teapot-coding-agent 0.37.173 → 0.37.174
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/agent/agent.js
CHANGED
|
@@ -22,11 +22,22 @@ export const AUTO_CONTINUE_NUDGE = "Continue working toward the current goal. If
|
|
|
22
22
|
"When the goal is genuinely met — or truly blocked — call finish() " +
|
|
23
23
|
"(finish(goalComplete=true) when met, finish(goalComplete=false) when blocked) " +
|
|
24
24
|
"instead of replying in prose; prose here does not stop the loop.";
|
|
25
|
+
/**
|
|
26
|
+
* How much of the goal the nudge repeats.
|
|
27
|
+
*
|
|
28
|
+
* The nudge is injected on EVERY auto-continue round, so an unbounded goal would re-enter the context over and over
|
|
29
|
+
* (#136). It is clipped — but the clip is ANNOUNCED, because a silently shortened goal is a requirement the model cannot
|
|
30
|
+
* see and cannot ask about: the same failure as the AGENTS.md slice. The full text stays one `get_goal()` away.
|
|
31
|
+
*/
|
|
32
|
+
const NUDGE_GOAL_CHARS = 8_000;
|
|
25
33
|
export function autoContinueNudge(goalText) {
|
|
26
34
|
const goal = goalText.trim();
|
|
27
35
|
if (!goal)
|
|
28
36
|
return AUTO_CONTINUE_NUDGE;
|
|
29
|
-
|
|
37
|
+
const shown = goal.length > NUDGE_GOAL_CHARS
|
|
38
|
+
? `${goal.slice(0, NUDGE_GOAL_CHARS)}\n… (clipped: ${goal.length - NUDGE_GOAL_CHARS} more characters — call get_goal() for the whole goal)`
|
|
39
|
+
: goal;
|
|
40
|
+
return `${AUTO_CONTINUE_NUDGE}\nCurrent goal:\n=====\n${shown}\n=====`;
|
|
30
41
|
}
|
|
31
42
|
export const PROGRESS_REQUEST = "Please give a brief progress report now: what you are doing, goal progress, " +
|
|
32
43
|
"what you recently tried, any problems, and your next step. Keep it under 10 lines. " +
|
|
@@ -116,9 +127,9 @@ Most session state is not injected into prompts — fetch it with tools instead:
|
|
|
116
127
|
- list_skills() / load_skill(name) / save_skill(...) → reusable playbooks.
|
|
117
128
|
|
|
118
129
|
## Rules
|
|
119
|
-
- Project instructions from AGENTS.md (workspace root) are EMBEDDED in
|
|
120
|
-
prompt below when present — read them here, don't re-read_file it.
|
|
121
|
-
file current by editing it when it goes stale.
|
|
130
|
+
- Project instructions from AGENTS.md (workspace root) are EMBEDDED **in full**
|
|
131
|
+
in this prompt below when present — read them here, don't re-read_file it.
|
|
132
|
+
Keep the file current by editing it when it goes stale.
|
|
122
133
|
- Work step by step with tools. Verify results (run tests/builds) before claiming progress.
|
|
123
134
|
- File changes — pick by scope: write_file (one new file / full rewrite) ·
|
|
124
135
|
edit_file (exactly one small unique replacement) · apply_patch (several
|
|
@@ -476,7 +487,16 @@ export class Agent {
|
|
|
476
487
|
if (!this.skillsCache.length)
|
|
477
488
|
return "";
|
|
478
489
|
const lines = this.skillsCache.map((s) => {
|
|
479
|
-
|
|
490
|
+
/**
|
|
491
|
+
* The description is passed WHOLE.
|
|
492
|
+
*
|
|
493
|
+
* It was cut at 200 characters, and a skill's description is the only thing the model sees when deciding whether to
|
|
494
|
+
* load it — "when a loaded skill matches your task, follow its playbook" is decidable only from that sentence. A
|
|
495
|
+
* truncated one ("…use when verifying work, writing test cases, or rev…") hides exactly the conditions that make
|
|
496
|
+
* the skill relevant, which is the same class of bug as the AGENTS.md truncation: the harness hiding its own
|
|
497
|
+
* instructions and then behaving as if it had given them.
