feat: RFC for token-efficient monitors

[kyleseaman]

Stacked change: PR 1 of 8

Stack: #5180 → #5181 → #5182 → #5183 → #5184 → #5185 → #5186 → #5305
Base: main
Next: #5181

Problem / Motivation

The existing babysit flow repeatedly wakes the model to poll remote state, owns completion in prompt text, and has inconsistent budgets and behavior across dashboard and messaging surfaces.

Why it matters

Unchanged polling consumes turns and tokens without advancing work. Voluntary model-side stopping also makes completion, auditability, and cost controls unreliable.

What changed (motivation → approach → change)

• Define a session-owned structured monitor that probes outside the model and wakes the session only for actionable evidence.
• Specify durable observations, fingerprints, terminal outcomes, finite runtime/turn/token/provider-error budgets, and completed-turn accounting.
• Record the stateless MCP/session boundary, security constraints, migration strategy, and the eight-PR delivery sequence.
• Keep legacy AutoNudge behavior isolated behind an explicit compatibility path.

Tests

• bash scripts/docs-lint.sh
• bash scripts/scrub-lint.sh --no-history
• Base-aware brand and stack-boundary checks

Manual verification

N/A — this PR is documentation-only; documentation and policy gates cover the changed surface.

Related Issues

N/A — no linked issue. This RFC is the design authority for #5181 through #5186.

Checklist

• Single commit with a Conventional Commits title
• Existing tests pass and new tests cover the new behavior
• Self-review completed; code follows project style guidelines
• Documentation updated where applicable
• No secrets, credentials, or internal references in the diff

Contribution License Agreement

N/A — the repository template does not yet supply final CLA wording.

[github-actions]

GPT 5.6 Review — ✅ no blocking findings

GPT 5.6 completed its review of f31efef29924750fb7458af76c1f217705ac3871 and found no blocking issues.

This comment is updated in place on each push.

Review details

No findings.
[GPT-REVIEWED] f31efef

False positive or not applicable? A repository writer can comment:
/ai-review override gpt f31efef29924750fb7458af76c1f217705ac3871: <one-sentence reason>

[github-actions]

Opus 4.8 Review — ✅ no blocking findings

Reviewed f31efef29924750fb7458af76c1f217705ac3871 — this comment is updated in place on each push.

Review details

No findings.

[OPUS-REVIEWED] f31efef

Verdict parsed from the review's SHA-scoped output markers for commit f31efef29924750fb7458af76c1f217705ac3871.

False positive or not applicable? A repository writer can comment:
/ai-review override fable f31efef29924750fb7458af76c1f217705ac3871: <one-sentence reason>

[github-actions]

Design Review (Fable 5) — ✅ PASS

Design-level review of f31efef29924750fb7458af76c1f217705ac3871 — updated in place on each push. A BLOCK verdict blocks PR readiness; PASS/CONCERNS are advisory.

Design-Verdict: PASS

A real, evidenced problem (all four observed loops died at cap) solved at its root — completion moves from prompt to runtime — with compat, budgets, and failure modes specified.

[DESIGN-REVIEWED] f31efef