loopforge 3.2.1 → 3.7.0

package/dist/prompt-assembler.js

@@ -5,11 +5,10 @@
  * are never truncated; budgets are soft and overflow is recorded.
  */
 import { createHash } from "node:crypto";
-import { hashCanonicalState } from "./canonical-state.js";
+import { buildRoadmap, hashCanonicalState, milestoneHeading, trustBarLine } from "./canonical-state.js";
 import { deriveItemId, STABLE_ID_RE, jaccardSimilarity } from "./token-utils.js";
 import { getPolicy } from "./policy.js";
 export const PROMPT_ARTIFACT_SCHEMA_VERSION = 1;
-export const BASE_PROMPT_VERSION = "2.0.0";
 export const DEFAULT_PROMPT_BUDGETS = {
     l0: 3000,
     l1: 7000,
@@ -39,19 +38,40 @@ function bulletsWithIds(values, prefix) {
         return `- [\`${id}\`] ${trimmed}`;
     }).join("\n");
 }
+/** v3.3.1: Render an identity-HETEROGENEOUS list — the merged active
+ *  constraint set holds hard/plan/criteria/discovered texts together, so no
+ *  single prefix fits. Each item's authoritative ID comes from constraint
+ *  metadata (criteria texts carry cr-XXXXXXXX, everything else
+ *  c-XXXXXXXX), keeping the active list in the same identity namespace as
+ *  the item's own section — one text never renders under two prefixes in
+ *  one prompt. Falls back to a c- derivation only when metadata is missing
+ *  (defensive: metadata covers every active text by construction). */
+function bulletsWithIdentities(values, metadata) {
+    const idEnabled = getPolicy().evolution.constraint_id_enabled;
+    if (!idEnabled)
+        return bullets(values);
+    const idByText = new Map(metadata.map((meta) => [meta.text, meta.id]));
+    return values.map((v) => {
+        const trimmed = v.trim();
+        if (STABLE_ID_RE.test(trimmed)) {
+            return `- [\`${trimmed}\`] ${trimmed}`;
+        }
+        const id = idByText.get(trimmed) ?? `c-${deriveItemId(trimmed)}`;
+        return `- [\`${id}\`] ${trimmed}`;
+    }).join("\n");
+}
 /** v3.2: Short imperative fix/action per verification check, appended to the
  *  L1/L2 Verification Gate lines. Deliberately shorter than the rejection-
  *  prompt wording in enforcement-gate.ts; keys are the plain check-name
  *  strings (NOT imported from verification-gate.ts — it already imports
  *  loop-compiler, which imports this module, so importing back would cycle).
- *  Full coverage of the 19 CHECK_* constants is enforced by a test. */
+ *  Full coverage of every CHECK_* constant is enforced by a test that
+ *  iterates the verification-gate module's CHECK_ exports (no hardcoded
+ *  list — a new constant can never silently lack an action entry). */
 const VERIFICATION_ACTIONS = {
-    progress_regression: "Correct progress_estimate or explain the change in output_summary.",
-    empty_change_with_passing: "Report the actual files_changed, or set success=false with the real state.",
     success_with_remaining_criteria: "Complete the remaining criteria before claiming success, or set success=false and report what remains.",
-    success_without_verified_evidence: "Back the success claim with machine-verifiable evidence: run tests or a required verification command and report the actual output.",
+    success_without_verified_evidence: "Back the success claim with machine-verifiable evidence: run tests or a required verification command and report the actual output. If you genuinely performed no work, provide execution_evidence and declare no_change_reason.",
     outcome_success_contradiction: "Reconcile the declared outcome with the success flag and resubmit an honest evaluation.",
-    success_claim_conflict: "Align the success claim with your own evaluation data before resubmitting.",
     blocked_without_blocker: "State the concrete blocker, or drop the blocked outcome.",
     retroactive_claim_bad_round: "Correct the referenced round number, or withdraw the retroactive claim.",
     retroactive_claim_unverified: "Provide evidence for the retroactive claim, or withdraw it.",
@@ -65,6 +85,17 @@ const VERIFICATION_ACTIONS = {
     subgoal_drift: "Reconcile the sub-goal list with your reported work before resubmitting.",
     backtrack_workspace_not_restored: "Restore the skipped-round files and the expected git HEAD, then resubmit.",
     criteria_claims_unverified: "Provide evidence for each claimed success criterion.",
+    // v3.3 — Round Contract checks (and the three v3.3 checks that predated
+    // this table entry and fell back to the generic action).
