ADR 0007 amendment: §21 trigger entries as Recursive-Reload Anchor (post-pruning-discipline-recall)

Context

PR #11434 added FAIR-band stance declaration per §1.3 + FAIR-band declaration verification to AGENTS.md §21 trigger table per operator-direction. The substantive insight surfaced during PR #11434 review-handoff (operator messages 2026-05-15 ~18:18Z + ~18:28Z): AGENTS.md §21 trigger entries function as recursive-reload anchors for post-pruning-discipline-recall, not just first-time-invocation pointers.

The current ADR 0007 baseline taxonomy (3-axis slot rule: trigger-frequency × failure-severity × enforceability) doesn't explicitly capture this dimension. Phase C compression planners need this annotation to preserve §21 entries as load-bearing under compression cycles.

The Problem

Currently, skill-payload-only discipline (the content loaded via view_file on skill payload at trigger-time) is fragile under context pruning in long sessions:

• AGENTS.md is L1 anchor; reloaded every turn
• Skill payloads loaded only on-invocation via view_file
• After context-pruning, skill content is pruned but AGENTS.md reloads
• Without §21 trigger, the agent cannot recover the pruned discipline until independently rediscovering the need + skill path

This means: any substrate-discipline that depends on skill-payload-only loading is empirically unreliable in long sessions. Even invoked-earlier disciplines (/lead-role's relaxed-planning mindset; /peer-role's convergence-pressure posture; FAIR-band author-lane discipline) silently forget post-compaction.

The Architectural Reality

ADR 0007 (Compaction Taxonomy) §2.1 baseline disposition table currently classifies sections by:

• Disposition (keep / move / compress-to-trigger / rewrite / retire)
• Tag (MACHINE-ENFORCEABLE-CANDIDATE / DISCIPLINE-ONLY)
• Rationale / Friction Capture

What's missing: annotation of which keep rows are recursive-reload-required (load-bearing for post-pruning discipline recall) vs which are merely high-frequency-trigger entries.

§21 trigger table specifically holds the recursive-reload anchors for the entire skill substrate. Phase C compression candidates must NOT touch §21 entries that have recursive-reload-required semantic, even if other axes (frequency / severity / enforceability) would otherwise justify compression.

The Fix

Amend learn/agentos/decisions/0007-agents-md-compaction-taxonomy.md with:

1. New §2.0.1 Post-Pruning-Recurrence-Rate Sub-Axis explaining the dimension:

• Trigger-frequency now decomposes into (a) per-turn-load-frequency (existing) + (b) post-pruning-discipline-recall-recurrence (new)
• Skills that shape ongoing-session-behavior (lead-role, peer-role, FAIR-band, blocked-task-state, others) NEED their §21 trigger as recursive-reload anchor
• Skills that are one-shot-and-done (ticket-create, structural-pre-flight, session-sunset) may be lower-priority on this axis

2. Annotate baseline disposition table with new recursive-reload-required column for keep rows in §21:

• pull-request, pr-review, post-review-pickup, lead-role, peer-role, FAIR-band: YES
• ticket-create, ticket-triage, ticket-intake, structural-pre-flight, session-sunset, unit-test, whitebox-e2e: NO (one-shot lifecycle skills; discipline doesn't persist across turns)

3. New §5.3 Anti-Pattern: Recursive-Reload Anchor Retirement — explicit prohibition on retiring §21 entries marked recursive-reload-required during Phase C compression cycles.

Acceptance Criteria

• ADR 0007 §2.0.1 (or equivalent) codifies post-pruning-recurrence-rate as sub-axis of trigger-frequency
• Baseline disposition table (§2.1) extended with recursive-reload-required annotation for relevant keep rows
• §5 Anti-Patterns adds explicit prohibition on retiring recursive-reload-required entries during compression
• PR body documents which §21 entries are annotated; cross-reference Phase C compression planning surface
• FAIR-band stance declaration per pull-request-workflow §1.3 (per the discipline that surfaced this insight)

Out of Scope

• AGENTS.md §21 content changes (the entries are already correct per recent PRs; this is metadata-amendment to ADR 0007)
• New skill creation
• Phase C compression itself (separate work tracked in archived Discussion #11419)
• Mechanical CI enforcement of the annotation (defer to future ticket if needed)

