friday-session-mode-recommendation

name	friday-session-mode-recommendation
description	Session-mode recommendation policy for FRIDAY. Load when FRIDAY needs to recommend clean session versus continue here after a completed specialist step and an approved meaningful continuation.

FRIDAY Session Mode Recommendation Policy

Purpose

This skill defines when FRIDAY should recommend clean session versus continue here.

When loaded, it is authoritative for:

• Whether a session-mode recommendation should be offered at all.
• The default recommendation bias.
• State-based counters and continuation signals.
• Practical thresholds and hard exceptions.
• The required recommendation output format.

FRIDAY recommends; the user decides.

When To Load

Load this skill only after a specialist step is complete and FRIDAY is evaluating whether the next meaningful work should continue in the same conversation or from saved state.

Do not load it for:

• Startup classification.
• Generic approval requests.
• Active analyst or architect loop questions that have not reached completion.
• Turns where no meaningful continuation is available yet.

Offer Gate

Offer a session-mode recommendation only when at least one of these is true:

• There is a meaningful next specialist step that is already approved or is the exact next approved continuation.
• There is substantial continuation still expected in the same workflow and session choice affects quality.

Do not offer a session-mode recommendation when any of these are true:

• FRIDAY is only waiting for output approval, rejection, or change requests.
• The workflow is complete and no next specialist step is being prepared.
• The user must answer an analyst or architect loop question before any new specialist step can happen.
• FRIDAY is blocked on missing domain input, repair approval, reset approval, or direction change.
• The next action is only a short direct FRIDAY answer with no specialist continuation.

Default Bias

Default to continue here.

Use clean session only when the continuation signals clearly show that the next step would benefit from narrower, fresher context.

Persisted Metrics Source

Use persisted session_metrics from .temper/friday-state.json as the primary source for recommendation decisions.

Expected fields:

• specialist_runs_total
• specialist_runs_current_thread
• distinct_specialists
• distinct_specialist_count
• cycle_rounds_total
• recovery_attempts
• prompt_failure_retries
• long_outputs_count
• large_summaries_count
• last_output_class
• next_action_type
• scope_confusion_signals
• mixed_scope_signal
• risk_score

If session_metrics is missing because the state file predates this policy, default missing counters to 0, default mixed_scope_signal to false, default last_output_class to none, default next_action_type from current checkpoint intent, recompute risk_score, and keep the default bias to continue here.

Practical Thresholds

Normalize next_action_type before evaluating thresholds:

• Continuation types that may justify an offer: approved-next-specialist, approved-exact-continuation.
• Types that suppress an offer: output-approval, analyst-cycle-input, architect-cycle-input, recovery-approval, repair-or-reset, blocked-input, direct-answer, change-direction, none.

Compute risk_score from persisted metrics using this additive model:

• +1 when specialist_runs_current_thread >= 2
• +1 when distinct_specialist_count >= 2
• +1 when cycle_rounds_total >= 3
• +1 when recovery_attempts >= 1
• +1 when prompt_failure_retries >= 1
• +1 when long_outputs_count >= 2
• +1 when large_summaries_count >= 1
• +1 when last_output_class is failure, partial-output, unclear-output, or recovery-report
• +1 when scope_confusion_signals >= 1
• +2 when mixed_scope_signal is true

Persist the computed total back to session_metrics.risk_score when FRIDAY writes state.

Recommend continue here when all of these are true:

• next_action_type is approved-next-specialist or approved-exact-continuation.
• risk_score <= 2.
• specialist_runs_current_thread <= 1.
• long_outputs_count <= 1.
• large_summaries_count = 0.
• recovery_attempts = 0.
• prompt_failure_retries = 0.
• scope_confusion_signals = 0.
• mixed_scope_signal = false.

Recommend clean session when any of these threshold groups is true:

• risk_score >= 4
• specialist_runs_current_thread >= 3
• distinct_specialist_count >= 3
• cycle_rounds_total >= 4
• recovery_attempts + prompt_failure_retries >= 2
• long_outputs_count + large_summaries_count >= 3

Hard Exceptions

Recommend continue here even if some noise exists when any of these are true:

• next_action_type is approved-exact-continuation, risk_score <= 3, and both mixed_scope_signal = false and scope_confusion_signals = 0.
• The next step is immediate, narrow, and tightly coupled to the just-completed specialist output, with specialist_runs_current_thread <= 2 and no recovery history.
• FRIDAY only needs to perform a short approved continuation and the only load signal is long_outputs_count = 1 or large_summaries_count = 1.

Recommend clean session immediately when any of these are true:

• mixed_scope_signal = true.
• scope_confusion_signals >= 2.
• risk_score >= 6.
• next_action_type = approved-next-specialist and the next specialist shifts focus after recovery-heavy or report-heavy context: recovery_attempts >= 1 plus long_outputs_count + large_summaries_count >= 2.

Output Format

Use exactly this structure:

Recommendation: clean session | continue here
Reason: [one concise reason tied to current continuation signals]
Choices: Reply "clean" to continue from saved state with focused context, or "continue here" to proceed in this session.

Keep the reason operational and short. Do not give a long essay.

Decision Rule

If the signals are mixed or borderline, prefer continue here.

Only recommend clean session when the practical thresholds or hard exceptions clearly justify it.

Decision order:

1. Apply the offer gate.
2. Read session_metrics and recompute risk_score if needed.
3. Apply continue here hard exceptions first when continuation is immediate and tightly scoped.
4. Apply clean session hard exceptions next.
5. If no hard exception fired, use the practical thresholds.
6. If still mixed or borderline, recommend continue here.