|
|
498
|
+
*/
|
|
499
|
+
const desc = (s.description || "(no description)").replace(/\s+/g, " ");
|
|
480
500
|
return `- ${s.name}: ${desc}`;
|
|
481
501
|
});
|
|
482
502
|
return lines.join("\n");
|
|
@@ -1194,18 +1214,18 @@ export class Agent {
|
|
|
1194
1214
|
});
|
|
1195
1215
|
}
|
|
1196
1216
|
async setGoalVerify(verify) {
|
|
1197
|
-
this.goal = { ...this.goal, verify
|
|
1217
|
+
this.goal = { ...this.goal, verify, updatedAt: new Date().toISOString() };
|
|
1198
1218
|
await this.writeGoalFile();
|
|
1199
1219
|
await this.log.append("goal", this.currentSession, this.currentBranch, {
|
|
1200
1220
|
event: "verify-set",
|
|
1201
|
-
verify
|
|
1221
|
+
verify,
|
|
1202
1222
|
});
|
|
1203
1223
|
}
|
|
1204
1224
|
async setGoalAudit(verdict, feedback) {
|
|
1205
1225
|
this.goal = {
|
|
1206
1226
|
...this.goal,
|
|
1207
1227
|
updatedAt: new Date().toISOString(),
|
|
1208
|
-
audit: { verdict, feedback
|
|
1228
|
+
audit: { verdict, feedback, at: new Date().toISOString() },
|
|
1209
1229
|
};
|
|
1210
1230
|
if (verdict === "approved")
|
|
1211
1231
|
this.goal.status = "done";
|
|
@@ -1222,7 +1242,11 @@ export class Agent {
|
|
|
1222
1242
|
await this.log.append("goal", this.currentSession, this.currentBranch, {
|
|
1223
1243
|
event: "audit",
|
|
1224
1244
|
verdict,
|
|
1225
|
-
|
|
1245
|
+
/**
|
|
1246
|
+
* Not cut: `feedback` is what the auditor tells the NEXT turn to fix, so a truncation silently drops requirements —
|
|
1247
|
+
* the same failure mode as the AGENTS.md slice. The bound is only a runaway guard.
|
|
1248
|
+
*/
|
|
1249
|
+
feedback: feedback.slice(0, 20_000),
|
|
1226
1250
|
});
|
|
1227
1251
|
}
|
|
1228
1252
|
snapshot() {
|
|
@@ -1251,7 +1275,7 @@ export class Agent {
|
|
|
1251
1275
|
goal: {
|
|
1252
1276
|
status: this.goal.status,
|
|
1253
1277
|
text: this.goal.text,
|
|
1254
|
-
...(this.goal.verify ? { verify: this.goal.verify
|
|
1278
|
+
...(this.goal.verify ? { verify: this.goal.verify } : {}),
|
|
1255
1279
|
...(this.goal.audit ? { audit: this.goal.audit } : {}),
|
|
1256
1280
|
},
|
|
1257
1281
|
latestProgress: this.latestProgress,
|
|
@@ -1753,7 +1777,7 @@ export class Agent {
|
|
|
1753
1777
|
}
|
|
1754
1778
|
const isSub = !!this.opts.parent;
|
|
1755
1779
|
const nudge = isSub
|
|
1756
|
-
? this.harnessMessage(`[harness] Auto-nudge. Re-anchor on YOUR task — and only that:\n\n${this.goal.text.slice(0,
|
|
1780
|
+
? this.harnessMessage(`[harness] Auto-nudge. Re-anchor on YOUR task — and only that:\n\n${this.goal.text.length > NUDGE_GOAL_CHARS ? this.goal.text.slice(0, NUDGE_GOAL_CHARS) + "\n… (clipped — call get_goal() for the whole goal)" : this.goal.text}\n\nThe inherited conversation is reference only; other agents own any older tasks in it. If your task is complete, call finish() now instead of continuing. If blocked, explain why briefly and finish().`)
|
|
1757
1781
|
: this.harnessMessage(autoContinueNudge(this.goal.text));
|
|
1758
1782
|
await this.log.append("prompt", this.currentSession, this.currentBranch, {
|
|
1759
1783
|
source: "harness",
|
|
@@ -1959,7 +1983,7 @@ export class Agent {
|
|
|
1959
1983
|
if (text)
|
|
1960
1984
|
await this.setGoal(text);
|
|
1961
1985
|
if (typeof a.verify === "string" && a.verify.trim())
|
|
1962
|
-
await this.setGoalVerify(a.verify.trim().slice(0,
|
|
1986
|
+
await this.setGoalVerify(a.verify.trim().slice(0, 20_000));
|
|
1963
1987
|
await this.answerMeta(call, text ? "goal updated" : "empty goal rejected");
|
|
1964
1988
|
return "next";
|
|
1965
1989
|
},
|
|
@@ -1974,7 +1998,8 @@ export class Agent {
|
|
|
1974
1998
|
},
|
|
1975
1999
|
ask_operator: async (call, m) => {
|
|
1976
2000
|
const a = safeParse(call.function.arguments);
|
|
1977
|
-
|
|
2001
|
+
// a question cut at 2,000 characters asks the operator something else than what the model meant
|
|
2002
|
+
const question = String(a.question ?? "").slice(0, 20_000);
|
|
1978
2003
|
const options = Array.isArray(a.options)
|
|
1979
2004
|
? a.options
|
|
1980
2005
|
.slice(0, 6)
|
|
@@ -2558,19 +2583,34 @@ export class Agent {
|
|
|
2558
2583
|
const head = [{ role: "system", content: sys }];
|
|
2559
2584
|
return hasSystem ? [...head, ...this.messages.slice(1)] : [...head, ...this.messages];
|
|
2560
2585
|
}
|
|
2561
|
-
|
|
2562
|
-
|
|
2563
|
-
|
|
2564
|
-
|
|
2586
|
+
/**
|
|
2587
|
+
* AGENTS.md is passed WHOLE, and read in one place.
