aegis-desktop 1.0.12 → 1.0.14

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.
@@ -0,0 +1,229 @@
1
+ // SPDX-License-Identifier: AGPL-3.0-or-later
2
+ // Copyright (C) 2026 Niklas Borneklint — AEGIS Code (aegiscode) / Aegiscloud.
3
+ // AEGIS Code (aegiscode) — source: https://github.com/aegisinfo/aegiscode-plugin
4
+ // Conveying a modified copy, or running one as a network service, owes the
5
+ // corresponding source under this licence (§5/§13) — see LICENSING.md.
6
+ 'use strict';
7
+
8
+ /**
9
+ * google-connect.js — the desktop's Google identity + Drive flow.
10
+ *
11
+ * Two distinct steps, run in this order because the second depends on the
12
+ * first having actually landed:
13
+ *
14
+ * 1. Sign in with Google (`client/google-signin.js`'s local round-trip) —
15
+ * replaces this machine's saved AEGIS key with the one tied to that
16
+ * Google identity, exactly as `aegiscode connect` already does for the
17
+ * CLI (cli/src/connect.js). Desktop never had this wiring before this
18
+ * file; the only Google sign-in that existed anywhere in this product
19
+ * was the CLI's.
20
+ * 2. Connect Drive (`client/drive-connect.js`, which opens `/connect/drive`
21
+ * in the system browser and polls `/api/drive/token`) — grants the
22
+ * `drive.file` scope to WHICHEVER Google identity that browser's
23
+ * consent screen uses. That is only the SAME account as step 1 if step
24
+ * 1 ran first on this machine: `/connect/drive`'s callback resolves an
25
+ * account purely by Google identity, with no notion of "the account
26
+ * this machine's API key already belongs to" — so connecting Drive
27
+ * before signing in attaches the tokens to a DIFFERENT account than the
28
+ * one this machine will poll, and the poll just times out.
29
+ *
30
+ * "Signed in with Google" is NOT "has an API key". Every install
31
+ * auto-provisions an anonymous device key, so a key always exists and an
32
+ * `!apiKey` guard never fires. The only honest source for the real
33
+ * question is the server's `google_linked` (users.google_id), carried on
34
+ * `/api/drive/token` and surfaced as `status().googleLinked` — so
35
+ * `connectDrive` refuses on that, and the renderer disables the button
36
+ * from the same flag.
37
+ *
38
+ * Electron-free and fully injected — same shape as `account-key.js` next to
39
+ * it — so the whole flow runs under a test with a stub `aegis` client, a
40
+ * stub `googleSignIn`/`driveConnect`/`driveUpload`, and no browser, window or
41
+ * network.
42
+ */
43
+
44
+ function createGoogleConnect({
45
+ aegis,
46
+ credentials,
47
+ persistKey,
48
+ openUrl,
49
+ googleSignIn,
50
+ driveConnect,
51
+ driveUpload,
52
+ onAccountChange,
53
+ }) {
54
+ let signInInFlight = null;
55
+ let connectInFlight = null;
56
+ let lastSyncAt = null;
57
+ let lastSyncError = null;
58
+ // Did a Google round-trip complete in THIS process? That is first-hand
59
+ // evidence of a linked account and does not depend on the server reporting
60
+ // it — which older servers do not (see `status`).
61
+ let signedInViaGoogle = false;
62
+
63
+ /** Sign this machine in with Google, replacing whatever key it had. */
64
+ function signInWithGoogle() {
65
+ if (!signInInFlight) {
66
+ signInInFlight = (async () => {
67
+ try {
68
+ // Everything from here on (binding the local callback server,
69
+ // opening the browser, waiting for the round-trip) can fail, and
70
+ // every caller of this module — the IPC handler in main.js chief
71
+ // among them — is promised `{ ok, … }`/`{ ok: false, error }`,
72
+ // never a rejection. One try covers the whole flow instead of only
73
+ // the `session.wait()` half of it.
74
+ let session;
75
+ try {
76
+ session = await googleSignIn.startGoogleSignIn({ apiBase: aegis.apiBase });
77
+ } catch (e) {
78
+ return { ok: false, error: `could not start the local sign-in server — ${(e && e.message) || e}` };
79
+ }
80
+ const opened = Boolean(openUrl(session.url));
81
+ let result;
82
+ try {
83
+ result = await session.wait();
84
+ } catch (e) {
85
+ return { ok: false, opened, url: session.url, error: (e && e.message) || String(e) };
86
+ }
87
+ // Live client first: the very next call this process makes (e.g.
