@pugi/cli 0.1.0-beta.2 → 0.1.0-beta.20

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.
Files changed (130) hide show
  1. package/THIRD_PARTY_NOTICES.md +40 -0
  2. package/assets/pugi-mascot.ansi +15 -40
  3. package/bin/run.js +33 -1
  4. package/dist/commands/jobs-watch.js +201 -0
  5. package/dist/commands/jobs.js +15 -0
  6. package/dist/core/agent-progress/cleanup.js +134 -0
  7. package/dist/core/agent-progress/schema.js +144 -0
  8. package/dist/core/agent-progress/writer.js +101 -0
  9. package/dist/core/compact/auto-trigger.js +96 -0
  10. package/dist/core/compact/buffer-rewriter.js +115 -0
  11. package/dist/core/compact/summarizer.js +196 -0
  12. package/dist/core/compact/token-counter.js +108 -0
  13. package/dist/core/consensus/diff-capture.js +73 -0
  14. package/dist/core/context/index.js +7 -0
  15. package/dist/core/context/markdown-traverse.js +255 -0
  16. package/dist/core/cost/rate-card.js +129 -0
  17. package/dist/core/cost/tracker.js +221 -0
  18. package/dist/core/denial-tracking/index.js +8 -0
  19. package/dist/core/denial-tracking/state.js +264 -0
  20. package/dist/core/diagnostics/probe-runner.js +93 -0
  21. package/dist/core/diagnostics/probes/api.js +46 -0
  22. package/dist/core/diagnostics/probes/auth.js +86 -0
  23. package/dist/core/diagnostics/probes/cli-version.js +127 -0
  24. package/dist/core/diagnostics/probes/config.js +72 -0
  25. package/dist/core/diagnostics/probes/denial-tracking.js +57 -0
  26. package/dist/core/diagnostics/probes/disk.js +81 -0
  27. package/dist/core/diagnostics/probes/git.js +65 -0
  28. package/dist/core/diagnostics/probes/mcp.js +75 -0
  29. package/dist/core/diagnostics/probes/node.js +59 -0
  30. package/dist/core/diagnostics/probes/pnpm.js +36 -0
  31. package/dist/core/diagnostics/probes/session.js +74 -0
  32. package/dist/core/diagnostics/probes/status-snapshot.js +442 -0
  33. package/dist/core/diagnostics/probes/workspace.js +63 -0
  34. package/dist/core/diagnostics/types.js +70 -0
  35. package/dist/core/edits/dispatch.js +218 -2
  36. package/dist/core/edits/journal.js +199 -0
  37. package/dist/core/edits/layer-d-ast.js +557 -14
  38. package/dist/core/edits/verify-hook.js +273 -0
  39. package/dist/core/edits/worktree.js +111 -18
  40. package/dist/core/engine/anvil-client.js +115 -5
  41. package/dist/core/engine/budgets.js +89 -0
  42. package/dist/core/engine/context-prefix.js +155 -0
  43. package/dist/core/engine/intent.js +260 -0
  44. package/dist/core/engine/native-pugi.js +744 -210
  45. package/dist/core/engine/prompts.js +61 -6
  46. package/dist/core/engine/strip-internal-fields.js +124 -0
  47. package/dist/core/engine/tool-bridge.js +818 -31
  48. package/dist/core/file-cache.js +113 -1
  49. package/dist/core/init/scaffold.js +195 -0
  50. package/dist/core/lsp/client.js +174 -29
  51. package/dist/core/mcp/client.js +75 -6
  52. package/dist/core/mcp/http-server.js +553 -0
  53. package/dist/core/mcp/permission.js +190 -0
  54. package/dist/core/mcp/registry.js +24 -2
  55. package/dist/core/mcp/server-tools.js +219 -0
  56. package/dist/core/mcp/server.js +397 -0
  57. package/dist/core/permissions/gate.js +187 -0
  58. package/dist/core/permissions/index.js +18 -0
  59. package/dist/core/permissions/mode.js +102 -0
  60. package/dist/core/permissions/state.js +160 -0
  61. package/dist/core/permissions/tool-class.js +93 -0
  62. package/dist/core/repl/codebase-survey.js +308 -0
  63. package/dist/core/repl/history.js +11 -1
  64. package/dist/core/repl/init-interview.js +457 -0
  65. package/dist/core/repl/model-pricing.js +135 -0
  66. package/dist/core/repl/onboarding-state.js +297 -0
  67. package/dist/core/repl/session.js +719 -29
  68. package/dist/core/repl/slash-commands.js +133 -9
  69. package/dist/core/retry-budget/budget.js +284 -0
  70. package/dist/core/retry-budget/index.js +5 -0
  71. package/dist/core/settings.js +71 -0
  72. package/dist/core/skills/defaults.js +457 -0
  73. package/dist/core/subagents/dispatcher-real.js +600 -0
  74. package/dist/core/subagents/dispatcher.js +113 -24
