Timeboxing regression: explicit in-thread constraints do not override conflicting defaults

[hugolytics]

Goal

Make same-day explicit session facts override conflicting profile/default memory, including conditional defaults such as "go to the gym unless another sport is already planned or explicitly stated for that day."

Problem

In the latest planning thread, explicit session instructions are not overriding conflicting default/profile constraints. The agent continues using stale defaults even after correction.

Evidence

Slack thread: https://biolyticsai.slack.com/archives/C0AA6HC1RJL/p1772840876953279

• User says gym should be at 12:00 (ts=1772841791.541729), but agent repeatedly reverts to default evening rhythm before partially correcting.
• User says planning session is already on calendar at 16:00 (ts=1772841893.355329), but the next reply does not reconcile against that anchor.
• Broader desired behavior: a default activity preference should be suppressible when same-day context contains a stronger conflicting activity (for example another sport on calendar or explicitly stated in-thread).

Scope

• Use existing aspect_classification metadata to model conditional applicability during timeboxing memory selection.
• Evaluate conditional_on_present, conditional_on_absent, and excludes_aspect_ids against same-day session/calendar context before selecting active constraints.
• Suppress weaker default/profile constraints when a stronger same-day explicit/session constraint or anchor conflicts in the same planning domain.
• Emit deterministic debug/snapshot fields for kept vs suppressed memory candidates and suppression reasons.
• Add regression tests covering the reproduced gym/planning-time conflict and the gym-vs-other-sport conditional-default case.

Non-goals

• Dual extraction deduplication per Refine turn (Timeboxing Refine: dual constraint extraction per turn (graphflow_turn + refine_background_memory both extracting) #104).
• Graphiti DB deployment/runtime wiring audit (Canonical Graphiti memory cutover: Neo4j backend, no Mem0, no file artifacts #90).
• Full cross-agent memory architecture redesign (Memory architecture continuation: shared kernel + cross-agent reconciliation rollout #64).

Acceptance Criteria

1. Given a conflict between a session constraint and a profile/default constraint, the session constraint wins for this thread.
2. Given a default/profile activity preference that is conditional, when same-day context contains a blocking/conflicting activity, the default/profile constraint is suppressed before active selection.
3. Given no conflicting same-day blocker, the same default/profile constraint remains eligible and can still be selected.
4. Agent/session debug output includes deterministic kept vs suppressed counts and suppression reasons for this selection pass.
5. Regression tests reproduce:
   • the gym/planning-time override pattern from the reported thread,
   • a gym default suppressed by another same-day sport,
   • a non-conflict case where gym remains active.

Design / Ownership Notes

• Primary ownership: src/fateforger/agents/timeboxing/agent.py selection/reconciliation path.
• Reuse existing aspect_classification schema and avoid regex/substring heuristics.
• Keep storage/canonicalization behavior from Replace Mem0 with temporal memory backend (Graphiti path) + reconciliation policy #81 intact unless a narrowly scoped fix is required.

Validation Plan

• Add focused unit tests for same-day conditional suppression and precedence rules.
• Run targeted timeboxing memory/constraint test subset, including new regressions.
• Verify debug snapshot fields expose suppression results without introducing heuristic behavior.

Links

• Epic: EPIC: Timeboxing reliability stabilization (memory reconciliation -> planner fidelity -> audit hardening) #74
• Related: Timeboxing local constraint mirror: canonicalize shared constraints + intelligent pruning #69, Replace Mem0 with temporal memory backend (Graphiti path) + reconciliation policy #81, timeboxing capture-inputs: respect explicit scope and suppress legacy GTD/admin task bleed-through #82, Timeboxing Refine: dual constraint extraction per turn (graphflow_turn + refine_background_memory both extracting) #104

[hugolytics]

Additional evidence from exhaustive rerun (thread 1772840876.953279):

• ts=1772841791.541729 user: "No I said gym at 12!"
• ts=1772841893.355329 user: "planning session is already on my calendar at 16:00"
• Despite corrections, subsequent turns continue to show strong default/profile influence and drift.

Open Items

• To decide: precedence table should be strict override or conflict-confirmation mode when defaults conflict with explicit session instructions.
• To do: implement deterministic precedence contract and render conflict-resolution summary.
• Blocked by: none

[hugolytics]

Implementation kickoff (Option 3 selected by user)

Goal

• Preserve all session constraints during Stage 4/5 patching, while capping only durable constraints.
• Keep patching as primary action; memory extraction/update remains asynchronous background work.
• Ensure latest user message semantics supersede conflicting remembered defaults for this thread.

