Mode Selector

Mode-Selector Skill

Status

Heuristic v1 active (issue #291, shipped 2026-04-25). Wired into session-start Phase 7.5 (issue #292, shipped 2026-04-25). Backlog signal source live (issue #293, shipped 2026-04-25): signals.backlog is populated by scripts/lib/backlog-scan.mjs::scanBacklog. Accuracy feedback loop live (issue #294, shipped 2026-04-25): scripts/lib/mode-selector-accuracy.mjs::recordAccuracy writes a mode-selector-accuracy learning after the user confirms/overrides the Phase 7.5 banner. Phase B contract is closed. Phase C (#277) /autopilot Loop Command is the next epic and owns its own PRD.

Purpose

Mode-Selector centralizes the session-mode decision across all consumers: session-start Phase 1.5 banner, /autopilot (Phase C), and any future caller that needs a structured recommendation rather than ad-hoc heuristics inline at the call site. Before this skill existed, mode-picking logic was either implicit (user-typed free text) or embedded directly in session-start with no reuse path.

Phase A (state-md.mjs::parseRecommendations, issue #272) established the recommended-mode frontmatter field written by session-end Phase 3.7a. Phase B is the skill that reads that field (plus future signals) and returns a structured recommendation. The key output is a four-field tuple: {mode, rationale, confidence, alternatives}. mode is the recommended session type. rationale is a ≤120-char human-readable explanation. confidence is a float (0.0–1.0) indicating how strongly the selector commits to the recommendation. alternatives is an ordered list of {mode, confidence} objects representing the next-best choices, enabling callers to offer override options without re-running the selector.

The selector is a pure function: given the same signals object it always returns the same output. No file I/O, no network calls, no global state. This makes it trivially testable and safe to call from any skill without side-effect risk.

Contract

Input: signals object

• recommendedMode (string|null) — Phase A frontmatter field; the recommended-mode key from parseRecommendations()
• topPriorities (number[]|null) — issue numbers from the top-priorities frontmatter field
• carryoverRatio (number|null) — float 0.0–1.0 from Phase A; fraction of issues carried over from previous session
• completionRate (number|null) — float 0.0–1.0 from Phase A; ratio of planned issues completed
• previousRationale (string|null) — the rationale string written by session-end Phase 3.7a
• learnings (object[]|null) — RESERVED; not consumed in scaffold; Phase B-1 heuristic input
• recentSessions (object[]|null) — RESERVED; not consumed in scaffold; recent-sessions trend input
• backlog (object|null) — {criticalCount, highCount, staleCount, byLabel, total, vcs, limit} from scripts/lib/backlog-scan.mjs::scanBacklog (Phase B-3, #293). null when CLI missing or no git origin — contributes 0 delta.
• bootstrapLock (object|null) — RESERVED; not consumed in scaffold; tier-aware sizing hints

Output: Recommendation object

Invocation Points

Current

• skills/session-start/SKILL.md Phase 7.5 — first wired invocation (issue #292). Renders 📊 Mode-Selector suggests: when confidence < 0.5 (informational, no pre-selection) or 📊 Mode-Selector recommends: when confidence >= 0.5 (pre-selects AUQ option 1). Eight graceful no-op conditions documented inline. Note: Phase 1.5 📋 banner is NOT a Mode-Selector invocation — it reads Phase A STATE.md frontmatter directly via parseRecommendations; the Mode-Selector lives at Phase 7.5.
• tests/lib/mode-selector.test.mjs — 75 tests (7 describe blocks) exercising SPIRAL, CARRYOVER, high-confidence path, conflicting-signals, stale-signals, alternatives generation, and defensive parsing. mode-selector.mjs coverage 100%/100%/100%/100%. Issue #291.

Future

• /autopilot (Phase C, #277) — auto-execute when confidence >= 0.85 AND SPIRAL/FAILED/carryover-50% kill-switches pass. No user prompt in that path.

Companion modules (Phase B closure)

• scripts/lib/backlog-scan.mjs::scanBacklog — feeds signals.backlog. Phase B-3 (#293). Module-level cache, glab/gh auto-detection, returns null on graceful-degradation paths.
• scripts/lib/mode-selector-accuracy.mjs::recordAccuracy — post-AUQ feedback writer. Phase B-4 (#294). Subject pattern <recommended>-selected-vs-<chosen>; agreement and override land at distinct subjects so the existing learning lifecycle can confirm/contradict them independently.

Scaffold Heuristic (v0)

The v0 scaffold implements a minimal three-branch passthrough. It is intentionally thin so the contract is exercisable by tests before the full Phase B-1 rule-set lands.

selectMode(signals):
  if signals is null/undefined:
    → {mode: 'feature', rationale: 'scaffold: null signals → default', confidence: 0.0, alternatives: []}
  if signals.recommendedMode is valid mode:
    → {mode: <recommendedMode>, rationale: 'scaffold: passthrough of Phase A recommended-mode', confidence: 0.5, alternatives: []}
  otherwise:
    → {mode: 'feature', rationale: 'scaffold: missing/invalid recommendedMode → default', confidence: 0.0, alternatives: []}

Note: the full Phase B heuristic — rule-set consuming learnings.jsonl, recent sessions trend, VCS backlog priority-weighting, and bootstrap.lock tier — is the Phase B-1 follow-up sub-issue.

Fallback Behavior

• confidence = 0.0 means the selector is declining to choose; caller should fall back to its own logic (v0 heuristic) or prompt the user without pre-selecting any option.
• 0.0 < confidence < 0.5 means low-confidence; caller should present as a suggestion, never auto-execute; AUQ should show the recommended mode without marking it as "Recommended".
• confidence ≥ 0.5 means accept as default; present as the pre-selected AUQ option; user can still override.
• confidence ≥ 0.85 (future Phase C threshold) means suitable for autonomous execution in /autopilot mode without user prompt, subject to kill-switch guards.

