_SELECTOR.test(name)) {
|
|
46
|
+
throw new ArgvRefusalError({
|
|
47
|
+
issue: "invalid-tool-grant",
|
|
48
|
+
harness: h.name,
|
|
49
|
+
supported: [`must match ${TOOL_SELECTOR.source}`],
|
|
50
|
+
detail: name,
|
|
51
|
+
});
|
|
52
|
+
}
|
|
53
|
+
}
|
|
54
|
+
return [...names];
|
|
55
|
+
};
|
|
56
|
+
|
|
57
|
+
/** Curated-name mapping plus extensible pass-through; returns the effective
|
|
58
|
+
* names and which of them the descriptor could not vouch for. */
|
|
59
|
+
const resolveNames = (
|
|
60
|
+
h: HarnessDescriptor,
|
|
61
|
+
names: readonly string[],
|
|
62
|
+
): { mapped: string[]; unmapped: string[] } => {
|
|
63
|
+
const known = new Set(h.tools.builtins.map((t) => t.name));
|
|
64
|
+
const mapped: string[] = [];
|
|
65
|
+
const unmapped: string[] = [];
|
|
66
|
+
for (const name of names) {
|
|
67
|
+
if (known.has(name)) mapped.push(name);
|
|
68
|
+
else unmapped.push(name);
|
|
69
|
+
}
|
|
70
|
+
return { mapped, unmapped };
|
|
71
|
+
};
|
|
72
|
+
|
|
73
|
+
export const renderToolSelection = (
|
|
74
|
+
h: HarnessDescriptor,
|
|
75
|
+
selection: ToolSelection,
|
|
76
|
+
): RenderedToolSelection => {
|
|
77
|
+
const hasInclude = selection.include !== undefined;
|
|
78
|
+
const hasExclude = selection.exclude !== undefined;
|
|
79
|
+
|
|
80
|
+
// D1: mutual exclusivity, structured refusal.
|
|
81
|
+
if (hasInclude && hasExclude) {
|
|
82
|
+
throw new ArgvRefusalError({
|
|
83
|
+
issue: "mutually-exclusive-options",
|
|
84
|
+
harness: h.name,
|
|
85
|
+
option: "tools",
|
|
86
|
+
supported: ["--tools (exact allowlist) or --exclude-tools (complement), never both"],
|
|
87
|
+
detail: "mutual exclusion",
|
|
88
|
+
});
|
|
89
|
+
}
|
|
90
|
+
if (!hasInclude && !hasExclude) return { tokens: [], unmapped: [] };
|
|
91
|
+
|
|
92
|
+
if (h.tools.includeFlag === null || h.tools.excludeFlag === null) {
|
|
93
|
+
const option = hasInclude ? "tools" : "excludeTools";
|
|
94
|
+
const by = supportedBy(defaultDescriptors(), option);
|
|
95
|
+
throw new ArgvRefusalError({
|
|
96
|
+
issue: "unsupported-option",
|
|
97
|
+
harness: h.name,
|
|
98
|
+
option,
|
|
99
|
+
supported: ["per-tool name lists"],
|
|
100
|
+
supportedBy: by,
|
|
101
|
+
hint:
|
|
102
|
+
h.name === "codex"
|
|
103
|
+
? "nearest control on codex: category switches via config keys (features.shell_tool, web_search) or sandbox modes - see `hcn inspect codex`"
|
|
104
|
+
: "nearest control on muse: category switches (--disable-write, --disable-shell, --disable-web-tools) gate tool execution per session",
|
|
105
|
+
});
|
|
106
|
+
}
|
|
107
|
+
|
|
108
|
+
if (hasInclude) {
|
|
109
|
+
const names = validateNames(h, selection.include!);
|
|
110
|
+
const { mapped, unmapped } = resolveNames(h, names);
|
|
111
|
+
if (!h.tools.includeIsStrictAllowlist) {
|
|
112
|
+
// claude: exact allowlist must reshape the visible set via the deny
|
|
113
|
+
// complement (probe 2b). The include flag also carries the granted
|
|
114
|
+
// names (curated + pass-throughs) so they skip approval prompts;
|
|
115
|
+
// unmapped names ride along - the deny complement is computed over
|
|
116
|
+
// curated names only, so an extension tool in an include is granted,
|
|
117
|
+
// never denied.
|
|
118
|
+
const known = h.tools.builtins.map((t) => t.name);
|
|
119
|
+
const excluded = known.filter((n) => !mapped.includes(n));
|
|
120
|
+
const tokens: string[] = [h.tools.includeFlag!, [...mapped, ...unmapped].join(",")];
|
|
121
|
+
tokens.push(h.tools.excludeFlag!, excluded.join(","));
|
|
122
|
+
return { tokens, unmapped };
|
|
123
|
+
}
|
|
124
|
+
// pi: strict over built-ins, direct include; unmapped (extension/MCP)
|
|
125
|
+
// names ride along in the same list - pi governs them too.