Scope

• Timeboxing agent extraction + refine selection paths only.
• No schema migration in this slice.
• No new handoff/routing behavior in this slice.

Acceptance criteria (this slice)

1. Stage 4/5 patcher input always includes all session-scoped active constraints.
2. Durable constraints are bounded by existing refine cap and treated as secondary lane.
3. Constraint extraction context includes existing session constraints so the interpreter can patch/add against current state.
4. Background memory update remains non-blocking for patch execution path.
5. Tests cover selection behavior and extraction-context behavior.

Design/ownership notes

• Coordinator/agent owns orchestration order and patcher input selection.
• NLU interpreter owns extraction output shape; prompt updated to include existing constraints context.
• Local constraint store remains source for in-thread session constraints.

Validation plan

• tests/unit/test_timeboxing_constraint_selection.py
• targeted new/updated tests for extraction payload context + background flow

User waiver

• User explicitly requested direct implementation now (“update the tickets and go and implement it”), so options-comparison step is waived for this checkpoint.

Open Items

• To decide: none
• To do: implement + run targeted tests
• Blocked by: none

[hugolytics]

Kickoff Checkpoint

Active branch: codex/issue-91-option3-session-constraints
Notebook mode: code-only-mode
Status: Roadmap

Summary

• Confirmed Replace Mem0 with temporal memory backend (Graphiti path) + reconciliation policy #81 baseline is already merged into the active branch and its targeted test subset passes.
• Refined Timeboxing regression: explicit in-thread constraints do not override conflicting defaults #91 scope to focus on conditional applicability and same-day conflict suppression in the timeboxing memory-selection path.
• Selected implementation direction: use existing aspect_classification metadata at retrieval/selection time rather than pushing this logic into planner heuristics.

Planned validation

• Add regression tests for:
  • session constraint overriding profile/default constraint,
  • gym default suppressed by another same-day sport,
  • non-conflict case where gym remains eligible.
• Run the targeted timeboxing memory/constraint test subset after implementation.

Open Items

• To decide: none
• To do: implement conditional suppression in stage-context reconciliation and add regression coverage
• Blocked by: none

[hugolytics]

Implementation checkpoint (Option 3)

Shipped in branch codex/issue-91-option3-session-constraints:

1. Session-preserving Stage 4/5 constraint selection

• _select_constraints_for_refine_patcher now always keeps all session-scope constraints.
• Durable constraints are capped separately (TIMEBOXING_LIMITS.refine_patcher_constraint_limit) and appended as secondary lane.
• Added richer debug counters: session_selected, durable_total, durable_selected, durable_dropped.

2. Extraction context now includes current session constraints

• _ConstraintInterpretationPayload now carries existing_constraints.
• _interpret_constraints now injects serialized in-thread session constraints so extraction can reconcile against current state.
• NLU prompt updated to require full updated session set when scope=session and to apply latest user message precedence over existing constraints.

3. Session snapshot replacement persistence path

• Added ConstraintStore.replace_session_constraints(...).
• Background extraction path now uses replace (for session scope) instead of additive upsert/add, preventing session snapshot drift.

Tests added/updated

• tests/unit/test_timeboxing_constraint_selection.py
  • now verifies all session constraints are preserved and durable is capped.
• tests/unit/test_timeboxing_stateless_agents.py
  • verifies _interpret_constraints payload includes existing_constraints.
• tests/unit/test_timeboxing_constraint_extraction_background.py
  • verifies session extraction path calls replace_session_constraints.
• tests/unit/test_timeboxing_constraint_store_canonicalization.py
  • verifies replace-session operation replaces thread snapshot.

Validation run

• poetry run pytest -q tests/unit/test_timeboxing_constraint_selection.py tests/unit/test_timeboxing_stateless_agents.py tests/unit/test_timeboxing_constraint_extraction_background.py tests/unit/test_timeboxing_constraint_store_canonicalization.py
• poetry run pytest -q tests/unit/test_timeboxing_durable_constraints.py
• Result: all passing.

Open Items

• To decide: whether we also want explicit extractor delta ops (patch/archive) instead of full-session-set replacement
• To do: run Slack end-to-end replay and inspect session logs for extraction+patch latency
• Blocked by: none

[hugolytics]

Checkpoint: Option 3 implementation completed and tested on branch codex/issue-91-option3-session-constraints.

Status

• Code changes merged into working branch (not yet committed/pushed by agent).
• Targeted unit suites are green.
• Ready for Slack replay validation and PR creation.