  75. package/dist/core/subagents/index.js +18 -5
  76. package/dist/core/subagents/isolation-matrix.js +213 -0
  77. package/dist/core/subagents/spawn.js +19 -4
  78. package/dist/core/transport/version-interceptor.js +166 -0
  79. package/dist/index.js +28 -0
  80. package/dist/runtime/bootstrap.js +190 -0
  81. package/dist/runtime/cli.js +1588 -266
  82. package/dist/runtime/commands/compact.js +296 -0
  83. package/dist/runtime/commands/cost.js +199 -0
  84. package/dist/runtime/commands/delegate.js +289 -0
  85. package/dist/runtime/commands/doctor.js +369 -0
  86. package/dist/runtime/commands/lsp.js +187 -5
  87. package/dist/runtime/commands/mcp.js +824 -0
  88. package/dist/runtime/commands/patch.js +17 -0
  89. package/dist/runtime/commands/permissions.js +87 -0
  90. package/dist/runtime/commands/report.js +299 -0
  91. package/dist/runtime/commands/review-consensus.js +17 -2
  92. package/dist/runtime/commands/roster.js +117 -0
  93. package/dist/runtime/commands/status.js +178 -0
  94. package/dist/runtime/commands/worktree.js +50 -6
  95. package/dist/runtime/headless.js +543 -0
  96. package/dist/runtime/load-hooks-or-exit.js +71 -0
  97. package/dist/runtime/plan-decompose.js +531 -0
  98. package/dist/runtime/version.js +65 -0
  99. package/dist/tools/agent-tool.js +206 -0
  100. package/dist/tools/apply-patch.js +281 -39
  101. package/dist/tools/ask-user-question.js +213 -0
  102. package/dist/tools/ask-user.js +115 -0
  103. package/dist/tools/file-tools.js +85 -14
  104. package/dist/tools/mcp-tool.js +260 -0
  105. package/dist/tools/multi-edit.js +361 -0
  106. package/dist/tools/registry.js +22 -2
  107. package/dist/tools/skill-tool.js +96 -0
  108. package/dist/tools/tasks.js +208 -0
  109. package/dist/tools/web-fetch.js +147 -2
  110. package/dist/tools/web-search.js +458 -0
  111. package/dist/tui/agent-progress-card.js +111 -0
  112. package/dist/tui/agent-tree.js +10 -0
  113. package/dist/tui/ask-modal.js +2 -2
  114. package/dist/tui/ask-user-question-prompt.js +192 -0
  115. package/dist/tui/compact-banner.js +54 -0
  116. package/dist/tui/conversation-pane.js +69 -8
  117. package/dist/tui/cost-table.js +111 -0
  118. package/dist/tui/doctor-table.js +31 -0
  119. package/dist/tui/input-box.js +1 -1
  120. package/dist/tui/markdown-render.js +4 -4
  121. package/dist/tui/repl-render.js +276 -37
  122. package/dist/tui/repl-splash.js +2 -2
  123. package/dist/tui/repl.js +25 -6
  124. package/dist/tui/splash.js +1 -1
  125. package/dist/tui/status-bar.js +94 -16
  126. package/dist/tui/status-table.js +7 -0
  127. package/dist/tui/tool-stream-pane.js +7 -0
  128. package/dist/tui/update-banner.js +20 -2
  129. package/docs/examples/codegraph.mcp.json +10 -0
  130. package/package.json +9 -6
@@ -0,0 +1,144 @@
1
+ /**
2
+ * Agent progress JSON schema — α7 live-progress sprint (2026-05-27).
3
+ *
4
+ * Long-running agents (spawned externally OR via `pugi /agent`) emit a
5
+ * single JSON document per agent to `~/.pugi/agent-progress/<id>.json`.
6
+ * `pugi jobs --watch` tails the directory via chokidar and re-renders
7
+ * an Ink TUI that mirrors the Claude Code `/compact` visual pattern
8
+ * (header with elapsed + token counter, unicode progress bar, milestone
9
+ * list with done/active/pending status icons).
10
+ *
11
+ * Schema is intentionally optimistic — every numeric field is clamped
12
+ * by the writer/reader so a malformed document degrades to a partial
13
+ * card instead of crashing the watcher. The `pendingCount` /
14
+ * `completedCount` fields are pre-computed by the agent for the
15
+ * "… +N pending, M completed" footer; the renderer never re-counts
16
+ * (the agent may have collapsed milestone history to save bytes).
17
+ */
18
+ /**
19
+ * Validate an unknown payload as an `AgentProgress` document. Returns
20
+ * the typed value on success and a string error otherwise. We deliberately
21
+ * keep this hand-rolled (no zod) — every field is checked exactly once,
22
+ * the error message is human-readable, and zero runtime deps.