88
+ // the renderer asking for drive status right after this resolves)
89
+ // must see the new identity, not whatever key this process started
90
+ // with.
91
+ aegis.setApiKey(result.apiKey);
92
+ persistKey(result.apiKey);
93
+ if (result.memoryToken && credentials && typeof credentials.saveMemoryToken === 'function') {
94
+ try {
95
+ credentials.saveMemoryToken(result.memoryToken);
96
+ } catch {
97
+ /* best-effort — the account key itself is already saved */
98
+ }
99
+ }
100
+ // Before the notification, which is what makes the renderer re-read
101
+ // `status()` — otherwise that refresh races this assignment and
102
+ // reports the sign-in as not having happened.
103
+ signedInViaGoogle = true;
104
+ if (typeof onAccountChange === 'function') onAccountChange();
105
+ return { ok: true, opened, url: session.url, username: result.username || null };
106
+ } finally {
107
+ signInInFlight = null;
108
+ }
109
+ })();
110
+ }
111
+ return signInInFlight;
112
+ }
113
+
114
+ /** Open `/connect/drive` and wait for this account to show connected. */
115
+ async function connectDrive() {
116
+ if (!aegis.apiKey) {
117
+ return {
118
+ ok: false,
119
+ error: 'sign in with Google first — this device has no AEGIS account key yet',
120
+ };
121
+ }
122
+ if (!connectInFlight) {
123
+ connectInFlight = (async () => {
124
+ try {
125
+ // The key exists but may be an anonymous device key. Ask the server
126
+ // whether this account is actually tied to a Google identity before
127
+ // opening a consent screen that would attach the tokens elsewhere.
128
+ // Inside the in-flight promise so a double click cannot open two
129
+ // consent tabs. If the status call fails, fall through: a transport
130
+ // blip must not be reported as "not signed in with Google".
131
+ let linked = true;
132
+ try {
133
+ const st = await driveConnect.driveStatus(aegis);
134
+ // Refuse only on a definite "not linked". A server that predates
135
+ // `google_linked` answers the same shape as an unlinked account,
136
+ // and refusing on that would make a real Google sign-in unable to
137
+ // ever connect Drive — the exact dead end this guard exists to
138
+ // prevent, only inverted.
139
+ linked = !(st.googleLinkedKnown && !st.googleLinked);
140
+ } catch {
141
+ /* unknown — let the flow try rather than misreport the reason */
142
+ }
143
+ if (!linked) {
144
+ return {
145
+ ok: false,
146
+ error: 'sign in with Google first — this account is not linked to a Google identity yet, '
147
+ + 'and connecting Drive now would attach it to a different account',
148
+ };
149
+ }
150
+ const url = driveConnect.driveConnectUrl(aegis.apiBase);
151
+ const opened = Boolean(openUrl(url));
152
+ try {
153
+ const status = await driveConnect.waitForDriveConnected(aegis);
154
+ return { ok: true, opened, url, connected: status.connected, email: status.email };
155
+ } catch (e) {
156
+ return { ok: false, opened, url, error: (e && e.message) || String(e) };
157
+ }
158
+ } finally {
159
+ connectInFlight = null;
160
+ }
161
+ })();
162
+ }
163
+ return connectInFlight;
164
+ }
165
+
166
+ /**
167
+ * A one-shot read: signed in with Google, and if so, is Drive connected.
168
+ *
169
+ * `signedIn` is the server's `google_linked`, not "has an API key" — see
170
+ * the header. The old `apiKey ? true` form reported every anonymous
171
+ * auto-provisioned device as signed in, which both mislabelled the card
172
+ * ("signed in · Drive not connected yet" on a machine that never touched
173
+ * Google) and enabled Connect Drive ahead of the sign-in it depends on.
174
+ */
175
+ async function status() {
176
+ if (!aegis.apiKey) {
177
+ return { signedIn: false, googleLinked: false, googleLinkedKnown: true, connected: false, email: null };
178
+ }
179
+ const s = await driveConnect.driveStatus(aegis);
180
+ // A completed round-trip outranks a server that cannot answer: without
181
+ // this, signing in against a server that predates `google_linked` leaves
182
+ // the card reading "not signed in" forever, because the sign-in that just
183
+ // succeeded is the one thing the flag was supposed to prove.
