@saccolabs/pi-claude-cli 0.4.15 → 0.5.0
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 +56 -5
- package/package.json +1 -1
- package/src/autocompact.ts +99 -0
- package/src/process-manager.ts +25 -5
- package/src/prompt-builder.ts +5 -4
- package/src/system-prompt-mode.ts +8 -4
package/README.md
CHANGED
|
@@ -95,15 +95,54 @@ allowlists). Model access and your subscription login are unaffected.
|
|
|
95
95
|
Related knobs: `PI_CLAUDE_CLI_TIMEOUT_MS` overrides the 300s inactivity
|
|
96
96
|
timeout (CLI-side tools can be silent on stdout for minutes).
|
|
97
97
|
|
|
98
|
+
### Auto-compact window
|
|
99
|
+
|
|
100
|
+
The provider resumes **one** CLI session for a pi session's whole life, and
|
|
101
|
+
nothing else ever shrinks it. On 1M-context models the CLI's own auto-compact
|
|
102
|
+
default lets that session ratchet toward a million tokens — measured across 26
|
|
103
|
+
real sessions, contexts reached 480k+, the average request carried 202k
|
|
104
|
+
tokens, and every request re-reads the full context. So the provider passes
|
|
105
|
+
`--autocompact 200000` **by default**: Claude Code compacts the session itself
|
|
106
|
+
when its context nears 200k, keeping the cached system-prompt prefix and full
|
|
107
|
+
transcript fidelity.
|
|
108
|
+
|
|
109
|
+
`PI_CLAUDE_CLI_AUTOCOMPACT` configures it (read per spawn, like the flags
|
|
110
|
+
above):
|
|
111
|
+
|
|
112
|
+
| Value | Behaviour |
|
|
113
|
+
| ----------------------- | ---------------------------------------------------------------------------------------------- |
|
|
114
|
+
| _(unset)_ | `--autocompact 200000` — the 200k budget these models run under everywhere the 1M beta is off. |
|
|
115
|
+
| `400k`, `400000`, `400` | Any window from 100k to 1M; bare numbers are thousands (CLI shorthand). |
|
|
116
|
+
| `auto` | `--autocompact auto` — the CLI's own default (≈ the model's full window). |
|
|
117
|
+
| `off` | Omit the flag entirely (use on CLIs that predate `--autocompact`). |
|
|
118
|
+
|
|
119
|
+
The value is a token **count**, not a percentage: cache read/write bill per
|
|
120
|
+
token and every request re-reads the whole context, so the sane budget is the
|
|
121
|
+
same on a 200k model and a 1M one. Invalid values warn and fall back to the
|
|
122
|
+
default instead of reaching the CLI, which rejects them by refusing to start.
|
|
123
|
+
|
|
124
|
+
Note for pre-existing sessions: the first resumed turn of a session already
|
|
125
|
+
past the window compacts immediately — one summarization pass, then the
|
|
126
|
+
session continues small. That is the remediation, not a bug. pi's own
|
|
127
|
+
compaction is separate (it rewrites pi's transcript, never the CLI session's)
|
|
128
|
+
and with this cap it should rarely trigger.
|
|
129
|
+
|
|
98
130
|
### Which system prompt
|
|
99
131
|
|
|
100
132
|
`PI_CLAUDE_CLI_SYSTEM_PROMPT` chooses whose system prompt the subprocess
|
|
101
133
|
runs under. It is read per spawn, so a host can change it between sessions.
|
|
102
134
|
|
|
103
|
-
| Value | Behaviour
|
|
104
|
-
| -------------------- |
|
|
105
|
-
| `claude` _(default)_ | `--append-system-prompt`: pi's prompt layers on top of Claude Code's own. |
|
|
106
|
-
| `pi` | `--system-prompt`: pi's prompt replaces Claude Code's entirely. |
|
|
135
|
+
| Value | Behaviour |
|
|
136
|
+
| -------------------- | ------------------------------------------------------------------------------ |
|
|
137
|
+
| `claude` _(default)_ | `--append-system-prompt-file`: pi's prompt layers on top of Claude Code's own. |
|
|
138
|
+
| `pi` | `--system-prompt-file`: pi's prompt replaces Claude Code's entirely. |
|
|
139
|
+
|
|
140
|
+
The `-file` suffix matters: `--system-prompt` / `--append-system-prompt`
|
|
141
|
+
(unsuffixed) take a **literal string**, not a path. Passing a temp-file path
|
|
142
|
+
to the unsuffixed flag makes the path itself the prompt — pi's instructions
|
|
143
|
+
never reach the model, silently, with no error. This shipped unnoticed since
|
|
144
|
+
the provider's system-prompt support was first added; see the correction
|
|
145
|
+
below.