23
+ */
24
+ export function validateAgentProgress(value) {
25
+ if (typeof value !== 'object' || value === null) {
26
+ return { ok: false, error: 'progress payload must be a JSON object' };
27
+ }
28
+ const raw = value;
29
+ const requiredString = (field) => {
30
+ const v = raw[field];
31
+ return typeof v === 'string' && v.length > 0 ? v : null;
32
+ };
33
+ const agentId = requiredString('agentId');
34
+ if (!agentId)
35
+ return { ok: false, error: 'agentId required (non-empty string)' };
36
+ if (!/^[a-zA-Z0-9_-]+$/.test(agentId)) {
37
+ return { ok: false, error: 'agentId must match [a-zA-Z0-9_-]+ (filename-safe)' };
38
+ }
39
+ const agentType = requiredString('agentType');
40
+ if (!agentType)
41
+ return { ok: false, error: 'agentType required (non-empty string)' };
42
+ const task = requiredString('task');
43
+ if (!task)
44
+ return { ok: false, error: 'task required (non-empty string)' };
45
+ const startedAt = requiredString('startedAt');
46
+ if (!startedAt)
47
+ return { ok: false, error: 'startedAt required (ISO string)' };
48
+ if (Number.isNaN(Date.parse(startedAt))) {
49
+ return { ok: false, error: 'startedAt must be a parseable ISO timestamp' };
50
+ }
51
+ const lastUpdate = requiredString('lastUpdate');
52
+ if (!lastUpdate)
53
+ return { ok: false, error: 'lastUpdate required (ISO string)' };
54
+ if (Number.isNaN(Date.parse(lastUpdate))) {
55
+ return { ok: false, error: 'lastUpdate must be a parseable ISO timestamp' };
56
+ }
57
+ const elapsedMs = raw.elapsedMs;
58
+ if (typeof elapsedMs !== 'number' || !Number.isFinite(elapsedMs) || elapsedMs < 0) {
59
+ return { ok: false, error: 'elapsedMs required (non-negative number)' };
60
+ }
61
+ const percentComplete = raw.percentComplete;
62
+ if (typeof percentComplete !== 'number' ||
63
+ !Number.isFinite(percentComplete)) {
64
+ return { ok: false, error: 'percentComplete required (number 0..100)' };
65
+ }
66
+ const clampedPercent = Math.max(0, Math.min(100, percentComplete));
67
+ const status = raw.status;
68
+ if (status !== 'running' && status !== 'completed' && status !== 'failed') {
69
+ return { ok: false, error: 'status must be running | completed | failed' };
70
+ }
71
+ const currentStep = raw.currentStep;
72
+ if (typeof currentStep !== 'number' || currentStep < 0) {
73
+ return { ok: false, error: 'currentStep required (non-negative number)' };
74
+ }
75
+ const totalSteps = raw.totalSteps;
76
+ if (typeof totalSteps !== 'number' || totalSteps < 0) {
77
+ return { ok: false, error: 'totalSteps required (non-negative number)' };
78
+ }
79
+ const stepDescription = typeof raw.stepDescription === 'string'
80
+ ? raw.stepDescription
81
+ : '';
82
+ const milestonesRaw = raw.milestones;
83
+ if (!Array.isArray(milestonesRaw)) {
84
+ return { ok: false, error: 'milestones required (array, may be empty)' };
85
+ }
86
+ const milestones = [];
87
+ for (let i = 0; i < milestonesRaw.length; i += 1) {
88
+ const m = milestonesRaw[i];
89
+ if (typeof m !== 'object' || m === null) {
90
+ return { ok: false, error: `milestones[${i}] must be an object` };
91
+ }
92
+ const mr = m;
93
+ const label = typeof mr.label === 'string' ? mr.label : '';
94
+ if (!label) {
95
+ return { ok: false, error: `milestones[${i}].label required` };
96
+ }
97
+ const mStatus = mr.status;
98
+ if (mStatus !== 'done' && mStatus !== 'active' && mStatus !== 'pending') {
99
+ return {
100
+ ok: false,
101
+ error: `milestones[${i}].status must be done | active | pending`,
102
+ };
103
+ }
104
+ const ts = typeof mr.ts === 'string' ? mr.ts : undefined;
105
+ milestones.push({ label, status: mStatus, ts });
106
+ }
107
+ const tokensUsed = typeof raw.tokensUsed === 'number' && Number.isFinite(raw.tokensUsed)
108
+ ? Math.max(0, raw.tokensUsed)
109
+ : undefined;
110
+ const pendingCount = typeof raw.pendingCount === 'number' && Number.isFinite(raw.pendingCount)
111
+ ? Math.max(0, Math.floor(raw.pendingCount))
112
+ : undefined;
113
+ const completedCount = typeof raw.completedCount === 'number' && Number.isFinite(raw.completedCount)
114
+ ? Math.max(0, Math.floor(raw.completedCount))
115
+ : undefined;
116
+ return {
117
+ ok: true,
118
+ value: {
119
+ agentId,
120
+ agentType,
121
+ task,
122
+ startedAt,
123
+ lastUpdate,
124
+ elapsedMs,
125
+ tokensUsed,
126
+ percentComplete: clampedPercent,
127
+ status,
128
+ currentStep,
129
+ totalSteps,
130
+ stepDescription,
131
+ milestones,
132
+ pendingCount,
133
+ completedCount,
134
+ },
135
+ };
136
+ }
137
+ /**
138
+ * Default directory the writer and watcher share. Lives under the
139
+ * project workspace by convention so worktree-isolated agents can emit
140
+ * progress without crossing tenant boundaries. Operators can override
141
+ * via `PUGI_AGENT_PROGRESS_DIR` env var.