Open Items

• To decide: extractor delta-op model (patch/archive) vs full-session snapshot replacement
• To do: run live Slack replay + log audit; open PR from branch
• Blocked by: none

[hugolytics]

Progress Checkpoint

Status: Tested
Branch: codex/issue-91-option3-session-constraints
Notebook mode: code-only-mode

Summary

• Added session-time constraint suppression for same-day conflicts in src/fateforger/agents/timeboxing/agent.py.
• Reused existing aspect_classification metadata instead of adding heuristic parsing.
• Fixed the specific failure mode where a same-aspect session override with a different time survived alongside the profile default.
• Added conditional suppression for defaults like gym when another stronger same-day sport aspect is present.
• Added deterministic suppression evidence to the constraints_active_snapshot debug event.
• Documented the behavior in src/fateforger/agents/timeboxing/README.md.

Tests run

• poetry run pytest tests/unit/test_timeboxing_constraint_selection.py -q
• poetry run pytest tests/unit/test_timeboxing_constraint_selection.py tests/unit/test_timeboxing_durable_constraints.py tests/unit/test_constraint_retriever.py tests/unit/test_timeboxing_memory_backend_selection.py -q

Observed result

• Focused constraint-selection suite: 7 passed
• Broader targeted subset: 27 passed

Scope / cleanliness

• The worktree was already dirty before this ticket slice. I kept #91 scoped to:
  • src/fateforger/agents/timeboxing/agent.py
  • src/fateforger/agents/timeboxing/README.md
  • tests/unit/test_timeboxing_constraint_selection.py
  • docs/plans/2026-03-08-timeboxing-session-constraint-override-design.md
  • docs/plans/2026-03-08-timeboxing-session-constraint-override.md
• Unrelated pre-existing modified files remain out of scope for this checkpoint.

Open Items

• To decide: none
• To do: run any broader integration or Slack audit replay you want before PR creation
• Blocked by: none

[hugolytics]

Ticket Sync Checkpoint

Status: Tested
Branch: codex/issue-91-option3-session-constraints
Head: be3f820

Summary

• #91 now contains the active implementation for same-day session override and conditional default suppression.
• #81 has been re-validated as the merged baseline under this work, and the relationship between the two issues is now explicit in GitHub.
• Current implementation/test evidence remains on this branch; no PR exists yet.

Validation state

• Focused regression suite and broader targeted subset are green on this branch.
• Latest implementation checkpoint: Timeboxing regression: explicit in-thread constraints do not override conflicting defaults #91 (comment)

Open Items

• To decide: none
• To do: create PR or continue with additional related follow-on issues if desired
• Blocked by: none

[hugolytics]

Merge / Closeout Handoff

Status: Tested
Closure vehicle: PR #106 (after PR #105 lands)

Acceptance Criteria validation

• Session constraint wins over conflicting profile/default constraint in-thread.
• Conditional default/profile activity is suppressed when same-day context contains a blocker/conflict.
• Non-conflict case keeps the default/profile constraint eligible.
• Session debug output now includes deterministic suppression counts/reasons.
• Regression tests cover gym/planning-time override, gym-vs-other-sport suppression, and non-conflict preservation.

Validation evidence

• poetry run pytest tests/unit/test_timeboxing_constraint_selection.py -q
• poetry run pytest tests/unit/test_timeboxing_constraint_selection.py tests/unit/test_timeboxing_durable_constraints.py tests/unit/test_constraint_retriever.py tests/unit/test_timeboxing_memory_backend_selection.py -q
• Result: focused suite 7 passed, targeted subset 27 passed

No-code-loss merge instructions for adjacent agents

1. Do not merge or close #91 until PR issue/96+81: calendar sync idempotency + graphiti memory backend #105 is merged first.
2. After PR issue/96+81: calendar sync idempotency + graphiti memory backend #105 lands, retarget or rebase PR Codex/issue-91-option3-session-constraints #106 onto main.
3. Before merging PR Codex/issue-91-option3-session-constraints #106, verify its diff still contains only #91 scope:
   • timeboxing selection/suppression logic
   • test_timeboxing_constraint_selection.py
   • timeboxing README note
   • plan docs (if you keep them in the PR)
4. Explicitly check that no #91 code was dropped during rebase/retarget by re-running the targeted test subset after the base change.
5. Close #91 only after the rebased/retargeted PR is merged to main, with the PR reference and test evidence in the close comment.

Open Items

• To decide: whether the plan docs should remain in the #91 PR or be omitted if your maintainers prefer code-only PRs
• To do: retarget/rebase PR Codex/issue-91-option3-session-constraints #106 after PR issue/96+81: calendar sync idempotency + graphiti memory backend #105 merge, rerun targeted tests, merge, close #91
• Blocked by: PR issue/96+81: calendar sync idempotency + graphiti memory backend #105 not merged yet

[hugolytics]

Substantial checkpoint

Current status

• Option-3 core behavior is implemented on codex/issue-91-option3-session-constraints.
• Fresh targeted validation re-run is green.
• Issue is NOT ready to close yet (AC2/AC3 still open).