|
|
107
146
|
|
|
108
147
|
`minimal` is accepted as an alias for `pi`, `append` for `claude`; anything
|
|
109
148
|
unrecognised falls back to the default rather than failing a session.
|
|
@@ -127,7 +166,7 @@ restyles its prompt so the `Available tools:` / `Guidelines:` anchors are
|
|
|
127
166
|
missing, the prompt passes through untouched rather than being mangled.
|
|
128
167
|
|
|
129
168
|
The system prompt goes on **every** spawn, not just the session-creating one:
|
|
130
|
-
the CLI does not keep `--system-prompt` across `--resume`, and a resumed
|
|
169
|
+
the CLI does not keep `--system-prompt-file` across `--resume`, and a resumed
|
|
131
170
|
session without it silently reverts to Claude Code's default prompt from turn
|
|
132
171
|
2 onwards. Because an identical prefix is what keeps the prompt cache warm,
|
|
133
172
|
the prompt a session was created with is stored in the sidecar
|
|
@@ -135,6 +174,18 @@ the prompt a session was created with is stored in the sidecar
|
|
|
135
174
|
verbatim rather than rebuilt. A change to the mode therefore takes effect on
|
|
136
175
|
the next new session, not the current one.
|
|
137
176
|
|
|
177
|
+
> **Correction (2026-08-29).** Both bullets above named the unsuffixed flags
|
|
178
|
+
> until this date. They were wrong the whole time the provider has supported a
|
|
179
|
+
> system prompt: `--system-prompt` / `--append-system-prompt` take a literal
|
|
180
|
+
> string, and the provider was handing them a temp-file path. That path string
|
|
181
|
+
> either became the entire "system prompt" (`pi` mode) or got appended as
|
|
182
|
+
> noise Claude Code's model ignored (`claude` mode) — either way, pi's actual
|
|
183
|
+
> instructions never reached the model, on ANY turn, since the very first spawn.
|
|
184
|
+
> Fixed by switching to `--system-prompt-file` / `--append-system-prompt-file`,
|
|
185
|
+
> which take a path. See
|
|
186
|
+
> [pidex's write-up](https://github.com/agustinsacco/pidex/blob/main/specs/log/2026-08-29-claude-cli-lifecycle-verification.md)
|
|
187
|
+
> for the live before/after.
|
|
188
|
+
|
|
138
189
|
## License
|
|
139
190
|
|
|
140
191
|
MIT
|
package/package.json
CHANGED
|
@@ -0,0 +1,99 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Auto-compact window resolution (PI_CLAUDE_CLI_AUTOCOMPACT).
|
|
3
|
+
*
|
|
4
|
+
* Claude Code compacts a session automatically when its context approaches
|
|
5
|
+
* the auto-compact window (`--autocompact <auto|tokens>`, claude 2.1.x). On
|
|
6
|
+
* 1M-context models the CLI's own default lets a long-lived session ratchet
|
|
7
|
+
* toward a million tokens, and this provider resumes ONE CLI session for a
|
|
8
|
+
* pi session's whole life — nothing else ever shrinks it. Measured across
|
|
9
|
+
* 26 real pidex sessions (2026-08-30): contexts ratcheted to 480k+, the
|
|
10
|
+
* average request carried 202k tokens, and cache reads alone were ~53% of
|
|
11
|
+
* total spend. Sessions that stayed near 100–150k did the same work at a
|
|
12
|
+
* fraction of the cost.