142
+ */
143
+ export const DEFAULT_AGENT_PROGRESS_DIRNAME = '.pugi/agent-progress';
144
+ //# sourceMappingURL=schema.js.map
@@ -0,0 +1,101 @@
1
+ /**
2
+ * Agent progress writer — atomic file ops so the chokidar watcher
3
+ * never reads a half-written JSON document.
4
+ *
5
+ * Pattern (POSIX-portable):
6
+ * 1. Build the canonical JSON body.
7
+ * 2. Write to `<dir>/<agentId>.json.tmp-<pid>-<seq>`.
8
+ * 3. Rename (atomic on the same filesystem) to `<dir>/<agentId>.json`.
9
+ * 4. Cleanup is a no-op on success; on failure the caller bubbles the
10
+ * error and the .tmp file is removed best-effort.
11
+ *
12
+ * The auto-cleanup of completed entries lives in `cleanup.ts` (commit 4);
13
+ * the writer here is the minimal surface the agent prompt boilerplate
14
+ * needs to copy + paste.
15
+ */
16
+ import { existsSync, mkdirSync, renameSync, rmSync, writeFileSync } from 'node:fs';
17
+ import { homedir } from 'node:os';
18
+ import { dirname, join, resolve } from 'node:path';
19
+ import { DEFAULT_AGENT_PROGRESS_DIRNAME, validateAgentProgress, } from './schema.js';
20
+ let writeSequence = 0;
21
+ /**
22
+ * Resolve the effective progress directory. Public so the watcher can
23
+ * share the same resolution rules without duplicating env-var logic.
24
+ */
25
+ export function resolveProgressDir(override) {
26
+ const candidate = override
27
+ ?? process.env.PUGI_AGENT_PROGRESS_DIR
28
+ ?? join(process.cwd(), DEFAULT_AGENT_PROGRESS_DIRNAME);
29
+ if (candidate.startsWith('~/')) {
30
+ return join(homedir(), candidate.slice(2));
31
+ }
32
+ return resolve(candidate);
33
+ }
34
+ /**
35
+ * Write a progress document atomically.
36
+ *
37
+ * Throws on validation failure so the caller is loud about a schema
38
+ * regression. File-system errors propagate untouched.
39
+ */
40
+ export function writeProgress(progress, options = {}) {
41
+ const validation = validateAgentProgress(progress);
42
+ if (!validation.ok) {
43
+ throw new Error(`invalid agent-progress payload: ${validation.error}`);
44
+ }
45
+ const dir = resolveProgressDir(options.dir);
46
+ if (!existsSync(dir)) {
47
+ mkdirSync(dir, { recursive: true });
48
+ }
49
+ const finalPath = join(dir, `${progress.agentId}.json`);
50
+ writeSequence += 1;
51
+ const tmpPath = `${finalPath}.tmp-${process.pid}-${writeSequence}`;
52
+ const body = `${JSON.stringify(progress, null, 2)}\n`;
53
+ try {
54
+ // mkdirSync above may have created `dir`, but a parallel agent could
55
+ // delete it between the existsSync and writeFile — ensure the parent
56
+ // of the tmp file exists right before the write.
57
+ const parent = dirname(tmpPath);
58
+ if (!existsSync(parent)) {
59
+ mkdirSync(parent, { recursive: true });
60
+ }
61
+ writeFileSync(tmpPath, body, 'utf8');
62
+ renameSync(tmpPath, finalPath);
63
+ }
64
+ catch (err) {
65
+ try {
66
+ rmSync(tmpPath, { force: true });
67
+ }
68
+ catch {
69
+ // best-effort cleanup; surface the original error
70
+ }
71
+ throw err;
72
+ }
73
+ return { path: finalPath };
74
+ }
75
+ /**
76
+ * Build a fresh progress document with sensible defaults. Mostly a
77
+ * convenience for tests + the agent prompt boilerplate — the writer
78
+ * does NOT call this implicitly because agents have richer context
79
+ * about their own milestone state.
