session-flow: aihero course lane 3: compaction doctrine (fork-beats-compaction, auto-compact stance, evidence-degraded marker)

[kyle-sexton]

Context

Third of six vetting lanes for the AI Hero course lessons; contract in docs/topics/pocock-course-lanes/PLAN.md (branch claude/plan-mode-discussion-55kszx). Merges the Compaction and Auto-Compaction lessons: he treats steered compaction as the default continuation mechanism (and compact-before-QA as its ideal case); our doctrine is compact-as-last-resort (session-flow:handoff "Fork beats compaction when the window is deep") and context-guard marks every post-compaction session evidence-degraded (effective zone dumb regardless of numbers). Consumes lane 2's verdicts and the background harness-claims research (.work/pocock-course-lanes/harness-claims/, dispatched pre-lane).

Proposed work

• Vet his steered-compact-for-QA case against the evidence-degraded marker: is post-compaction-equals-dumb too blunt for a boundary-timed, steered compact validating finished work, or does that QA belong in a subagent anyway (his own decision-tree lesson suggests the latter)
• Vet the auto-compact stance: his "if auto-compact fires, something went wrong; the human owns the boundary" vs our instrumented approach (zones observable, menus operator-only, optional blocking mode)
• Act on the verified harness claims (autoCompactWindow, compaction mechanics, message queueing) per the claim ladder: doc-verified claims only; empirical figures recorded as anchors, never adopted as numbers
• Record the primary/secondary-source framing verdict (feeds lane 6 vocabulary)

Acceptance criteria

• Adopt / reject-with-reason / track-on-event rows recorded in docs/upstream/aihero-course.md for every compaction-lesson and auto-compaction-lesson claim
• Work items filed for each decided plugin change
• Lane summary written to the lane's topic slice
• Every harness-behavior claim carries its verification verdict (CONFIRMED / REFUTED / PARTIALLY TRUE / UNDOCUMENTED) from the research artifact, none repeated unverified

References

• docs/topics/pocock-course-lanes/PLAN.md (contract + claim ladder)
• plugins/session-flow/skills/handoff/SKILL.md (fork-beats-compaction section)
• plugins/context-guard/README.md (PostCompact evidence-degraded marker)
• Harness-claims research artifact: .work/pocock-course-lanes/harness-claims/RESEARCH.md (memory tier; re-dispatch /discovery:research if the slice is gone)

Metadata

Field	Value
Category	unspecified
Area	session-flow, context-guard
Ecosystem	unspecified

[kyle-sexton]

Lane 3 closed (2026-08-17, commit e064f23f). Twelve rows + house-decisions paragraph in docs/upstream/aihero-course.md "Lane 3"; summary at docs/topics/pocock-lane-3/SUMMARY.md; register Q24-Q30 in the topic ledger (answers locked under the user's standing acceptance of the session's recommendations, stated in the meta session).

Headlines: compact-as-default REJECTED (fork-beats-compaction stands; his own phase-boundaries lesson converges); steered-compact-for-QA carve-out REJECTED with a track-on-event reopen condition (the PostCompact marker already records trigger: manual|auto|unknown, so consumer differentiation is buildable the day evidence justifies it); auto-compact stance ADOPTED as convergent ("human owns the boundary"; we instrument further); C4 refuted mechanics and the transcript-lossless refinement recorded; C5 stays untaught-unknown; C6 keeps its single-pool label. One work item filed: #2995 (context-guard documents the verified auto-compact config surfaces and zones-below-trigger guidance).

---

Generated by Claude Code

[kyle-sexton]

Lane 3 closed (2026-08-17, core-lane chain on claude/pocock-steering-course-00zkvd). Interview-first per contract: 1 round, 5 questions, register gate clean, user confirmed.

Verdicts (full rows in docs/upstream/aihero-core-lanes.md, commit 032e4ea4 — lane 6 harvests): compact-for-QA REJECTED with the evidence-degraded marker affirmed unconditional (boundary QA routes to a fresh-context subagent, which his own tree concedes); auto-compact stance SPLIT — diagnostic half adopted as an anchor phrase (firing = missed boundary, convergent with context-guard zones), prescriptive half rejected (unattended sessions have no human; the instrumented ladder is the licensed version, auto-compact stays as safety net); C4 recorded REFUTED with the corrected same-session mechanism (no doc change needed); C5 tracked-with-cure, never taught until cured; transcript-primary/summary-secondary vocabulary routed to lane 6 (#2904).

Acceptance criteria: ☑ decision rows recorded (in the core-lane record pending lane 6's aihero-course.md collection, per the graduated steering-record precedent) · ☑ work items — zero decided (parity lane; #2973 already covers the known context-guard doc drift) · ☑ lane summary (this comment + the record's lane section) · ☑ every harness claim carries its verdict from the contract's C1–C9 table, none repeated unverified.

---

Generated by Claude Code