|
|
2588
|
+
*
|
|
2589
|
+
* It was `readFileSync(...).slice(0, 12_000)` in three places — MEASURED on this repository: 56,290 characters, so
|
|
2590
|
+
* ~79% of the project's rules never reached the model, while the prompt line says "EMBEDDED in this prompt below when
|
|
2591
|
+
* present — read them here, don't re-read_file it". The rules that were cut are exactly the ones written after
|
|
2592
|
+
* repeated bugs (reproduce before fixing, verify the PATH a helper is called from, prefer the real measurement over a
|
|
2593
|
+
* synthetic test), so the harness hid its own hard-won knowledge and then behaved as if it had been told.
|
|
2594
|
+
*
|
|
2595
|
+
* `reloadAgentsMd` compared the same truncated prefix, so an edit in the second half was invisible to the reload too.
|
|
2596
|
+
*/
|
|
2597
|
+
readAgentsMd() {
|
|
2565
2598
|
try {
|
|
2566
2599
|
const p = path.join(this.opts.workspace, "AGENTS.md");
|
|
2567
2600
|
if (!existsSync(p))
|
|
2568
|
-
return
|
|
2569
|
-
|
|
2601
|
+
return "";
|
|
2602
|
+
return readFileSync(p, "utf8");
|
|
2570
2603
|
}
|
|
2571
2604
|
catch {
|
|
2572
|
-
return
|
|
2605
|
+
return "";
|
|
2573
2606
|
}
|
|
2607
|
+
}
|
|
2608
|
+
reloadAgentsMd() {
|
|
2609
|
+
if (!this.agentsMdLoaded)
|
|
2610
|
+
return null;
|
|
2611
|
+
const fresh = this.readAgentsMd();
|
|
2612
|
+
if (!fresh)
|
|
2613
|
+
return null;
|
|
2574
2614
|
if (fresh === this.agentsMd)
|
|
2575
2615
|
return null;
|
|
2576
2616
|
this.agentsMd = fresh;
|
|
@@ -2595,14 +2635,7 @@ export class Agent {
|
|
|
2595
2635
|
this.wsSnapshot = "(empty workspace)";
|
|
2596
2636
|
}
|
|
2597
2637
|
if (!this.wsSnapshot) {
|
|
2598
|
-
|
|
2599
|
-
this.agentsMd = existsSync(path.join(this.opts.workspace, "AGENTS.md"))
|
|
2600
|
-
? readFileSync(path.join(this.opts.workspace, "AGENTS.md"), "utf8").slice(0, 12_000)
|
|
2601
|
-
: "";
|
|
2602
|
-
}
|
|
2603
|
-
catch {
|
|
2604
|
-
this.agentsMd = "";
|
|
2605
|
-
}
|
|
2638
|
+
this.agentsMd = this.readAgentsMd();
|
|
2606
2639
|
try {
|
|
2607
2640
|
const entries = readdirSync(this.opts.workspace, { withFileTypes: true });
|
|
2608
2641
|
const lines = [];
|
|
@@ -2633,21 +2666,20 @@ export class Agent {
|
|
|
2633
2666
|
}
|
|
2634
2667
|
}
|
|
2635
2668
|
else {
|
|
2636
|
-
|
|
2637
|
-
this.agentsMd = existsSync(path.join(this.opts.workspace, "AGENTS.md"))
|
|
2638
|
-
? readFileSync(path.join(this.opts.workspace, "AGENTS.md"), "utf8").slice(0, 12_000)
|
|
2639
|
-
: "";
|
|
2640
|
-
}
|
|
2641
|
-
catch {
|
|
2642
|
-
this.agentsMd = "";
|
|
2643
|
-
}
|
|
2669
|
+
this.agentsMd = this.readAgentsMd();
|
|
2644
2670
|
}
|
|
2645
2671
|
this.skillCatalogue = this.skillCatalogueText();
|
|
2646
2672
|
this.memoryDigest = this.memoryDigestText();
|
|
2647
2673
|
}
|
|
2648
2674
|
let sys = SYSTEM_TEMPLATE;
|
|
2649
2675
|
if (this.agentsMd.trim()) {
|
|
2650
|
-
|
|
2676
|
+
// the whole file, with its size stated so the model can tell "this is everything" from "this was cut"
|
|
2677
|
+
// the heading stays stable (tests and the model both key on it); the size line says "this is the whole file", so a
|
|
2678
|
+
// model cannot mistake it for an excerpt the way it could when 79% was silently cut
|
|
2679
|
+
sys +=
|
|
2680
|
+
`\n\n## Project instructions (AGENTS.md)\n` +
|
|
2681
|
+
`(complete file, ${this.agentsMd.length.toLocaleString()} characters)\n\n` +
|
|
2682
|
+
this.agentsMd;
|
|
2651
2683
|
}
|
|
2652
2684
|
if (this.wsSnapshot) {
|
|
2653
2685
|
sys += `\n\n## Workspace snapshot (top level)\n${this.wsSnapshot}\nUse list_dir/read_file for anything deeper — do NOT list again what is already here.`;
|