Integration with Other Skills

• state-md.mjs::parseRecommendations → read Phase A frontmatter fields; consumed via signals.recommendedMode, signals.carryoverRatio, signals.completionRate, etc.
• recommendations-v0.mjs::isValidMode → mode enum validation; import and use — do not redefine the six-value enum inline.
• learnings.mjs::readLearnings → reserved for Phase B-1 heuristic input via signals.learnings
• session-schema.mjs::normalizeSession → reserved for recent-sessions trend input via signals.recentSessions
• bootstrap-lock-freshness.mjs::parseBootstrapLock → reserved for tier-aware sizing hints via signals.bootstrapLock
• gitlab-ops.md → reserved for VCS backlog scan (Phase B-3) via signals.backlog

Critical Rules

• Pure function only. No I/O, no side effects, no throws, no dynamic imports. selectMode must be synchronous and referentially transparent.
• Never write STATE.md. session-start or Phase C writes any derived state; the selector is read-only. session-end Phase 3.7a is the sole writer of recommended-mode.
• Every return path returns all 4 keys. The {mode, rationale, confidence, alternatives} shape is enforced by tests; missing keys are a contract violation.
• alternatives is always an array. Never null, never undefined, may be empty ([]).
• Use isValidMode from recommendations-v0.mjs. Do not redefine the mode enum; drift between selector and validator is a schema bug.

Anti-Patterns

• Do not call selectMode from inside session-end. session-end Phase 3.7a is the SOLE producer of the recommended-mode frontmatter field; the selector is a consumer only.
• Do not add logging inside selectMode — the function must stay pure. Logging (breadcrumbs, sweep.log events) happens at the call site, not inside the selector.
• Do not expand the scaffold to consume learnings inside this session. That is Phase B-1 follow-up work; the RESERVED fields in signals are intentionally ignored here.
• Do not treat confidence as binary. Threshold semantics (0.5 accept-as-default, 0.85 auto-execute) live at the call site, not in the selector. The selector emits a float; the caller decides what to do with it.

References

• Implementation: scripts/lib/mode-selector.mjs
• Tests: tests/lib/mode-selector.test.mjs (selector core, 75 tests), tests/lib/backlog-scan.test.mjs (signal source), tests/lib/mode-selector-accuracy.test.mjs (feedback loop)
• Backlog signal source: scripts/lib/backlog-scan.mjs (Phase B-3, #293)
• Accuracy feedback writer: scripts/lib/mode-selector-accuracy.mjs (Phase B-4, #294)
• PRD: docs/prd/2026-04-25-mode-selector.md
• Epic: #271 v3.2 Autopilot
• Issue: #276 Phase B Mode-Selector
• Phase A Contract PRD: docs/prd/2026-04-24-state-md-recommendations-contract.md
• Phase A parser: scripts/lib/state-md.mjs::parseRecommendations (issue #272)
• Phase A writer: skills/session-end/SKILL.md Phase 3.7a (issue #273)
• Mode enum: scripts/lib/recommendations-v0.mjs::isValidMode

Context-Pressure Signal (#332)

Context-pressure is a single 0.0–1.0 score combining scope size, cross-cutting keyword presence, and recent-session carryover-ratio. It modulates selectMode() decisions without overriding them:

Components

• Scope: (priorityCount - 3) / 10, clamped to [0, 0.5] — 0 issues = 0, 3 = 0, 8 = 0.5, 13+ = 0.5
• Cross-cutting keywords: +0.25 fixed bonus if task description matches /across all|every (skill|agent|repo)|repo-?wide|cross-cutting|rename across|massive refactor/i
• Carryover: mean of effectiveness.carryover / effectiveness.planned_issues over the last 5 sessions, then ratio - 0.3 clamped to [0, 0.25]

Examples

• 2 issues, no cross-cutting keywords, low carryover → score ≈ 0.0 (low) → feature +0.05
• 8 issues, "rename across all skills", carryover 0.4 → score ≈ 0.5 + 0.25 + 0.1 = 0.85 (high) → feature −0.15, housekeeping −0.10
• 5 issues, no keywords, carryover 0.2 → score ≈ 0.2 (low) → feature +0.05

Output

selectMode() now returns context_pressure: { score, components: { scope, keywords, carryover }, level } on all return paths. The rationale string is annotated with ; pressure:<level>(<score>) when level is medium or high.

API

computeContextPressure(signals) is exported as a standalone pure function for testing and future callers (e.g. session-start Phase 7.5 AUQ rendering — deferred to W3-C5).

Constraints

Context-pressure is NOT an Express Path gate (Express Path is structural: housekeeping + ≤3 issues). It is a complementary heuristic that surfaces in the session-start AUQ alongside the recommended mode. Missing signal fields contribute 0 to the score — no NaN propagation.

Open Questions (for Phase B-1 follow-up)

• Learnings freshness window — default 30d? Configurable per-type or a single global TTL?
• Backlog priority weighting — rule-based (priority:critical = +0.2 confidence bonus) vs. learned from historical completion rates?
• Alternative-generation algorithm — top-N non-selected modes scored by partial signal match, or fixed set derived from v0 heuristic branches?
• Confidence computation — additive bonuses per signal (each signal adds a fixed delta) or multiplicative penalties (each contradicting signal scales down a base score)?
• Per-session-type thresholds — should housekeeping require higher confidence for auto-execution than deep given the asymmetry in effort and risk?