184
+ const signedIn = s.googleLinked === true || signedInViaGoogle;
185
+ return {
186
+ signedIn,
187
+ ...s,
188
+ googleLinked: s.googleLinked === true || signedIn,
189
+ googleLinkedKnown: s.googleLinkedKnown || signedIn,
190
+ };
191
+ }
192
+
193
+ /** Pull this account's AEGIS cloud memory and back it up to Drive. */
194
+ async function syncMemoryNow() {
195
+ try {
196
+ const tokenInfo = await aegis.driveToken();
197
+ if (!tokenInfo || !tokenInfo.connected || !tokenInfo.access_token) {
198
+ throw new Error('Google Drive is not connected yet');
199
+ }
200
+ const pulled = await aegis.memoryPull(0);
201
+ const entries = (pulled && pulled.entries) || [];
202
+ const content = JSON.stringify({ exportedAt: new Date().toISOString(), entries }, null, 2);
203
+ const result = await driveUpload.uploadMemoryBackup(tokenInfo.access_token, content);
204
+ lastSyncAt = Date.now();
205
+ lastSyncError = null;
206
+ return {
207
+ ok: true,
208
+ count: entries.length,
209
+ fileId: result.fileId,
210
+ webViewLink: result.webViewLink,
211
+ updated: result.updated,
212
+ lastSyncAt,
213
+ };
214
+ } catch (e) {
215
+ lastSyncError = (e && e.message) || String(e);
216
+ return { ok: false, error: lastSyncError, lastSyncAt };
217
+ }
218
+ }
219
+
220
+ return {
221
+ signInWithGoogle,
222
+ connectDrive,
223
+ status,
224
+ syncMemoryNow,
225
+ state: () => ({ lastSyncAt, lastSyncError }),
226
+ };
227
+ }
228
+
229
+ module.exports = { createGoogleConnect };
@@ -167,6 +167,32 @@ function normalizeToolArgs(T, name, args, ctx) {
167
167
  return fn(name, args, ctx);
168
168
  }
169
169
 
170
+ /**
171
+ * Is `dir` inside a git work tree? Walks up the way git itself does, so a turn
172
+ * running in a subdirectory of a repo is still told it is one.
173
+ *
174
+ * Returns `'git'`, `''` (definitely not a repo), or `null` when the probe could
175
+ * not answer — the caller omits the line rather than asserting "not a repo",
176
+ * because a wrong negative is exactly the kind of fact a model will act on.
177
+ */
178
+ function detectVcs(dir) {
179
+ if (!dir) return null;
180
+ try {
181
+ const fs = require('node:fs');
182
+ const path = require('node:path');
183
+ let probe = path.resolve(String(dir));
184
+ for (let i = 0; i < 8; i += 1) {
185
+ if (fs.existsSync(path.join(probe, '.git'))) return 'git';
186
+ const up = path.dirname(probe);
187
+ if (up === probe) break;
188
+ probe = up;
189
+ }
190
+ return '';
191
+ } catch {
192
+ return null;
193
+ }
194
+ }
195
+
170
196
  function hostEnv(base = {}, supplied = {}, payload = null) {
171
197
  const b = base || {};
172
198
  const s = supplied || {};
@@ -250,6 +276,16 @@ function hostEnv(base = {}, supplied = {}, payload = null) {
250
276
  // a `{ key, value }` facet, or a bare string.
251
277
  languages: pick('languages', b.languages),
252
278
  cwd: pick('cwd', cwd),
279
+ // Whether the turn's directory is a git work tree, probed once per turn.
280
+ // Measured reason: the work rules order the model to check prior work with
281
+ // `git log` before it starts, and a benchmark workspace (a temp dir holding
282
+ // three files) is NOT a repo. On every task the model dutifully spent its
283
+ // second round on `cd … && git log`, paid a failed exec, and carried the
284
+ // error text into every later round's prompt — one wasted round in seven,
285
+ // and a bigger prefix for the rest. The host already knew; this states it
286
+ // so the rule can say "skip the sources the workspace does not have" and
287
+ // mean something the model can act on without guessing.