+    verification_entrypoint_modified: "Re-run the verification command on the changed entrypoint, or revert the entrypoint change and re-run.",
+    test_files_modified: "Re-run the verification command after the test-file changes and report the fresh output.",
+    round_underspecified: "Declare this round's contract with a non-empty done_when list — what must be true when the round is done?",
+    round_unverifiable: "Name only configured, enabled evidence.commands in verification_plan, or drop done_when items that cannot be machine-verified.",
+    round_scope_drift: "Revert the out-of-scope file changes, or extend the contract scope to cover them and explain why in drift_clarification.",
+    premature_boundary: "Run the contract's verification_plan commands and report real output before claiming done_when met; or set success=false and list the items in success_criteria_remaining (no_change_reason does not apply to contract claims).",
+    // v3.5 — contract completion machine-backing + premature-replacement warn
+    contract_completion_unverified: "Completion under the ACTIVE Round Contract requires its verification_plan commands to pass this round — fix the underlying failure so they pass and resubmit. Completion claims are not accepted without machine verification (no_change_reason does not apply).",
+    contract_premature: "The ACTIVE contract is still open — restate it unchanged to continue it; a different contract is ignored until the active one is completed or blocked.",
 };
 /** Fallback for unmapped check names (forward-compat with future checks). */
 const VERIFICATION_ACTION_FALLBACK = "Re-examine the flagged claim and correct your evaluation before resubmitting.";
@@ -250,7 +281,7 @@ function l1Sections(state, baseline, emphasized) {
         if (renderFull) {
             sections.push({
                 id: "active_constraints",
-                text: section("Active Constraints / Success Criteria", bulletsWithIds(active, "c")),
+                text: section("Active Constraints / Success Criteria", bulletsWithIdentities(active, state.constraintMetadata)),
                 mandatory: false,
             });
         }
@@ -366,6 +397,15 @@ function l1Sections(state, baseline, emphasized) {
             mandatory: false,
         });
     }
+    // v3.3: Forward-looking roadmap (same buildRoadmap as state.md and L2).
+    const roadmapLines = buildRoadmap(state);
+    if (roadmapLines.length > 0) {
+        sections.push({
+            id: "roadmap",
+            text: section("Roadmap", roadmapLines.join("\n")),
+            mandatory: false,
+        });
+    }
     // v3.2: Recurring Issues carries the deterministic violation lessons (with
     // counts and rounds) when available; falls back to the rolling window text.
     if (state.recurringIssues.length > 0 || state.lessons.length > 0) {
@@ -416,12 +456,19 @@ function l2Sections(state, fullStateMarkdown) {
         ["failed_patterns", "Failed Patterns", state.failedPatterns, null],
     ];
     for (const [id, title, values, idPrefix] of groups) {
-        if (values.length > 0) {
-            const body = idPrefix
+        if (values.length === 0)
+            continue;
+        // v3.3.1: the active set is identity-heterogeneous (hard/plan/criteria/
+        // discovered merged) — render through constraint metadata so criteria
+        // texts keep their cr- identity here too (same namespace as their own
+        // Success Criteria section). Every other group is homogeneous and keeps
+        // its single-prefix renderer.
+        const body = id === "active_constraints"
+            ? bulletsWithIdentities(values, state.constraintMetadata)
+            : idPrefix
                 ? bulletsWithIds(values, idPrefix)
                 : bullets(values);
-            sections.push({ id, text: section(title, body), mandatory: false });
-        }
+        sections.push({ id, text: section(title, body), mandatory: false });
     }
     // v3.2: Lessons learned — repeated violations / verification failures
     // across the whole loop (not just the rolling window).
@@ -441,14 +488,22 @@ function l2Sections(state, fullStateMarkdown) {
             mandatory: false,
         });
     }
+    // v3.3: Forward-looking roadmap — rendered right before the (backward-
+    // looking) Phase History so the agent sees where it is before where it's
+    // been. Same buildRoadmap as state.md and L1.