80
+ */
81
+ export function makeInitialProgress(input) {
82
+ const now = input.now ?? Date.now;
83
+ const iso = new Date(now()).toISOString();
84
+ return {
85
+ agentId: input.agentId,
86
+ agentType: input.agentType,
87
+ task: input.task,
88
+ startedAt: iso,
89
+ lastUpdate: iso,
90
+ elapsedMs: 0,
91
+ percentComplete: 0,
92
+ status: 'running',
93
+ currentStep: 0,
94
+ totalSteps: Math.max(0, Math.floor(input.totalSteps)),
95
+ stepDescription: '',
96
+ milestones: input.milestones ?? [],
97
+ pendingCount: input.milestones?.filter((m) => m.status === 'pending').length,
98
+ completedCount: input.milestones?.filter((m) => m.status === 'done').length,
99
+ };
100
+ }
101
+ //# sourceMappingURL=writer.js.map
@@ -0,0 +1,96 @@
1
+ /**
2
+ * Auto-compact threshold gate.
3
+ *
4
+ * Decides whether the conversation buffer has crossed the threshold
5
+ * percent of the active model's context window. The check runs after
6
+ * every operator/persona turn before the NEXT operator input lands so
7
+ * the compaction completes BEFORE the model would have rejected the
8
+ * request with a context-overflow error.
9
+ *
10
+ * Design choices:
11
+ *
12
+ * - Pure function. The caller passes (tokenCount, windowSize, env).
13
+ * The gate returns a verdict; the session module owns the side
14
+ * effect of invoking the summariser. Pure-function shape keeps the
15
+ * spec exhaustive and the call site readable.
16
+ *
17
+ * - Hysteresis: once a compaction lands, the marker resets the
18
+ * baseline token count to "summary + tail" — the gate looks at the
19
+ * POST-marker tokens only. This is enforced upstream by the caller
20
+ * passing the post-marker count; the gate itself has no memory.
21
+ *
22
+ * - Two env knobs:
23
+ * PUGI_AUTOCOMPACT_DISABLED=1 — kill switch
24
+ * PUGI_AUTOCOMPACT_THRESHOLD=N — float in (0, 1] (default 0.75)
25
+ * Anything outside (0, 1] is rejected and the gate falls back to
26
+ * the default. Bad input never crashes the REPL.
27
+ */
28
+ /** Default trip point as a fraction of the context window. */
29
+ export const DEFAULT_THRESHOLD = 0.75;
30
+ /**
31
+ * Decide whether to fire `/compact` automatically. Pure; safe to call
32
+ * after every turn.
33
+ */
34
+ export function evaluateAutoCompact(input) {
35
+ const env = input.env ?? process.env;
36
+ const threshold = resolveThreshold(env);
37
+ if (input.windowSize <= 0 || !Number.isFinite(input.windowSize)) {
38
+ return {
39
+ kind: 'skip',
40
+ reason: 'invalid-window',
41
+ tokenCount: input.tokenCount,
42
+ windowSize: input.windowSize,
43
+ threshold,
44
+ pressure: 0,
45
+ };
46
+ }
47
+ if (env['PUGI_AUTOCOMPACT_DISABLED'] === '1') {
48
+ return {
49
+ kind: 'skip',
50
+ reason: 'disabled',
51
+ tokenCount: input.tokenCount,
52
+ windowSize: input.windowSize,
53
+ threshold,
54
+ pressure: roundPressure(input.tokenCount / input.windowSize),
55
+ };
56
+ }
57
+ const pressure = roundPressure(input.tokenCount / input.windowSize);
58
+ if (pressure >= threshold) {
59
+ return {
60
+ kind: 'fire',
61
+ tokenCount: input.tokenCount,
62
+ windowSize: input.windowSize,
63
+ threshold,
64
+ pressure,
65
+ };
66
+ }
67
+ return {
68
+ kind: 'skip',
69
+ reason: 'below-threshold',
70
+ tokenCount: input.tokenCount,
71
+ windowSize: input.windowSize,
72
+ threshold,
73
+ pressure,
74
+ };
75
+ }
76
+ /**
77
+ * Resolve the threshold from env, clamping to the (0, 1] open-closed
78
+ * interval. Bad input silently falls back to DEFAULT_THRESHOLD so the
79
+ * REPL never crashes on a malformed environment variable.
80
+ */
81
+ function resolveThreshold(env) {
82
+ const raw = env['PUGI_AUTOCOMPACT_THRESHOLD'];
83
+ if (!raw)
84
+ return DEFAULT_THRESHOLD;
85
+ const parsed = Number.parseFloat(raw);
86
+ if (!Number.isFinite(parsed) || parsed <= 0 || parsed > 1) {
87
+ return DEFAULT_THRESHOLD;
88
+ }
89
+ return parsed;
90
+ }
91
+ function roundPressure(raw) {
92
+ if (!Number.isFinite(raw) || raw < 0)
93
+ return 0;
94
+ return Math.round(raw * 1000) / 1000;
95
+ }
96
+ //# sourceMappingURL=auto-trigger.js.map
@@ -0,0 +1,115 @@
1
+ /**
2
+ * Type guard: discriminate a SessionEvent against the `compaction`
3
+ * kind. Used by replay code so the boundary marker drives a different
4
+ * code path than `user`/`persona`/`system` transcript rows.