288
+ vcs: pick('vcs', b.vcs != null ? b.vcs : detectVcs(cwd)),
253
289
  shell: pick('shell', shellName),
254
290
  shellKind: pick('shellKind', shellKind),
255
291
  roots: pick('roots', b.roots),
@@ -709,6 +745,7 @@ const { offerableCatalog } = requireSharedBrain();
709
745
  const { withByokAdditions } = requireSharedByokCatalog();
710
746
  const {
711
747
  withUserModels,
748
+ unmatchedUserProviders,
712
749
  load: loadUserModels,
713
750
  addUserModel,
714
751
  removeUserModel,
@@ -1635,6 +1672,16 @@ function createLocalEngine({
1635
1672
  // it via `fee.require_balance` below.
1636
1673
  let providers = [];
1637
1674
  let fee = null;
1675
+ // Did the CATALOG itself fail to arrive, as opposed to arriving empty?
1676
+ // The two are different states and the picker must not confuse them: a
1677
+ // 5xx / offline relay used to fall into the bare `catch` below and hand
1678
+ // back `providers: []`, which reads on screen exactly like an account
1679
+ // with nothing configured — while the user's provider row and their
1680
+ // `models.json` declaration sat intact on disk. Reported instead, the
1681
+ // same way the 'local' branch reports `listed: false` for a tag list
1682
+ // that did not arrive.
1683
+ let providersFailed = false;
1684
+ let providersError = null;
1638
1685
  try {
1639
1686
  const data = await aegis.byokProviders();
1640
1687
  // The server's catalog, plus the ids it does not name but the vendor
@@ -1652,8 +1699,16 @@ function createLocalEngine({
1652
1699
  // Absent until the server publishes one, and the UI must then say
1653
1700
  // nothing rather than show a guess.
1654
1701
  fee = (data && data.fee) || null;
1655
- } catch {
1702
+ } catch (err) {
1703
+ // Keep the empty list (nothing to offer) but SAY why, so the renderer
1704
+ // can distinguish "catalog unreachable" from "catalog said none". The
1705
+ // message is carried rather than a boolean alone because the two cases
1706
+ // need different copy: offline wants a retry, a server error wants the
1707
+ // status. `err.message` is the transport's own text, already shaped
1708
+ // elsewhere in this file.
1656
1709
  providers = [];
1710
+ providersFailed = true;
1711
+ providersError = (err && err.message) || String(err);
1657
1712
  }
1658
1713
  const models = [];
1659
1714
  for (const p of providers) {
@@ -1689,6 +1744,18 @@ function createLocalEngine({
1689
1744
  needsProviderKey: models.length > 0 && !models.some((m) => m.configured),
1690
1745
  needsAegisKey: !aegis.apiKey,
1691
1746
  fee,
1747
+ // True only when the catalog did not arrive — never merely because it
1748
+ // arrived with no providers. `withUserModels` never fabricates a
1749
+ // provider, so this list is the same either side of the fold above.
1750
+ providersFailed,
1751
+ providersError,
1752
+ // Declarations in `models.json` naming a provider the catalog does not
1753
+ // list — a typo, or a provider renamed/dropped server-side. `withUserModels`
1754
+ // deliberately does NOT invent a row for them (it would 401 on its first
1755
+ // turn), but silence is the other wrong answer: the model stays in the
1756
+ // file and the user is entitled to know the picker cannot show it.
1757
+ // Carried as the ids alone; the renderer names them.
1758
+ unmatchedProviders: unmatchedUserProviders(providers),
1692
1759
  };
1693
1760
  }
1694
1761
  if (cls === 'local') {
@@ -2408,6 +2475,32 @@ function createLocalEngine({
2408
2475
  const refusal = localT.remoteRefusal(cfg.baseURL || localT.DEFAULT_BASE);
2409
2476
  if (refusal) throw refusal;
2410
2477
  }
2478
+ // `custom` is the mirror of `local`: both are rows whose payload is an
2479
+ // ADDRESS rather than a credential, and both are unusable without it.
2480
+ // They differ in what absence MEANS. A blank `local` row falls back to
2481
+ // DEFAULT_BASE (the stock Ollama port) because there is exactly one
2482
+ // sensible address to guess. A blank `custom` row has no default at all —
2483
+ // the catalog descriptor pins none, so the server refuses the call with a
2484
+ // bare 400 that says nothing about which half of the setup is missing.