|
|
126
|
+
return { tokens: [h.tools.includeFlag!, [...mapped, ...unmapped].join(",")], unmapped };
|
|
127
|
+
}
|
|
128
|
+
|
|
129
|
+
// exclude (D2): all descriptor-known names minus the excluded ones.
|
|
130
|
+
const names = validateNames(h, selection.exclude!);
|
|
131
|
+
const { mapped, unmapped: excludedUnmapped } = resolveNames(h, names);
|
|
132
|
+
if (excludedUnmapped.length > 0) {
|
|
133
|
+
// Excluding a name the descriptor does not know: permitted (it may be a
|
|
134
|
+
// runtime-registered tool), rendered natively where the harness's deny
|
|
135
|
+
// list can carry it; on claude only, since its include path consumes
|
|
136
|
+
// unknown names without effect (silent-acceptance hazard).
|
|
137
|
+
if (!h.tools.includeIsStrictAllowlist) {
|
|
138
|
+
const tokens = [h.tools.excludeFlag!, [...mapped, ...excludedUnmapped].join(",")];
|
|
139
|
+
return { tokens, unmapped: excludedUnmapped };
|
|
140
|
+
}
|
|
141
|
+
// pi: a D2 exclude renders as a computed include (its native exclude
|
|
142
|
+
// subtracts from the default set, which cannot express "all minus X"
|
|
143
|
+
// when off-by-default tools are involved). An unknown excluded name
|
|
144
|
+
// cannot be complemented - refuse rather than silently grant it.
|
|
145
|
+
throw new ArgvRefusalError({
|
|
146
|
+
issue: "unknown-tool-name",
|
|
147
|
+
harness: h.name,
|
|
148
|
+
option: "excludeTools",
|
|
149
|
+
supported: [`known tool names: ${h.tools.builtins.map((t) => t.name).join(", ")}`],
|
|
150
|
+
detail: `cannot exclude unknown name(s) ${excludedUnmapped.join(", ")}: the complement cannot be computed`,
|
|
151
|
+
});
|
|
152
|
+
}
|
|
153
|
+
const known = h.tools.builtins.filter((t) => !mapped.includes(t.name));
|
|
154
|
+
const kept = known.map((t) => t.name);
|
|
155
|
+
if (!h.tools.includeIsStrictAllowlist) {
|
|
156
|
+
// claude: deny list IS the complement expression.
|
|
157
|
+
return { tokens: [h.tools.excludeFlag!, mapped.join(",")], unmapped: [] };
|
|
158
|
+
}
|
|
159
|
+
// pi: computed include over every known name minus the excluded.
|
|
160
|
+
return { tokens: [h.tools.includeFlag!, kept.join(",")], unmapped: [] };
|
|
161
|
+
};
|
|
@@ -12,6 +12,7 @@
|
|
|
12
12
|
import type { HarnessDescriptor, OptionRender, SpecBase } from "../knowledge/descriptor.js";
|
|
13
13
|
import { DISCOVERY_FACETS, resolveRender, TURN_OPTION_KEYS } from "../knowledge/descriptor.js";
|
|
14
14
|
import type { DiscoveryOptions, TurnOptions } from "./argv.js";
|
|
15
|
+
import { hintFor } from "./hints.js";
|
|
15
16
|
import { ArgvRefusalError } from "./refusal.js";
|
|
16
17
|
import { CLEAN_SELECTOR, resolveModel, validateEffort } from "./vocabulary.js";
|
|
17
18
|
|
|
@@ -62,6 +63,7 @@ export const renderTurnOptions = (
|
|
|
62
63
|
option: "discovery",
|
|
63
64
|
supported: Object.keys(h.turnOptions).length ? Object.keys(h.turnOptions) : ["(none)"],
|
|
64
65
|
detail: String(requested[0]),
|
|
66
|
+
hint: hintFor(h.name, `discovery.${String(requested[0])}`),
|
|
65
67
|
});
|
|
66
68
|
}
|
|
67
69
|
if (spec.kind !== "discovery") {
|
|
@@ -109,6 +111,7 @@ export const renderTurnOptions = (
|
|
|
109
111
|
option: "discovery",
|
|
110
112
|
facet,
|
|
111
113
|
supported: supportedFacets.length ? supportedFacets : ["(none)"],
|
|
114
|
+
hint: hintFor(h.name, `discovery.${facet}`),
|
|
112
115
|
});
|
|
113
116
|
}
|
|
114
117
|
continue;
|
|
@@ -199,6 +202,7 @@ export const renderTurnOptions = (
|
|
|
199
202
|
option: key,
|
|
200
203
|
supported: Object.keys(h.turnOptions).length ? Object.keys(h.turnOptions) : ["(none)"],
|
|
201
204
|
detail: String(raw),
|
|
205
|
+
hint: hintFor(h.name, key),
|
|
202
206
|
});
|
|
203
207
|
}
|
|
204
208
|
|
|
@@ -145,10 +145,43 @@ export const claudeCode: HarnessDescriptor = deepFreeze({
|
|
|
145
145
|
render: { kind: "flag-list", flags: ["--setting-sources", "project"] },
|
|
146
146
|
},
|
|
147
147
|
skills: {
|
|
148
|
+
// Phase 0 (claude-tool-interplay.md probe 4): --disable-slash-commands
|
|
149
|
+
// removes the Skill tool AND all skills listing - verified a full
|
|
150
|
+
// skills-off switch, not just command dispatch. --setting-sources
|
|
151
|
+
// project stays the extensions-facet spelling (settings scope).