|
|
13
|
+
*
|
|
14
|
+
* So the provider caps the window at 200k tokens BY DEFAULT — the budget
|
|
15
|
+
* these models run under everywhere the 1M beta is not enabled — and lets
|
|
16
|
+
* the host raise, lower, or disable it:
|
|
17
|
+
*
|
|
18
|
+
* PI_CLAUDE_CLI_AUTOCOMPACT=200k cap at 200k (the default)
|
|
19
|
+
* PI_CLAUDE_CLI_AUTOCOMPACT=400000 plain token counts work too
|
|
20
|
+
* PI_CLAUDE_CLI_AUTOCOMPACT=400 bare numbers are thousands (CLI shorthand)
|
|
21
|
+
* PI_CLAUDE_CLI_AUTOCOMPACT=auto the CLI's own default behaviour
|
|
22
|
+
* PI_CLAUDE_CLI_AUTOCOMPACT=off omit the flag entirely (also: 0, none,
|
|
23
|
+
* disable, disabled — for CLIs that
|
|
24
|
+
* predate --autocompact)
|
|
25
|
+
*
|
|
26
|
+
* The value is a token COUNT, not a percentage of the model's window. The
|
|
27
|
+
* costs this guards against are absolute — cache read/write bill per token,
|
|
28
|
+
* and every request re-reads the whole context — so "half the window" means
|
|
29
|
+
* something completely different on a 200k model than on a 1M one, while
|
|
30
|
+
* 200k tokens costs the same everywhere.
|
|
31
|
+
*
|
|
32
|
+
* An invalid or out-of-range value falls back to the default with a warning
|
|
33
|
+
* rather than being passed through: the CLI rejects bad values by refusing
|
|
34
|
+
* to start, which would kill every turn of every session over a typo.
|
|
35
|
+
*/
|
|
36
|
+
|
|
37
|
+
/** Default auto-compact window, in tokens. */
|
|
38
|
+
export const DEFAULT_AUTOCOMPACT_TOKENS = 200_000;
|
|
39
|
+
|
|
40
|
+
/** The CLI accepts 100k–1M (claude 2.1.231: "It must be 'auto', or between 100k and 1M"). */
|
|
41
|
+
const MIN_TOKENS = 100_000;
|
|
42
|
+
const MAX_TOKENS = 1_000_000;
|
|
43
|
+
|
|
44
|
+
const OFF_VALUES = new Set([
|
|
45
|
+
"off",
|
|
46
|
+
"0",
|
|
47
|
+
"none",
|
|
48
|
+
"disable",
|
|
49
|
+
"disabled",
|
|
50
|
+
"false",
|
|
51
|
+
]);
|
|
52
|
+
|
|
53
|
+
/**
|
|
54
|
+
* Parse a user-supplied window size into a token count.
|
|
55
|
+
* Mirrors the CLI's accepted forms: `500k`, `1M`, `200000`, and bare `200`
|
|
56
|
+
* meaning thousands. Returns undefined when unparseable.
|
|
57
|
+
*/
|
|
58
|
+
export function parseAutocompactTokens(raw: string): number | undefined {
|
|
59
|
+
const m = /^(\d+(?:\.\d+)?)\s*([km])?$/i.exec(raw.trim());
|
|
60
|
+
if (!m) return undefined;
|
|
61
|
+
const n = Number(m[1]);
|
|
62
|
+
if (!Number.isFinite(n) || n <= 0) return undefined;
|
|
63
|
+
const suffix = (m[2] ?? "").toLowerCase();
|
|
64
|
+
if (suffix === "k") return Math.round(n * 1_000);
|
|
65
|
+
if (suffix === "m") return Math.round(n * 1_000_000);
|
|
66
|
+
// Bare number: the CLI reads `200` as 200k shorthand; anything that already
|
|
67
|
+
// looks like a token count (>= 100k) is taken literally.
|
|
68
|
+
return n < MIN_TOKENS ? Math.round(n * 1_000) : Math.round(n);
|
|
69
|
+
}
|
|
70
|
+
|
|
71
|
+
/**
|
|
72
|
+
* Resolve the `--autocompact` argument from the environment.
|
|
73
|
+
*
|
|
74
|
+
* Returns the string to pass to the flag, or undefined to omit the flag
|
|
75
|
+
* entirely (explicit off). Unset resolves to the 200k default; `auto` is
|
|
76
|
+
* passed through so the CLI applies its own default; invalid values warn
|
|
77
|
+
* and fall back to the default.