2485
+ //
2486
+ // Caught here, before any byte goes out, so the message can name the
2487
+ // command that fixes it. Reachable in practice: `/byok-key custom <key>`
2488
+ // stores a key with no address, and the CLI explicitly trusts an
2489
+ // unreachable catalog rather than refusing the id — so a user who typed
2490
+ // the key first and the endpoint never has a row that looks configured.
2491
+ // Keyed on the provider, not the class: every other byok provider's host
2492
+ // is a server-side constant and needs no address from us.
2493
+ if (cls === 'byok' && byokParts.provider === 'custom' && !(cfg.baseURL || '').trim()) {
2494
+ const err = new Error(
2495
+ 'No endpoint saved for the "custom" provider — its key is on this machine but the address ' +
2496
+ 'it dials is not, and no default exists for a vendor AEGIS does not know. Fix: ' +
2497
+ '/model-add custom <https-base-url> <api-key> <model-id> (writes both halves), or the ' +
2498
+ 'endpoint field in Settings › Provider (custom).'
2499
+ );
2500
+ err.status = 400;
2501
+ err.code = 'custom_endpoint_missing';
2502
+ throw err;
2503
+ }
2411
2504
  const base = {
2412
2505
  cls, model, mode: payload && payload.mode, maxTokens, statedMaxTokens, autonomous, sessionId, signal, onDelta, cfg, apiKey, toolChoice,
2413
2506
  // The turn's working directory, computed above as `turnCwd` — `dispatch`
@@ -2563,7 +2656,34 @@ function createLocalEngine({
2563
2656
  // dispatch is summed, including the truncation retry and the synthesis
2564
2657
  // re-dispatch: both are real, separately billed provider calls.
2565
2658
  const turnUsage = { calls: 0 };
2566
- const USAGE_FIELDS = ['input_tokens', 'output_tokens', 'prompt_tokens', 'completion_tokens', 'total_tokens'];
2659
+ // The cache fields MUST be listed here, not just summed at the top level.
2660
+ // DeepSeek reports its hit count as `prompt_cache_hit_tokens` and its
2661
+ // miss count as `prompt_cache_miss_tokens`, and neither name was in this
2662
+ // list — so every tool-using turn dropped them on the way out and handed
2663
+ // its caller a usage object with no cache information at all. The CLI's
2664
+ // `--json` republishes exactly this accumulator, `usageBuckets` then had
2665
+ // nothing to read, and the published cost of an agentic run was priced
2666
+ // entirely at the miss rate. Measured on the 3-task sweep: `cache_read: 0`
2667
+ // on all three trials (tokens_in 210,904) while the row's own baseline
2668
+ // showed 72,448 cached tokens. The same defect the CLI's per-turn readout
2669
+ // was fixed for, here one layer below it.
2670
+ const USAGE_FIELDS = [
2671
+ 'input_tokens', 'output_tokens', 'prompt_tokens', 'completion_tokens', 'total_tokens',
2672
+ // Anthropic spelling (disjoint from its input count).
2673
+ 'cache_read_input_tokens', 'cache_creation_input_tokens',
2674
+ // DeepSeek/OpenAI spelling (both SUBSETs of the prompt count, which is
2675
+ // what `usageBuckets` detects and subtracts rather than double-billing).
2676
+ 'prompt_cache_hit_tokens', 'prompt_cache_miss_tokens',
2677
+ // Hidden deliberation, which every provider here bills as OUTPUT but
2678
+ // none of them split out for free. These were absent from the list
2679
+ // while `reasoningTokensOf` (below) already knew how to read them, so
2680
+ // the one place that could have separated "wrote more code" from
2681
+ // "thought longer" threw the split away at the accumulator and handed
2682
+ // its caller a single `output_tokens` with CoT folded in. A benchmark
2683
+ // row could therefore see that a turn cost 1.9x another's output and
2684
+ // still have no way to say which half moved.
2685
+ 'reasoning_tokens', 'thinking_tokens',
2686
+ ];
2567
2687
  const addUsage = (res) => {
2568
2688
  const u = res && res.usage;
2569
2689
  if (!u || typeof u !== 'object') return;
@@ -2961,6 +3081,22 @@ function createLocalEngine({
2961
3081
  });
2962
3082
  }
2963
3083
  }
3084
+ } catch (e) {
3085
+ // Stamp the stop at the one place it is known for certain.