+    const roadmapLines = buildRoadmap(state);
+    if (roadmapLines.length > 0) {
+        sections.push({
+            id: "roadmap",
+            text: section("Roadmap", roadmapLines.join("\n")),
+            mandatory: false,
+        });
+    }
     if (state.milestones.length > 0) {
         const milestoneText = state.milestones
             .map((m) => {
-            const kindIcon = m.kind === "agent_declared"
-                ? "🏁"
-                : m.kind === "criteria_milestone" ? "✅" : "📍";
             const lines = [
-                `**${kindIcon} ${m.label}** (Rounds ${m.round_range.start}–${m.round_range.end}, ${(m.progress_at_boundary * 100).toFixed(0)}%)`,
+                milestoneHeading(m),
                 `  Outcome: ${m.outcome}`,
             ];
             if (m.resolved_constraints.length > 0) {
@@ -507,13 +562,6 @@ function l2Sections(state, fullStateMarkdown) {
             mandatory: false,
         });
     }
-    if (state.loopSynthesis) {
-        sections.push({
-            id: "loop_synthesis",
-            text: section("Loop Summary", state.loopSynthesis),
-            mandatory: false,
-        });
-    }
     if (state.nextAction) {
         sections.push({
             id: "next_action",
@@ -534,7 +582,8 @@ function l2Sections(state, fullStateMarkdown) {
         state.progress.criteriaMet.length > 0 ||
         state.progress.criteriaRemaining.length > 0 ||
         state.progress.filesChanged.length > 0 ||
-        state.progress.tests !== null) {
+        state.progress.tests !== null ||
+        state.machineStatus !== undefined) {
         const lines = [];
         const total = state.progress.criteriaMet.length + state.progress.criteriaRemaining.length;
         if (total > 0) {
@@ -542,6 +591,22 @@ function l2Sections(state, fullStateMarkdown) {
         }
         if (state.progress.estimate !== null) {
             lines.push(`**Estimated Completion**: ${(state.progress.estimate * 100).toFixed(0)}%`);
+            // v3.3: honest labeling — the estimate is the agent's own number.
+            lines.push("**Signal source**: self-reported estimate (unverified until machine-backed)");
+        }
+        // v3.3: machine side of the comparison — git motion over committed
+        // rounds (v3.6: the object always carries definite values — no
+        // "unavailable" arm).
+        if (state.machineStatus) {
+            const ms = state.machineStatus;
+            const motion = ms.gitMotion
+                ? `changes in ${ms.motionRounds}/${ms.windowRounds} recent committed rounds`
+                : `no git changes in the last ${ms.windowRounds} committed rounds`;
+            lines.push(`**Machine (git)**: ${motion}`);
+        }
+        if (state.criterionStatuses.length > 0) {
+            const metCount = state.criterionStatuses.filter((cs) => cs.status === "met").length;
+            lines.push(`**Machine (criteria)**: ${metCount}/${state.criterionStatuses.length} met across committed rounds`);
         }
         if (state.progress.tests) {
             lines.push(`**Tests**: ${state.progress.tests.passed} passed, ` +
@@ -577,6 +642,32 @@ function l2Sections(state, fullStateMarkdown) {
             mandatory: false,
         });
     }
+    // v3.3: Per-round statistics — lets the agent calibrate round granularity
+    // (files per round, rejected attempts, self-reported progress deltas).
+    // v3.6: source-labeled — files/Δ are agent-reported; only rejected
+    // attempts (and the Machine (git) row above) are machine-recorded.