|
|
148
152
|
polarity: "disables",
|
|
149
|
-
render: { kind: "flag-list", flags: ["--
|
|
153
|
+
render: { kind: "flag-list", flags: ["--disable-slash-commands"] },
|
|
150
154
|
},
|
|
151
155
|
},
|
|
152
156
|
},
|
|
153
157
|
},
|
|
158
|
+
// Phase 0 fixtures: claude-tool-interplay.md. include is a permission
|
|
159
|
+
// grant (Bash, Edit stay visible under --allowedTools Read); only the
|
|
160
|
+
// disallow flag reshapes the model-visible set. Both flags together
|
|
161
|
+
// compose, deny winning on overlap. Patterns (Bash(git *)) valid in both
|
|
162
|
+
// lists; unknown PATTERN spellings warn on stderr, unknown exact names
|
|
163
|
+
// are the silent-acceptance hazard the curated vocabulary guards.
|
|
164
|
+
tools: {
|
|
165
|
+
includeFlag: "--allowedTools",
|
|
166
|
+
excludeFlag: "--disallowedTools",
|
|
167
|
+
includeIsStrictAllowlist: false,
|
|
168
|
+
composable: true,
|
|
169
|
+
builtins: [
|
|
170
|
+
{ name: "Bash", defaultEnabled: true },
|
|
171
|
+
{ name: "Edit", defaultEnabled: true },
|
|
172
|
+
{ name: "Glob", defaultEnabled: true },
|
|
173
|
+
{ name: "Grep", defaultEnabled: true },
|
|
174
|
+
{ name: "Read", defaultEnabled: true },
|
|
175
|
+
{ name: "Write", defaultEnabled: true },
|
|
176
|
+
{ name: "WebFetch", defaultEnabled: true },
|
|
177
|
+
{ name: "WebSearch", defaultEnabled: true },
|
|
178
|
+
{ name: "Monitor", defaultEnabled: true },
|
|
179
|
+
{ name: "Task", defaultEnabled: true },
|
|
180
|
+
{ name: "Skill", defaultEnabled: true },
|
|
181
|
+
{ name: "NotebookEdit", defaultEnabled: true },
|
|
182
|
+
{ name: "LSP", defaultEnabled: true },
|
|
183
|
+
],
|
|
184
|
+
categories: [],
|
|
185
|
+
denySemantics: "remove-from-set",
|
|
186
|
+
},
|
|
154
187
|
});
|
package/src/knowledge/codex.ts
CHANGED
|
@@ -114,4 +114,23 @@ export const codexCli: HarnessDescriptor = deepFreeze({
|
|
|
114
114
|
resumeRender: null,
|
|
115
115
|
},
|
|
116
116
|
},
|
|
117
|
+
// Phase 0 fixtures: codex-tool-surface.md. No name lists anywhere -
|
|
118
|
+
// not on the CLI, not in config.toml. Control is feature booleans
|
|
119
|
+
// (reachable per-call via -c key=value / --enable/--disable FEATURE),
|
|
120
|
+
// sandbox, and approval policy. MCP servers do have per-tool keys
|
|
121
|
+
// (mcp_servers.<id>.tools.<tool>) but built-ins do not.
|
|
122
|
+
tools: {
|
|
123
|
+
includeFlag: null,
|
|
124
|
+
excludeFlag: null,
|
|
125
|
+
includeIsStrictAllowlist: false,
|
|
126
|
+
composable: false,
|
|
127
|
+
builtins: [],
|
|
128
|
+
categories: [
|
|
129
|
+
{ key: "shell", disableFlag: null, configKey: "features.shell_tool" },
|
|
130
|
+
{ key: "exec", disableFlag: null, configKey: "features.unified_exec" },
|
|
131
|
+
{ key: "web", disableFlag: null, configKey: "web_search" },
|
|
132
|
+
{ key: "view-image", disableFlag: null, configKey: "tools.view_image" },
|
|
133
|
+
],
|
|
134
|
+
denySemantics: "no-lists",
|
|
135
|
+
},
|
|
117
136
|
});
|