3086
+ //
3087
+ // `cancel(sessionId)` aborts THIS turn's controller, and the in-flight
3088
+ // transport unwinds as a rejection. The host has to tell that rejection
3089
+ // apart from a genuine provider/transport failure before it can stop
3090
+ // printing Ctrl-X as an application crash — and it used to guess from the
3091
+ // message text, which is exactly how a real failure ("connection aborted
3092
+ // by peer", "upstream aborted the connection") gets relabelled as a
3093
+ // deliberate stop and leaves the user with a silently empty answer. The
3094
+ // message is not evidence. This is: the signal reads aborted only because
3095
+ // the user asked for it, so it cannot be confused with anything upstream
3096
+ // did. Attached to the error object itself so it survives any rethrow,
3097
+ // and read by `isAbortShapedError` in desktop/main.js.
3098
+ if (signal.aborted && e && typeof e === 'object') e.aegisCancelled = true;
3099
+ throw e;
2964
3100
  } finally {
2965
3101
  if (shell) shell.dispose();
2966
3102
  controllers.delete(sessionId);
@@ -203,9 +203,15 @@ function byokFeeLine(fee) {
203
203
  // default-off gate after it. Joined with ' — ' only when there IS a second
204
204
  // clause: a sentence-ending dash with nothing after it ("…provider's bill —.")
205
205
  // is what the first live run of this line actually printed.
206
+ //
207
+ // This rate is the deployment's published RATE, not a quote for the reader
208
+ // specifically — an Unlimited subscriber is waived from this fee entirely
209
+ // (billing_service.is_unlimited on aegis1), so the number above is what a
210
+ // Free-tier BYOK caller pays, not what everyone pays.
206
211
  const price = `Bring your own key: relayed through AEGIS, handled at `
207
212
  + `${ratePerMillion(inRate)} in / ${ratePerMillion(outRate)} out, `
208
- + `on top of your provider's bill`;
213
+ + `on top of your provider's bill — no fee at all on Unlimited, `
214
+ + `and no fee at all during the free 1-month Unlimited trial`;
209
215
  let clause = '';
210
216
  if (f.require_account === true) {
211
217
  clause = 'and needs an AEGIS account key (X-AEGIS-Key) for the fee to be billed';
@@ -30,50 +30,62 @@
30
30
 
31
31
  /** Identity + work rules. Ported from the upstream engine reference. */
32
32
  const MAIN_CHAT_PROMPT =
33
- `You are Aegiscodex, a terminal coding assistant that works in the user's repository. ` +
34
- `Help with software engineering tasks: read and reason about code, write and edit files, ` +
35
- `run shell commands, and investigate bugs. Work rules:\n` +
33
+ `You are Aegiscodex, a terminal coding assistant working in the user's repository: read and ` +
34
+ `reason about code, write and edit files, run shell commands, investigate bugs. Work rules:\n` +
35
+ `- Output is billed by the token and is the largest line on the bill, so spend it where it is ` +
36
+ ` read. Before a tool call, one short sentence on what you are about to find out. Do NOT ` +
37
+ ` restate file contents, diffs, command output or test results the user can already see. ` +
38
+ ` Finish with the answer in a few lines — expand only when asked to explain, or when the ` +
39
+ ` change's correctness depends on a detail the user must check.\n` +
40
+ `- Batch independent tool calls into ONE turn: read several files, or grep for several ` +
41
+ ` patterns, at once, or use one exec command that does the whole job. Every round ` +
42
+ ` re-sends the prefix and bills fresh output; one call per round is the expensive shape.\n` +
36
43
  `- Use tools silently. A one-line reason is enough; do not narrate your plan as a story. ` +
37
44
  `- Never claim what a tool found or what a command returned before the tool actually runs. ` +
38
45
  ` Report only the results you really received. ` +
39
- `- Before you start a fix or feature the user asks for, check whether they have asked for ` +
40
- ` it before and whether it was already delivered: look in recalled memory, ` +
41
- ` ~/.aegiscode/queue.jsonl, PLAN.md/notes, and \`git log\` for the same demand. If it was ` +
42
- ` resolved before and the problem is back, report how it was resolved (the commit, file, or ` +
43
- ` mechanism that fixed it) and why it regressed, then fix the cause — never re-fix blind as ` +
44