+    if (state.roundStats && state.roundStats.length > 0) {
+        const statLines = state.roundStats.map((stat) => {
+            const parts = [];
+            if (stat.filesChangedCount !== null) {
+                parts.push(`${stat.filesChangedCount} file${stat.filesChangedCount === 1 ? "" : "s"}`);
+            }
+            if (stat.rejectedAttempts !== null && stat.rejectedAttempts > 0) {
+                parts.push(`${stat.rejectedAttempts} rejected attempt${stat.rejectedAttempts === 1 ? "" : "s"}`);
+            }
+            if (stat.progressDelta !== null) {
+                parts.push(`Δ${stat.progressDelta >= 0 ? "+" : ""}${stat.progressDelta.toFixed(2)}`);
+            }
+            return `- R${stat.round}: ${parts.length > 0 ? parts.join(", ") : "no data"}`;
+        });
+        statLines.push("> files & Δ are agent-reported; rejected attempts are machine-recorded " +
+            "(the Machine (git) row above is machine-observed)");
+        sections.push({
+            id: "round_stats",
+            text: section("Round Stats", statLines.join("\n")),
+            mandatory: false,
+        });
+    }
     // Retired Constraints
     if (state.retiredConstraints.length > 0) {
         const lines = state.retiredConstraints.map((c) => `- ~${c}~ (retired)`);
@@ -586,35 +677,10 @@ function l2Sections(state, fullStateMarkdown) {
             mandatory: false,
         });
     }
-    // Inactive Constraints (with decay metadata)
-    if (state.inactiveConstraints.length > 0) {
-        const lines = [
-            "> Demoted after prolonged inactivity. Auto-reactivated if violated again.",
-            "",
-        ];
-        for (const c of state.inactiveConstraints) {
-            const meta = state.constraintMetadata.find((m) => m.text === c);
-            const age = meta
-                ? state.round - (meta.last_violated_at_round || meta.discovered_at_round || 0)
-                : "?";
-            const lastV = meta?.last_violated_at_round
-                ? `last violated R${meta.last_violated_at_round}`
-                : "never violated";
-            lines.push(`- ${c} (${lastV}, ${age} rounds inactive)`);
-        }
-        sections.push({
-            id: "inactive_constraints",
-            text: section("Inactive Constraints", lines.join("\n")),
-            mandatory: false,
-        });
-    }
     // Agent Trust Score + Trend
     if (state.agentTrustScore !== undefined) {
-        const barLen = 10;
-        const filled = Math.round(state.agentTrustScore * barLen);
-        const bar = "█".repeat(filled) + "░".repeat(barLen - filled);
         const lines = [
-            `${bar} ${(state.agentTrustScore * 100).toFixed(0)}%`,
+            trustBarLine(state.agentTrustScore),
         ];
         if (state.agentTrustTrend.length > 0) {
             const avg = (state.agentTrustTrend.reduce((a, b) => a + b, 0)
@@ -631,9 +697,6 @@ function l2Sections(state, fullStateMarkdown) {
     return sections;
 }
 function selectSections(input, emphasized) {
-    if (input.mode === "full") {
-        return l2Sections(input.state, input.fullStateMarkdown);
-    }
     if (input.level === "l0")
         return l0Sections(input.state);
     if (input.level === "l1") {
@@ -693,7 +756,7 @@ function matchEmphasize(targets, candidates, cap) {
 }
 /** Render the "Confusion Alerts" section that appears at the prompt top.
  *  Compiler auto-matches each confusion point against state sections. */
-function renderConfusionAlerts(points, state, expandAll) {
+function renderConfusionAlerts(points, state, fullDensity) {
     if (points.length === 0)
         return "";
     const lines = [
@@ -712,11 +775,10 @@ function renderConfusionAlerts(points, state, expandAll) {
         { keyword: "criteria", label: "Success Criteria", ref: "## Success Criteria" },
         { keyword: "phase", label: "Phase History", ref: "## Phase History" },
         { keyword: "objective", label: "Loop Objective", ref: "## Loop Objective" },
-        { keyword: "loop", label: "Loop Summary", ref: "## Loop Summary" },
     ];
     // v2.14: L2 renders up to the policy cap (was: every confusion point,
     // unbounded by max_confusion_points); L1 keeps a single alert.
-    const shown = expandAll
+    const shown = fullDensity
         ? points.slice(0, getPolicy().prompt.max_confusion_points)
         : points.slice(0, 1);
     for (const point of shown) {
@@ -738,7 +800,7 @@ function renderConfusionAlerts(points, state, expandAll) {
         lines.push(`  ${pointer}`);
         lines.push("");
     }
-    if (!expandAll && points.length > 1) {
+    if (!fullDensity && points.length > 1) {
         lines.push(`*… and ${points.length - 1} more confusion point(s) — see state file.*`);
         lines.push("");
     }
@@ -761,121 +823,15 @@ function renderCriticalContext(emphasized, state) {
     lines.push("");
     return lines.join("\n");
 }
-/** Expand a specific section to L2-level detail within a L1 prompt.