|
|
78
|
+
*/
|
|
79
|
+
export function resolveAutocompact(
|
|
80
|
+
env: NodeJS.ProcessEnv = process.env,
|
|
81
|
+
): string | undefined {
|
|
82
|
+
const raw = (env.PI_CLAUDE_CLI_AUTOCOMPACT ?? "").trim();
|
|
83
|
+
if (raw === "") return String(DEFAULT_AUTOCOMPACT_TOKENS);
|
|
84
|
+
|
|
85
|
+
const lowered = raw.toLowerCase();
|
|
86
|
+
if (OFF_VALUES.has(lowered)) return undefined;
|
|
87
|
+
if (lowered === "auto") return "auto";
|
|
88
|
+
|
|
89
|
+
const tokens = parseAutocompactTokens(raw);
|
|
90
|
+
if (tokens === undefined || tokens < MIN_TOKENS || tokens > MAX_TOKENS) {
|
|
91
|
+
console.warn(
|
|
92
|
+
`[pi-claude-cli] PI_CLAUDE_CLI_AUTOCOMPACT=${JSON.stringify(raw)} is not ` +
|
|
93
|
+
`'auto', 'off', or a window between 100k and 1M — using the default ` +
|
|
94
|
+
`${DEFAULT_AUTOCOMPACT_TOKENS} tokens`,
|
|
95
|
+
);
|
|
96
|
+
return String(DEFAULT_AUTOCOMPACT_TOKENS);
|
|
97
|
+
}
|
|
98
|
+
return String(tokens);
|
|
99
|
+
}
|
package/src/process-manager.ts
CHANGED
|
@@ -16,14 +16,16 @@ import {
|
|
|
16
16
|
DEFAULT_SYSTEM_PROMPT_MODE,
|
|
17
17
|
type SystemPromptMode,
|
|
18
18
|
} from "./system-prompt-mode.js";
|
|
19
|
+
import { resolveAutocompact } from "./autocompact.js";
|
|
19
20
|
|
|
20
21
|
/**
|
|
21
22
|
* Spawn a Claude CLI subprocess with all required flags for stream-json communication.
|
|
22
23
|
*
|
|
23
24
|
* @param modelId - The model ID to pass via --model flag
|
|
24
25
|
* @param systemPrompt - Optional system prompt. In `claude` mode it is appended
|
|
25
|
-
* to Claude Code's own via --append-system-prompt; in `pi` mode it
|
|
26
|
-
* it via --system-prompt.
|
|
26
|
+
* to Claude Code's own via --append-system-prompt-file; in `pi` mode it
|
|
27
|
+
* replaces it via --system-prompt-file. The `-file` suffix is required: the
|
|
28
|
+
* unsuffixed flags take a literal string. See src/system-prompt-mode.ts.
|
|
27
29
|
* @param options - Optional cwd, AbortSignal, effort level and prompt mode
|
|
28
30
|
* @returns The spawned ChildProcess with piped stdin/stdout/stderr
|
|
29
31
|
*/
|
|
@@ -98,8 +100,16 @@ export function spawnClaude(
|
|
|
98
100
|
}
|
|
99
101
|
|
|
100
102
|
if (systemPrompt) {
|
|
101
|
-
// Write system prompt to a temp file
|
|
102
|
-
//
|
|
103
|
+
// Write the system prompt to a temp file and pass the FILE flags.
|
|
104
|
+
//
|
|
105
|
+
// `--system-prompt` / `--append-system-prompt` take a literal string, NOT
|
|
106
|
+
// a path: handing them a path makes the path itself the system prompt, so
|
|
107
|
+
// pi's instructions never reach the model and the session silently runs on
|
|
108
|
+
// Claude Code's defaults. Verified on claude 2.1.231 — with a path the
|
|
109
|
+
// model denied having the codename its prompt assigned; with
|
|
110
|
+
// `--system-prompt-file` it answered correctly. The file variants also keep
|
|
111
|
+
// the prompt off the command line, which is what avoids ENAMETOOLONG on
|
|
112
|
+
// Windows.
|
|
103
113
|
//
|
|
104
114
|
// Keyed by CLI session, not just pid: the prompt goes on every spawn
|
|
105
115
|
// (see provider.ts), and pi can run two turns of one process at once —
|
|
@@ -113,7 +123,7 @@ export function spawnClaude(
|
|
|
113
123
|
// pi's on top of it. See src/system-prompt-mode.ts for the trade-off.