- ` if the request were new. If it was never met, say so plainly before starting. ` +
45
- `- Match the response to the request. A greeting, a question about something you already ` +
46
- ` know, or a request for an opinion is answered in plain words with NO tools. Only reach ` +
47
- ` for a tool when the answer genuinely depends on something in this repo or on this ` +
48
- ` machine. "hey" is not a task.\n` +
46
+ `- Before starting a fix or feature, check whether they have asked for it before and whether ` +
47
+ ` it was delivered: recalled memory, \`queue.jsonl\`, \`PLAN.md\`/notes and \`git log\` — ` +
48
+ ` ONE batched command over sources that exist: if the Environment block says this ` +
49
+ ` directory has no version control, skip it. ` +
50
+ ` If it was resolved before and the problem is back, report how it was resolved and ` +
51
+ ` why it regressed, then fix the cause — never re-fix blind as if it were new. ` +
52
+ ` If it was never met, say so plainly before starting. ` +
53
+ `- Match the response to the request. A greeting, a question you already know the answer to, ` +
54
+ ` or a request for an opinion is answered in plain words with NO tools. Reach for a tool ` +
55
+ ` only when the answer depends on something in this repo or on this machine. ` +
56
+ ` "hey" is not a task.\n` +
49
57
  `- WHEN THE USER HAS ASKED FOR WORK: act, don't just inspect. After at most 2 rounds of ` +
50
58
  ` reading or exploration, start making changes with writeFile or editFile. Reconnaissance ` +
51
- ` is not progress — implement, then verify. This rule is about HOW to carry out a task you ` +
52
- ` were given; it is never a reason to invent one. Never write or modify a file the user ` +
59
+ ` is not progress — implement, then verify. This is about HOW to carry out a task you were ` +
60
+ ` given; it is never a reason to invent one. Never write or modify a file the user ` +
53
61
  ` did not ask you to touch.\n` +
54
- `- Zero tolerance for stalling: once the target file/cause is identified and only the edit ` +
55
- ` remains, make it in the same turn. Never end a turn with an empty or no-op response, or a ` +
56
- ` "let me continue"/"I'll now…" placeholder when the user says continue — that is not an ` +
57
- ` answer and it wastes their turn. This has been flagged as non-compliance before; treat it ` +
58
- ` as a hard rule. ` +
59
- `- When you have what you need, stop using tools and give a concise, direct answer to the ` +
60
- ` user's question. Never end your turn with an intention like "Let me check…" or "I'll now…" ` +
61
- ` — that is not an answer. ` +
62
+ `- Zero tolerance for stalling: once the cause is identified and only the edit remains, make ` +
63
+ ` it in the same turn. Never end a turn with an empty or no-op response, or a "let me ` +
64
+ ` continue"/"I'll now…" placeholder when the user says continue — that wastes their turn. ` +
65
+ ` This has been flagged as non-compliance before; treat it as a hard rule. ` +
66
+ `- When you have what you need, stop using tools and give a concise, direct answer. Never ` +
67
+ ` end a turn with an intention like "Let me check…" — that is not an answer. ` +
62
68
  `- When a focused multi-step sub-task can be delegated, use the task tool to spawn a ` +
63
- ` specialist subagent rather than doing everything inline. ` +
69
+ ` specialist subagent rather than doing it inline. ` +
64
70
  `- If the user references a workflow you don't recognize, inspect the repo/scripts for that ` +
65
- ` mechanism before acting — don't assume it means inline work.`;
71
+ ` mechanism before acting.`;
66
72
 
67
- /** The tools the desktop actually advertises (kept in step with tools.js). */
73
+ /**
74
+ * The tool line: what each tool is, and the two facts a schema cannot carry.
75
+ *
76
+ * It used to restate the path rule in full — absolute, `~` expansion, relative
77
+ * resolution — which `PATH_ARG_DESC` already says once per path argument in nine
78
+ * schemas. That duplication is what `d01068b` unwound from the schemas; this is
79
+ * the other end of it. What stays is the part no schema carries: the
80
+ * platform-specific example (the one form a model gets wrong) and the persistent
81
+ * shell, which is a property of the session rather than an argument.