5
+ */
6
+ export function isCompactBoundary(event) {
7
+ if (event.kind !== 'compaction')
8
+ return false;
9
+ const p = event.payload;
10
+ if (p === null || typeof p !== 'object')
11
+ return false;
12
+ if (p.version !== 1)
13
+ return false;
14
+ if (p.trigger !== 'manual' && p.trigger !== 'auto')
15
+ return false;
16
+ if (typeof p.summary !== 'string' || p.summary.length === 0)
17
+ return false;
18
+ if (typeof p.summaryTokenCount !== 'number')
19
+ return false;
20
+ if (typeof p.summaryTurnsBefore !== 'number')
21
+ return false;
22
+ if (typeof p.keptTailTurns !== 'number')
23
+ return false;
24
+ if (typeof p.coversUntilOffset !== 'number')
25
+ return false;
26
+ return true;
27
+ }
28
+ /**
29
+ * Append one `compaction` boundary marker to the SessionStore. Returns
30
+ * the SessionEvent we wrote so the caller can echo it into the in-
31
+ * memory transcript without a re-read. Throws on store error so the
32
+ * caller surfaces the failure inline.
33
+ */
34
+ export async function appendCompactBoundary(input) {
35
+ const ts = (input.now ?? (() => Date.now()))();
36
+ const payload = {
37
+ version: 1,
38
+ trigger: input.trigger,
39
+ summary: input.summary,
40
+ summaryTokenCount: input.summaryTokenCount,
41
+ summaryTurnsBefore: input.summaryTurnsBefore,
42
+ keptTailTurns: input.keptTailTurns,
43
+ coversUntilOffset: input.coversUntilOffset,
44
+ };
45
+ const event = {
46
+ t: ts,
47
+ kind: 'compaction',
48
+ payload,
49
+ };
50
+ await input.store.appendEvent(event);
51
+ return event;
52
+ }
53
+ /**
54
+ * Apply replay masking to a chronological event list. Given the full
55
+ * ordered events.jsonl content, return only the events the caller
56
+ * should render: every event AFTER the latest `compaction` boundary,
57
+ * plus the boundary itself (so the renderer can show the banner +
58
+ * summary), and the K kept-tail events that landed BEFORE the boundary
59
+ * but were preserved per the marker's `keptTailTurns`.
60
+ *
61
+ * Mask logic:
62
+ * 1. Walk events. Find the LATEST boundary by offset.
63
+ * 2. Index 0 .. coversUntilOffset-1 are masked, EXCEPT the last
64
+ * `keptTailTurns` of that range (which are the verbatim tail).
65
+ * 3. The boundary event itself + everything after it stays.
66
+ *
67
+ * Why we expose this here (and not in session.ts): keeping the mask
68
+ * logic next to the writer means the wire format is owned by one
69
+ * module. session.ts depends on this; this depends on nothing in
70
+ * session.ts. Unidirectional.
71
+ */
72
+ export function applyCompactMask(events) {
73
+ // Find latest compaction event.
74
+ let latestIdx = -1;
75
+ let latestPayload = null;
76
+ for (let i = events.length - 1; i >= 0; i -= 1) {
77
+ const ev = events[i];
78
+ if (isCompactBoundary(ev)) {
79
+ latestIdx = i;
80
+ latestPayload = ev.payload;
81
+ break;
82
+ }
83
+ }
84
+ if (latestIdx === -1 || latestPayload === null) {
85
+ return events;
86
+ }
87
+ // `coversUntilOffset` is the count of events that existed in the
88
+ // store immediately before the marker append. Events 0 ..
89
+ // coversUntilOffset-1 are summarised; events keptTailTurns of them
90
+ // are surfaced anyway as the verbatim tail.
91
+ const cap = Math.max(0, Math.min(latestPayload.coversUntilOffset, latestIdx));
92
+ const tailKeepCount = Math.max(0, Math.min(latestPayload.keptTailTurns, cap));
93
+ // Take the LAST tailKeepCount events from the masked range, but only
94
+ // those that represent renderable turns (user/persona/system).
95
+ // Boundary markers and tool stream events are NOT counted as turns
96
+ // for the tail-keep window — using them would let the keptTailTurns
97
+ // budget be consumed by infra events and the operator would lose
98
+ // the last K real turns. The spec is about "last K human-visible
99
+ // turns", not "last K events".