|
|
114
124
|
const mode = options?.systemPromptMode ?? DEFAULT_SYSTEM_PROMPT_MODE;
|
|
115
125
|
args.push(
|
|
116
|
-
mode === "pi" ? "--system-prompt" : "--append-system-prompt",
|
|
126
|
+
mode === "pi" ? "--system-prompt-file" : "--append-system-prompt-file",
|
|
117
127
|
tmpFile,
|
|
118
128
|
);
|
|
119
129
|
}
|
|
@@ -134,6 +144,16 @@ export function spawnClaude(
|
|
|
134
144
|
args.push("--mcp-config", options.mcpConfigPath);
|
|
135
145
|
}
|
|
136
146
|
|
|
147
|
+
// Auto-compact window (PI_CLAUDE_CLI_AUTOCOMPACT, default 200k — see
|
|
148
|
+
// src/autocompact.ts). The CLI does not keep flags across --resume, so
|
|
149
|
+
// like the system prompt this goes on every spawn. Resolved per spawn so
|
|
150
|
+
// a host can change the setting without restarting pi; the flag is
|
|
151
|
+
// config, not context, so changing it never invalidates the prompt cache.
|
|
152
|
+
const autocompact = resolveAutocompact();
|
|
153
|
+
if (autocompact !== undefined) {
|
|
154
|
+
args.push("--autocompact", autocompact);
|
|
155
|
+
}
|
|
156
|
+
|
|
137
157
|
const proc = spawn("claude", args, {
|
|
138
158
|
stdio: ["pipe", "pipe", "pipe"],
|
|
139
159
|
cwd: options?.cwd ?? process.cwd(),
|
package/src/prompt-builder.ts
CHANGED
|
@@ -409,10 +409,11 @@ export function rewritePiToolSections(systemPrompt: string): string {
|
|
|
409
409
|
* appending AGENTS.md content if found (walking up from cwd, then global fallback).
|
|
410
410
|
* Sanitizes .pi references to .claude for Claude Code compatibility.
|
|
411
411
|
*
|
|
412
|
-
* In `pi` mode the caller passes the prompt to `--system-prompt`,
|
|
413
|
-
* Claude Code's own, so pi's tool sections are rewritten into Claude
|
|
414
|
-
* names first. In `claude` mode the prompt is appended to Claude
|
|
415
|
-
* pi's wording is left exactly as
|
|
412
|
+
* In `pi` mode the caller passes the prompt to `--system-prompt-file`,
|
|
413
|
+
* replacing Claude Code's own, so pi's tool sections are rewritten into Claude
|
|
414
|
+
* Code's names first. In `claude` mode the prompt is appended to Claude
|
|
415
|
+
* Code's (`--append-system-prompt-file`) and pi's wording is left exactly as
|
|
416
|
+
* pi wrote it.
|
|
416
417
|
*/
|
|
417
418
|
export function buildSystemPrompt(
|
|
418
419
|
context: { systemPrompt?: string; messages: any[] },
|
|
@@ -2,12 +2,16 @@
|
|
|
2
2
|
* Which system prompt the Claude CLI subprocess runs under.
|
|
3
3
|
*
|
|
4
4
|
* `claude` (the default) appends pi's prompt to Claude Code's own via
|
|
5
|
-
* `--append-system-prompt`. Everything the CLI normally knows about its
|
|
5
|
+
* `--append-system-prompt-file`. Everything the CLI normally knows about its
|
|
6
6
|
* built-in tools stays in place, and pi's instructions ride on top.
|
|
7
7
|
*
|
|
8
|
-
* `pi` replaces Claude Code's prompt outright via `--system-prompt`,
|
|
9
|
-
* only pi's — the point of a minimal harness being that it does not
|
|
10
|
-
* another agent's preamble.
|
|
8
|
+
* `pi` replaces Claude Code's prompt outright via `--system-prompt-file`,
|
|
9
|
+
* leaving only pi's — the point of a minimal harness being that it does not
|
|
10
|
+
* inherit another agent's preamble.
|
|
11
|
+
*
|
|
12
|
+
* The `-file` suffix is not optional: the unsuffixed flags take a literal
|
|
13
|
+
* string, and passing them a path (as this provider did until 2026-08-29)
|
|
14
|
+
* makes the path itself the prompt, silently. See process-manager.ts.
|
|
11
15
|
*
|
|
12
16
|
* Sizing, measured rather than assumed: in a real session the CLI's fixed
|
|
13
17
|
* cached prefix sat at 17,475 tokens. That is Claude Code's system prompt
|