- *  Returns the expanded section text, or empty string if the section
- *  has no content to expand. */
-function renderExpandedSection(sectionName, state) {
-    switch (sectionName) {
-        case "milestones": {
-            if (state.milestones.length === 0)
-                return "";
-            const text = state.milestones
-                .map((m) => {
-                const icon = m.kind === "agent_declared" ? "🏁" : m.kind === "criteria_milestone" ? "✅" : "📍";
-                const lines = [
-                    `**${icon} ${m.label}** (R${m.round_range.start}–R${m.round_range.end}, ${(m.progress_at_boundary * 100).toFixed(0)}%)`,
-                    `  Outcome: ${m.outcome}`,
-                ];
-                if (m.carried_constraints.length > 0)
-                    lines.push(`  Carried: ${m.carried_constraints.slice(0, 3).join("; ")}`);
-                if (m.resolved_constraints.length > 0)
-                    lines.push(`  Resolved: ${m.resolved_constraints.slice(0, 3).join("; ")}`);
-                return lines.join("\n");
-            })
-                .join("\n\n");
-            return section("Expanded Phase History", text);
-        }
-        case "sub_goals": {
-            if (state.subGoals.length === 0)
-                return "";
-            const lines = [];
-            for (const sg of state.subGoals) {
-                const icon = sg.status === "in_progress" ? "🔄" :
-                    sg.status === "done" ? "✅" :
-                        sg.status === "blocked" ? "🚫" :
-                            sg.status === "canceled" ? "❌" : "⏳";
-                const stale = sg.status === "pending" && state.round - sg.declared_at_round >= 10 ? " ⚠️ stale" : "";
-                const detail = sg.status === "done" && sg.completed_at_round
-                    ? ` (done, R${sg.completed_at_round})`
-                    : sg.status === "in_progress" ? ` (since R${sg.status_changed_at_round})`
-                        : sg.status === "pending" ? ` (since R${sg.declared_at_round})`
-                            : sg.status === "blocked" ? ` (blocked R${sg.status_changed_at_round})`
-                                : ` (canceled R${sg.status_changed_at_round})`;
-                lines.push(`- ${icon} [\`${sg.id}\`] ${sg.description}${detail}${stale}`);
-            }
-            return section("Expanded Sub-Goal Dashboard", lines.join("\n"));
-        }
-        case "constraint_lifecycle": {
-            if (state.inactiveConstraints.length === 0 && state.retiredConstraints.length === 0)
-                return "";
-            const lines = [];
-            if (state.inactiveConstraints.length > 0) {
-                lines.push("**Inactive (auto-decayed):**");
-                for (const c of state.inactiveConstraints) {
-                    const meta = state.constraintMetadata.find((m) => m.text === c);
-                    const age = meta ? state.round - (meta.last_violated_at_round || meta.discovered_at_round || 0) : "?";
-                    lines.push(`- ${c} (${age} rounds inactive)`);
-                }
-            }
-            if (state.retiredConstraints.length > 0) {
-                lines.push("**Retired:**");
-                for (const c of state.retiredConstraints)
-                    lines.push(`- ~${c}~`);
-            }
-            return section("Constraint Lifecycle", lines.join("\n"));
-        }
-        case "agent_trust": {
-            if (state.agentTrustScore === undefined)
-                return "";
-            const barLen = 10;
-            const filled = Math.round(state.agentTrustScore * barLen);
-            const bar = "█".repeat(filled) + "░".repeat(barLen - filled);
-            const lines = [
-                `${bar} ${(state.agentTrustScore * 100).toFixed(0)}%`,
-            ];
-            if (state.agentTrustTrend.length > 0) {
-                lines.push(`Trend: ${state.agentTrustTrend.join(" → ")}`);
-            }
-            return section("Agent Trust (Expanded)", lines.join("\n"));
-        }
-        case "progress": {
-            const p = state.progress;
-            if (p.estimate === null && p.criteriaMet.length === 0 && p.criteriaRemaining.length === 0)
-                return "";
-            const lines = [];
-            const total = p.criteriaMet.length + p.criteriaRemaining.length;
-            if (total > 0)
-                lines.push(`**Criteria**: ${p.criteriaMet.length}/${total} met`);
-            if (p.estimate !== null)
-                lines.push(`**Estimate**: ${(p.estimate * 100).toFixed(0)}%`);
-            if (p.tests)
-                lines.push(`**Tests**: ${p.tests.passed} passed, ${p.tests.failed} failed, ${p.tests.skipped} skipped`);
-            if (p.filesChanged.length > 0) {
-                lines.push("**Files Changed**:");
-                for (const f of p.filesChanged.slice(0, 10))
-                    lines.push(`- ${f}`);
-            }
-            return section("Progress Dashboard (Expanded)", lines.join("\n"));
-        }
-        case "loop_synthesis": {
-            if (!state.loopSynthesis)
-                return "";
-            return section("Loop Summary", state.loopSynthesis);
-        }
-        default:
-            return "";
-    }
-}
 export function assemblePromptArtifact(input) {
-    const mode = input.mode ?? "adaptive";
     const budgets = { ...DEFAULT_PROMPT_BUDGETS, ...input.budgets };
     const budget = budgets[input.level];
     const stateHash = hashCanonicalState(input.state);
     const attempt = Math.max(1, input.attempt ?? 1);
     // ── State file reference + conditional read instruction ────────────
-    // L2 rounds (full rehydration), recovery, rejection, drift, and pointer
-    // mode all benefit from the model reading the durable state file rather
-    // than relying on potentially-compacted conversation history.