100
+ const tailSlice = [];
101
+ for (let i = cap - 1; i >= 0 && tailSlice.length < tailKeepCount; i -= 1) {
102
+ const ev = events[i];
103
+ if (ev.kind === 'user' || ev.kind === 'persona' || ev.kind === 'system') {
104
+ tailSlice.push(ev);
105
+ }
106
+ }
107
+ tailSlice.reverse();
108
+ // After the marker: everything that landed AFTER the boundary
109
+ // append. These are post-compaction events the user has not yet
110
+ // seen folded into a summary; they pass through verbatim.
111
+ const afterMarker = events.slice(latestIdx + 1);
112
+ const markerEvent = events[latestIdx];
113
+ return [...tailSlice, markerEvent, ...afterMarker];
114
+ }
115
+ //# sourceMappingURL=buffer-rewriter.js.map
@@ -0,0 +1,196 @@
1
+ import { estimateTokens } from './token-counter.js';
2
+ /**
3
+ * System prompt for the summarizer. Six closed sections + brand
4
+ * voice clamp + "no tool calls" sentinel. The headings are stable
5
+ * markdown so a downstream renderer can split + reformat without
6
+ * re-parsing the model output.
7
+ *
8
+ * Why the explicit `If a section has no content` rule: empty sections
9
+ * cost zero tokens but their absence is meaningful — the second model
10
+ * reading the memo learns "no decisions were made yet" from the empty
11
+ * `## Decisions` header. Without the rule we observed models silently
12
+ * dropping empty sections and the reader could not distinguish "no
13
+ * decisions" from "summarizer forgot the section".
14
+ */
15
+ const SUMMARIZE_SYSTEM_PROMPT = [
16
+ 'You are the Pugi conversation summarizer. Compress the supplied',
17
+ 'transcript into a six-section memo. Operator picks the work back up',
18
+ 'from this memo — accuracy and completeness matter more than brevity.',
19
+ '',
20
+ 'OUTPUT FORMAT (verbatim section headings, in this order):',
21
+ '',
22
+ "## Intent",
23
+ '(What the operator is trying to accomplish, in one paragraph.)',
24
+ '',
25
+ "## Decisions",
26
+ '(Bullet list of decisions made + the reasoning. Empty section means',
27
+ 'no decisions yet — render the heading anyway.)',
28
+ '',
29
+ "## Files",
30
+ '(Bullet list of file paths touched, with one-line "why".)',
31
+ '',
32
+ "## Errors",
33
+ '(Bullet list of errors encountered + how each was resolved. Empty',
34
+ 'means no errors — still render the heading.)',
35
+ '',
36
+ "## Tools",
37
+ '(Bullet list of notable tool calls and their outcomes. Group similar',
38
+ 'calls; the goal is to surface meaningful state changes, not log',
39
+ 'every Read.)',
40
+ '',
41
+ "## Next",
42
+ '(One paragraph: the immediate next planned action.)',
43
+ '',
44
+ 'RULES:',
45
+ '- Do not invent state. Only summarise what is in the transcript.',
46
+ '- Preserve file paths verbatim.',
47
+ '- Do not call any tools (you have none).',
48
+ '- Do not address the operator. Write in third person.',
49
+ '- No emoji. No em dashes.',
50
+ ].join('\n');
51
+ /**
52
+ * Convert a slice of session events into the user message body the
53
+ * summarizer ingests. We keep the format simple: one event per line,
54
+ * prefixed with the kind, so the model can see role boundaries. Tool
55
+ * outputs are length-capped at 4 KB each so a single 200 KB grep result
56
+ * does not blow the summarizer's own context budget.
57
+ */
58
+ const TOOL_PAYLOAD_CAP_BYTES = 4096;
59
+ export function renderEventsForSummary(events) {
60
+ const lines = [];
61
+ for (const event of events) {
62
+ const rendered = renderOneEvent(event);
63
+ if (rendered !== null)
64
+ lines.push(rendered);
65
+ }
66
+ return lines.join('\n');
67
+ }
68
+ function renderOneEvent(event) {
69
+ const payload = (event.payload ?? null);
70
+ switch (event.kind) {
71
+ case 'user': {
72
+ const text = stringField(payload, 'brief') ?? stringField(payload, 'text') ?? '';
73
+ if (text.length === 0)
74
+ return null;
75
+ return `[operator] ${text}`;
76
+ }
77
+ case 'persona': {
78
+ const text = stringField(payload, 'text') ?? '';
79
+ if (text.length === 0)
80
+ return null;
81
+ const slug = stringField(payload, 'personaSlug') ?? 'persona';
82
+ return `[${slug}] ${text}`;
83
+ }
84
+ case 'system': {
85
+ const text = stringField(payload, 'text') ?? '';
86
+ if (text.length === 0)
87
+ return null;
88
+ return `[system] ${text}`;
89
+ }
90
+ case 'tool.start': {
91
+ const toolName = stringField(payload, 'toolName') ?? 'unknown';
92
+ const args = stringField(payload, 'args') ?? '';
93
+ return `[tool.start ${toolName}] ${truncate(args, TOOL_PAYLOAD_CAP_BYTES)}`;
94
+ }
95
+ case 'tool.result': {
96
+ const toolName = stringField(payload, 'toolName') ?? 'unknown';
97
+ const result = stringField(payload, 'result') ?? '';
98
+ return `[tool.result ${toolName}] ${truncate(result, TOOL_PAYLOAD_CAP_BYTES)}`;
99
+ }
100
+ case 'agent.spawned': {
101
+ const slug = stringField(payload, 'personaSlug') ?? 'unknown';
102
+ return `[agent.spawned ${slug}]`;
103
+ }
104
+ case 'agent.completed': {
105
+ const slug = stringField(payload, 'personaSlug') ?? 'unknown';
106
+ return `[agent.completed ${slug}]`;
107
+ }
108
+ case 'compaction': {
109
+ // Pre-existing compact marker — its `summary` payload IS the
110
+ // condensed form of older history. We pass it through verbatim so
111
+ // the summarizer treats it as already-summarised state and folds
112
+ // it into the new memo (rather than re-summarising garbage).
