lobstah 0.7.9 → 0.7.11
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/dist/main.js +2386 -1633
- package/dist/runner.js +450 -301
- package/docs/configuration.md +65 -1
- package/docs/man.md +3 -1
- package/package.json +1 -1
package/docs/configuration.md
CHANGED
|
@@ -189,7 +189,10 @@ send a continuation.
|
|
|
189
189
|
|
|
190
190
|
The defaults wait 900, 900, 900, 1800, 3600, 7200, then 14400 seconds.
|
|
191
191
|
History is stored per session across re-arms. An event or user prompt resets
|
|
192
|
-
that session; a dispatch or send resets all this boat's waiters.
|
|
192
|
+
that session; a dispatch or send resets all this boat's waiters. A prompt the
|
|
193
|
+
harness injects when a background task finishes (a `<task-notification>`,
|
|
194
|
+
which is how a backgrounded `man wait` or `soak --wait` ending quietly comes
|
|
195
|
+
back) is not a user prompt and does not reset. Already-armed
|
|
193
196
|
waiters adopt a fresh base deadline on their next poll. Polling and liveness
|
|
194
197
|
heartbeats keep their usual cadence, so events still wake immediately.
|
|
195
198
|
`--peek` and zero-time trap probes do not advance the backoff.
|
|
@@ -402,6 +405,67 @@ as that identity, with its token. Any key except `enabled`, `repos` and
|
|
|
402
405
|
The `review` attention kind is not in the default `attentionKinds`; add it to
|
|
403
406
|
let review items walk in `man tend`, the glass, and the pet.
|
|
404
407
|
|
|
408
|
+
## `[permissions]` — the Claude Code auto-mode profile
|
|
409
|
+
|
|
410
|
+
Traps and the helm run Claude Code in auto mode. `lobstah permissions apply`
|
|
411
|
+
merges lobstah's profile into your Claude Code user settings
|
|
412
|
+
(`$CLAUDE_CONFIG_DIR/settings.json`, else `~/.claude/settings.json`). That is
|
|
413
|
+
the only file Claude Code reads `autoMode.environment` from, apart from managed
|
|
414
|
+
settings and `--settings`.
|
|
415
|
+
|
|
416
|
+
The base profile adds context for Claude Code's classifier:
|
|
417
|
+
- lobstah's worktrees are your development checkouts;
|
|
418
|
+
- the helm doing what you asked in its chat is not a bypass;
|
|
419
|
+
- merges to trunk, production deploys, secrets, accounts and lobstah's own
|
|
420
|
+
permission controls stay yours.
|
|
421
|
+
|
|
422
|
+
It also allows `Read` and `Edit` on `<lobstah home>/worktrees/**`.
|
|
423
|
+
`[permissions]` adds to it:
|
|
424
|
+
|
|
425
|
+
```toml
|
|
426
|
+
[permissions]
|
|
427
|
+
environment = ["**Trusted GitHub orgs**: acme and acme-labs are the user's own orgs."]
|
|
428
|
+
allow = ["Bash(wrangler dev *)"]
|
|
429
|
+
deny = ["Bash(vercel env pull *)"]
|
|
430
|
+
|
|
431
|
+
[permissions.codex]
|
|
432
|
+
forbidden = [["vercel", "env", "pull"]]
|
|
433
|
+
allow = [["wrangler", "dev"]]
|
|
434
|
+
```
|
|
435
|
+
|
|
436
|
+
| Key | Meaning |
|
|
437
|
+
|---|---|
|
|
438
|
+
| `environment` | Prose entries added to `autoMode.environment`, e.g. trusted orgs, cloud accounts, or env-pull policy. |
|
|
439
|
+
| `allow`, `deny` | Claude Code permission rules (`Tool` or `Tool(specifier)`) added to `permissions.allow` / `permissions.deny`. These are resolved before the classifier. |
|
|
440
|
+
| `codex.forbidden`, `codex.prompt`, `codex.allow` | Codex command rules, each an argv prefix (`["gh", "pr", "merge"]`). A Codex `allow` runs the command outside Codex's sandbox, so these are never translated from the Claude rules. |
|
|
441
|
+
|
|
442
|
+
**Codex.** Codex has its own home: `$CODEX_HOME`, else `~/.codex`. When it exists, the same verbs also change two things there:
|
|
443
|
+
- **The auto reviewer's added policy.** This is `[auto_review].extra_policy` in `config.toml`, which Codex's reviewer reads alongside its own policy. lobstah writes a block between `<!-- lobstah:begin … -->` and `<!-- lobstah:end -->` holding the same three facts and `environment`.