82
+ */
68
83
  const TOOL_LINE =
69
- 'Tools: readFile, writeFile, editFile, listDir, glob, grep, exec, task. Paths are absolute and ' +
70
- 'must be written in the form THIS machine uses — read the platform and home directory in the ' +
71
- 'Environment block below and use them: on win32 that means C:\\Users\\you\\Desktop\\file.txt, ' +
72
- 'not /home/you/Desktop/file.txt or /tmp/file.txt. "~" and "~/…" mean the home directory and are ' +
73
- 'expanded for you; a relative path resolves against the working directory. ' +
74
- 'exec runs in a persistent shell session on this machine — cd and exported env vars carry ' +
75
- 'across calls within the turn, like a real terminal. task spawns a specialist subagent ' +
76
- '(its own tool loop, same model) for a focused, self-contained piece of work.';
84
+ 'Tools: readFile, writeFile, editFile, listDir, glob, grep, exec, task. Paths are absolute ' +
85
+ 'in the form THIS machine uses (see the Environment block): on win32 that means ' +
86
+ 'C:\\Users\\you\\Desktop\\file.txt, not /home/you/Desktop/file.txt or /tmp/file.txt. ' +
87
+ 'exec runs in a persistent shell session — cd and exported env vars carry across calls ' +
88
+ 'within the turn. task spawns a specialist subagent (its own tool loop, same model).';
77
89
 
78
90
  /**
79
91
  * One durable-language entry as a sentence fragment, accepting the three shapes
@@ -109,6 +121,7 @@ function environmentPreamble(env = {}) {
109
121
  localeName,
110
122
  languages,
111
123
  cwd,
124
+ vcs,
112
125
  shell,
113
126
  shellKind,
114
127
  roots,
@@ -164,8 +177,7 @@ function environmentPreamble(env = {}) {
164
177
  if (desktopdir) {
165
178
  bits.push(
166
179
  `desktop directory: ${desktopdir} — use this exact path when the user says "my desktop"; ` +
167
- 'do NOT build it yourself from the home directory, because it is not always a folder ' +
168
- 'called "Desktop" there',
180
+ 'do NOT build it yourself from the home directory',
169
181
  );
170
182
  }
171
183
  // Same failure mode as the desktop line, one directory over, and reported
@@ -176,13 +188,31 @@ function environmentPreamble(env = {}) {
176
188
  // missing path cannot tell a typo from an empty directory.
177
189
  if (dataDir) {
178
190
  bits.push(
179
- `AEGIS data directory: ${dataDir} — this is the store itself (config, sessions, ` +
191
+ `AEGIS data directory: ${dataDir} — the store itself (config, sessions, ` +
180
192
  'credentials, queue). Use this exact path; do NOT append it to the home directory ' +
181
193
  'yourself, because a dropped separator ("C:\\Users\\you.aegiscode") names a path that ' +
182
194
  'does not exist and looks like an empty directory rather than a mistake',
183
195
  );
184
196
  }
185
197
  if (cwd) bits.push(`working directory: ${cwd}`);
198
+ // The version-control fact, which turns the work rules' "check whether this
199
+ // was already delivered" from a guess into a decision. Measured: in a temp
200
+ // workspace with no `.git`, the model still ran `git log` on its second round
201
+ // of all three tasks, spent a round on a guaranteed failure, and carried the
202
+ // error into the rest of the turn. Stated only when the host actually probed
203
+ // (`null`/`undefined` means "could not tell", and silence beats a wrong fact).
204
+ if (vcs === 'git') {
205
+ bits.push(
206
+ `version control: this directory IS inside a git work tree, so \`git log\`, ` +
207
+ '`git diff` and `git status` are usable for checking what was already done',
208
+ );
209
+ } else if (vcs === '') {
210
+ bits.push(
211
+ `version control: this directory is NOT a git repository and has no history to ` +
212
+ 'consult — do NOT run `git log`, `git diff`, `git blame` or `git status`; they ' +
213
+ 'fail here and cost a round. Read the files that exist instead',
214
+ );
215
+ }
186
216
  if (shell) bits.push(`exec shell: ${shell}${shellKind ? ` (${shellKind})` : ''}`);
187
217
  if (model) bits.push(`model: ${model}`);
188
218
  if (appVersion) bits.push(`AEGIS Desktop ${appVersion}`);