113
+ const summary = stringField(payload, 'summary') ?? '';
114
+ if (summary.length === 0)
115
+ return null;
116
+ return `[prior compaction]\n${summary}`;
117
+ }
118
+ default: {
119
+ // Forward-compat: unknown kinds get a structural fingerprint so
120
+ // the summarizer can still represent them.
121
+ const exhaustive = event.kind;
122
+ void exhaustive;
123
+ return null;
124
+ }
125
+ }
126
+ }
127
+ function stringField(payload, key) {
128
+ if (!payload)
129
+ return undefined;
130
+ const value = payload[key];
131
+ return typeof value === 'string' ? value : undefined;
132
+ }
133
+ function truncate(s, capBytes) {
134
+ if (Buffer.byteLength(s, 'utf8') <= capBytes)
135
+ return s;
136
+ // Naive cut on UTF-16 chars — good enough for the summarizer; we
137
+ // append a marker so the model knows content was elided.
138
+ return `${s.slice(0, Math.floor(capBytes / 2))}\n... [truncated]`;
139
+ }
140
+ /**
141
+ * Run one summarisation round. Throws on transport error (the caller
142
+ * surfaces a one-line message to the operator and aborts the compact).
143
+ * Returns the structured result on success.
144
+ */
145
+ export async function summarizeEvents(input) {
146
+ if (input.events.length === 0) {
147
+ throw new SummarizerError('refusing-to-summarize-empty-slice', 'No events to summarize.');
148
+ }
149
+ const userBody = renderEventsForSummary(input.events);
150
+ if (userBody.length === 0) {
151
+ throw new SummarizerError('refusing-to-summarize-empty-slice', 'All events rendered as empty.');
152
+ }
153
+ const messages = [
154
+ { role: 'system', content: SUMMARIZE_SYSTEM_PROMPT },
155
+ { role: 'user', content: userBody },
156
+ ];
157
+ const response = await input.client.send(messages, [], {
158
+ personaSlug: input.personaSlug,
159
+ tag: { tag: 'summarize' },
160
+ maxTokens: 2048,
161
+ temperature: 0.1,
162
+ ...(input.model !== undefined ? { model: input.model } : {}),
163
+ ...(input.signal !== undefined ? { signal: input.signal } : {}),
164
+ });
165
+ if (response.stop === 'error') {
166
+ throw new SummarizerError(response.code, `Summarizer transport failed: ${response.message}`);
167
+ }
168
+ if (response.stop === 'tool_use') {
169
+ // Sanity guard. We pass tools: [] so the model should not be able
170
+ // to invoke any; if Anvil's prompt template ever leaks tool defs
171
+ // we want a hard failure rather than a silent dropped summary.
172
+ throw new SummarizerError('unexpected-tool-call', 'Summarizer returned tool_use despite tools: []. Treating as failure.');
173
+ }
174
+ const summary = response.content.trim();
175
+ if (summary.length === 0) {
176
+ throw new SummarizerError('empty-summary', 'Summarizer returned an empty body.');
177
+ }
178
+ return {
179
+ summary,
180
+ tokensSummarised: estimateTokens(userBody),
181
+ eventsSummarised: input.events.length,
182
+ };
183
+ }
184
+ /**
185
+ * Error class so the caller can branch on `error instanceof
186
+ * SummarizerError` and surface `error.code` to the operator.
187
+ */
188
+ export class SummarizerError extends Error {
189
+ code;
190
+ constructor(code, message) {
191
+ super(message);
192
+ this.name = 'SummarizerError';
193
+ this.code = code;
194
+ }
195
+ }
196
+ //# sourceMappingURL=summarizer.js.map