|
|
444
|
+
- Text outside the markers is yours and stays.
|
|
445
|
+
- lobstah edits only the key's own lines, keeping comments, key order and the string's quoting. It parses the file before and after; anything else changing refuses the write.
|
|
446
|
+
- **`rules/lobstah.rules`.** This file is wholly lobstah's. It holds the base rules plus `[permissions.codex]`:
|
|
447
|
+
- forbidden: `lobstah grant`, `lobstah trap grant|bypass` and `lobstah permissions apply|remove`;
|
|
448
|
+
- prompt: `gh pr merge`.
|
|
449
|
+
- `codex execpolicy check` loads every rule before the file is put in place.
|
|
450
|
+
|
|
451
|
+
`remove` deletes `rules/lobstah.rules` and lobstah's block. Neither verb sets `approvals_reviewer` or `approval_policy`: `man throw` already starts Codex traps with `--approve-for-me`.
|
|
452
|
+
|
|
453
|
+
**What each form does:**
|
|
454
|
+
- Bare `lobstah permissions` shows the diff and writes nothing.
|
|
455
|
+
- `apply` merges, keeping every existing key and rule. A new
|
|
456
|
+
`autoMode.environment` starts with `"$defaults"`, so Claude Code's built-in
|
|
457
|
+
context stays.
|
|
458
|
+
- `remove` takes lobstah's entries out.
|
|
459
|
+
- Both `apply` and `remove` back up the previous file to
|
|
460
|
+
`<lobstah home>/backups/`, ask before writing (`--yes` skips the question),
|
|
461
|
+
and refuse inside an agent's shell.
|
|
462
|
+
|
|
463
|
+
lobstah records the entries it added in `<lobstah home>/claude-permissions.json`,
|
|
464
|
+
so a re-run after editing `[permissions]` removes only lobstah's own stale
|
|
465
|
+
entries. `lobstah doctor` reports whether the profile is applied and current.
|
|
466
|
+
There is one Claude Code user settings file, so there is one `[permissions]`
|
|
467
|
+
table rather than one per grounds.
|
|
468
|
+
|
|
405
469
|
## Environment
|
|
406
470
|
|
|
407
471
|
| Variable | Meaning |
|
package/docs/man.md
CHANGED
|
@@ -90,7 +90,9 @@ for writes. Files are editable Markdown at
|
|
|
90
90
|
|
|
91
91
|
The first user prompt of a new day adds a non-blocking reminder to the helm:
|
|
92
92
|
handle that message, then put yesterday's recap and a draft of today's goals
|
|
93
|
-
on one card.
|
|
93
|
+
on one card. A turn the harness injects when a background task finishes (a
|
|
94
|
+
`<task-notification>`, such as a quiet `man wait` ending) is not a user
|
|
95
|
+
prompt: it neither gets the reminder nor uses it up. Existing goals, a `Goals for YYYY-MM-DD` card, or an already
|
|
94
96
|
delivered reminder suppress another ask. The helm saves the answered goals
|
|
95
97
|
with `goals set`; the first reminder leaves a comment in that same daily file.
|
|
96
98
|
The helm uses them to order work and choose the next quiet step.
|