+    // L2 rounds (full rehydration), recovery, rejection, and drift all
+    // benefit from the model reading the durable state file rather than
+    // relying on potentially-compacted conversation history.
     const fullContextReasons = new Set([
         "first_round", "plan_boundary", "checkpoint_boundary", "goal_changed",
         "missing_previous_state", "verification_contradicted",
@@ -883,7 +839,6 @@ export function assemblePromptArtifact(input) {
         "periodic_refresh",
     ]);
     const needsFullContext = input.level === "l2"
-        || input.mode === "pointer"
         || input.reasons.some((r) => fullContextReasons.has(r));
     const stateFilePath = input.state.stateFilePath;
     // v3.2: L1 annotates the pointer with the round the freshly-regenerated
@@ -1005,11 +960,6 @@ export function assemblePromptArtifact(input) {
             milestoneRanges: input.state.milestones.map((m) => [m.round_range.start, m.round_range.end]),
         };
     }
-    // Expand: L1 only, max 1 section (L2 already expanded, L0 ignored)
-    const expandCap = levelIsL1 ? policy.max_expand_l1 : 0;
-    const expandedSection = (pr?.expand?.length && expandCap > 0)
-        ? renderExpandedSection(pr.expand[0], input.state)
-        : "";
     // Inject confusion alerts into the header area (before mandatory sections)
     const fixedText = header + confusionText;
     // Inject critical context as a mandatory section
@@ -1024,34 +974,17 @@ export function assemblePromptArtifact(input) {
             ...sections,
         ];
     }
-    const selected = mode === "pointer"
-        ? renderWithinBudget(augmentedSections.filter((item) => item.mandatory), fixedText, budget)
-        : renderWithinBudget(augmentedSections, fixedText, budget - footer.length - expandedSection.length);
-    const renderedPrompt = selected.rendered + expandedSection + footer;
-    const includedSections = [
-        ...selected.included,
-        ...(confusionText ? ["confusion_alerts"] : []),
-        ...(criticalContextText ? ["critical_context"] : []),
-        ...(expandedSection ? [`expand:${pr?.expand?.[0] ?? ""}`] : []),
-        ...(pointer ? ["state_pointer"] : []),
-        ...(evaluation ? ["self_evaluation"] : []),
-    ];
+    const selected = renderWithinBudget(augmentedSections, fixedText, budget - footer.length);
+    const renderedPrompt = selected.rendered + footer;
     const promptHash = createHash("sha256").update(renderedPrompt).digest("hex");
     return {
         schemaVersion: PROMPT_ARTIFACT_SCHEMA_VERSION,
         roundId: `loop:${input.state.loopId}:round:${input.state.round}`,
         attempt,
         level: input.level,
-        levelReasons: [...input.reasons],
         renderedPrompt,
         promptHash,
         stateHash,
-        basePromptVersion: BASE_PROMPT_VERSION,
-        includedSections,
-        budgetChars: budget,
-        charCount: renderedPrompt.length,
-        budgetExceeded: renderedPrompt.length > budget,
-        generatedAt: Date.now(),
         ...(presentedState ? { presentedState } : {}),
     };
 }