Avoided Traps / Gold Standards Rejected

• Adding the annotation to AGENTS.md inline — rejected per substrate-accretion-defense; the annotation belongs in the ADR substrate that GOVERNS the §21 entries, not the §21 entries themselves
• New ADR 0009 — rejected; this is a substantive amendment to ADR 0007's framework, not a new primitive. ADR 0007 already establishes the framework being extended.
• Discussion (ideation-sandbox) first — rejected; the architectural insight is surfaced via operator-direct framing 2026-05-15 with empirical anchors (PR #11434 context-pruning case + the lead-role/peer-role substrate-discipline-fragility examples). No high-blast ambiguity needing convergence; direct ticket path is appropriate.

Empirical Anchors

• Operator framing 2026-05-15 ~18:28Z: "think of context pruning in long sessions: even lead or peer role might be 'forgotten' (adding weight for agents.md)"
• PR #11434 — §21 trigger addition surfaced the recursive-reload framing during operator-correction cycle
• PR #11432 — primary FAIR-band codification at post-review-pickup-workflow.md §4; the substrate this ADR amendment governs
• Discussion #11419 — graduation source for ADR 0007 + Phase C compression substrate
• Ticket #11413 — Phase C compression implementation ticket (Phase C planners are the primary consumer of this amendment)
• AGENTS.md §15.6 Helpful Assistant Regression-Defense — empirical anchor for skill-invocation bypass shape that paired with context-pruning makes payload-only disciplines fragile

Related

• ADR 0007 — Compaction Taxonomy (the artifact being amended)
• ADR 0006 — ADRs as Graph-Queryable Entities (this amendment becomes graph-queryable via ADR 0006)
• ADR 0005 §2.3 amended — ADR-amendment lifecycle (Proposed → Accepted via operator merge)
• PR #11434 — FAIR-band PR-Pre-Flight Gate (surfaced the insight; Substrate Mutation Rationale section partially captures this rationale)
• PR #11432 — FAIR-band primary codification (merged 2026-05-15)
• Discussion #11419 — substrate-evolution archaeology (Phase C planning consumer)
• Ticket #11413 — Phase C compression implementation (load-bearing consumer of this amendment)
• feedback_skill_adherence_asymmetry.md family — empirical anchor for skill-payload-only-discipline fragility

Origin Session ID

656c0935-0b3e-4b06-9b14-548524275859

Handoff Retrieval Hint

Memory query: "AGENTS.md §21 recursive-reload anchor post-pruning-discipline-recall" OR "ADR 0007 amendment annotation column". Cross-reference: PR #11434 substrate mutation rationale + operator-direction 2026-05-15 ~18:28Z context-pruning framing.

Substrate Enforcement Sufficiency Audit (per ticket-intake-workflow.md §9.2)

Required for substrate tickets per Hot Context Fast-Path Substrate Exception.

Target failure mode: Phase C compression cycle (ticket #11413 + others) retires §21 entries based on the existing 3-axis disposition table without recognizing that some keep entries are load-bearing recursive-reload anchors for post-pruning-discipline-recall. The compression succeeds on the byte-budget axis but fragilities the swarm's behavioral substrate in long sessions.

Existing enforcement layers:

• Layer 4 (CI): lint-substrate-size.mjs (PR #11415) enforces byte cap; no awareness of which keep entries are recursive-reload-anchors
• Layer 1 (AGENTS.md): §21 trigger table exists but doesn't self-annotate which entries are recursive-reload-required
• Existing skill logic: ADR 0007 Compaction Taxonomy provides the 3-axis framework but doesn't capture the post-pruning-recurrence-rate sub-axis

Insufficiency proven: No existing substrate signals to Phase C compression planners that §21 entries with recursive-reload semantic must be preserved. The current framework would permit a compress-to-trigger disposition for, e.g., lead-role §21 trigger — which would empirically break post-pruning lead-role discipline recall.

Decision: Amend ADR 0007 to capture the post-pruning-recurrence-rate sub-axis + annotate baseline disposition table + add §5 anti-pattern. ROI Positive — preserves substrate-evolution velocity (Phase C compression can proceed) while protecting load-bearing recursive-reload anchors.