Branch

• codex/issue-91-option3-session-constraints

PR

• Codex/issue-91-option3-session-constraints #106 Codex/issue-91-option3-session-constraints #106
• Dependency: PR Codex/issue-91-option3-session-constraints #106 currently targets issue/96-calendar-sync-idempotency (stacked on issue/96+81: calendar sync idempotency + graphiti memory backend #105), not main.
• Final closure for this issue will only happen after a PR carrying this work is merged to main.

Scope cleanliness

• Worktree was already dirty before this checkpoint (contains unrelated modified files).
• In-scope files for this issue slice:
  • src/fateforger/agents/timeboxing/agent.py
  • src/fateforger/agents/timeboxing/nlu.py
  • src/fateforger/agents/timeboxing/preferences.py
  • tests/unit/test_timeboxing_constraint_selection.py
  • tests/unit/test_timeboxing_stateless_agents.py
  • tests/unit/test_timeboxing_constraint_extraction_background.py
  • tests/unit/test_timeboxing_constraint_store_canonicalization.py
• Unrelated dirty files are intentionally kept out of this issue scope.

Commands/tests run

• poetry run pytest -q tests/unit/test_timeboxing_constraint_selection.py tests/unit/test_timeboxing_stateless_agents.py tests/unit/test_timeboxing_constraint_extraction_background.py tests/unit/test_timeboxing_constraint_store_canonicalization.py tests/unit/test_timeboxing_durable_constraints.py
• Result: 42 passed.

Acceptance criteria status

• AC1 (session constraints override conflicting defaults in-thread): PARTIAL
  • Implemented precedence-critical mechanics (all session constraints preserved in Stage 4/5 patcher input; session extraction snapshot replacement).
  • Still need explicit end-to-end replay evidence tied to the gym/planning conflict thread.
• AC2 (deterministic conflict-resolution summary in response): NOT DONE
• AC3 (regression test reproduces gym/planning conflict pattern): NOT DONE

Exact next step

• Implement AC2 response summary + AC3 regression case, then run Slack replay + log correlation and post evidence.
• After dependent base PR(s) land, retarget/rebase Codex/issue-91-option3-session-constraints #106 to main path and re-run validation.

Open Items

• To decide: exact response-contract format for conflict-resolution summary (section text vs structured section block)
• To do: AC2 implementation, AC3 test, Slack replay evidence, retarget/rebase after stack unblocks
• Blocked by: stacked PR order (issue/96+81: calendar sync idempotency + graphiti memory backend #105 -> Codex/issue-91-option3-session-constraints #106) for final landing to main

[hugolytics]

Kickoff Checkpoint (2026-03-08)

Current status

• Resuming active work on issue Timeboxing regression: explicit in-thread constraints do not override conflicting defaults #91 using branch codex/issue-91-option3-session-constraints.
• Applying strict no-code-loss + stacked-PR safeguards before any rebase/retarget/cherry-pick/merge operations.

Branch

• codex/issue-91-option3-session-constraints

PR

• Codex/issue-91-option3-session-constraints #106 Codex/issue-91-option3-session-constraints #106
• Current base: issue/96-calendar-sync-idempotency (stacked, not final landing to main yet).

Start-of-work scope cleanliness

• Worktree is already dirty at kickoff.
• Baseline dirty set captured from git status --porcelain before new edits in this cycle.
• I will keep Timeboxing regression: explicit in-thread constraints do not override conflicting defaults #91 behavior work scoped and avoid mixing unrelated files into any merge/closure recommendation.

Commands run

• git branch --show-current
• git status --porcelain
• gh pr status
• gh pr view 106 --json number,title,url,baseRefName,headRefName,state,isDraft,mergeStateStatus,body,commits,files,labels
• gh issue view 91 --json number,title,url,state,body,labels,assignees,comments

Acceptance criteria status (from issue #91)

• AC1: Partial (implementation exists; needs current-cycle revalidation evidence after any further code changes).
• AC2: Not done (deterministic conflict-resolution summary still open).
• AC3: Not done (explicit regression case still open).
• AC4: Partial (debug snapshot fields exist; needs verification against final AC wording).
• AC5: Partial (coverage exists but still requires explicit AC3 evidence in this cycle).

Exact next step

• Add/confirm failing-first regression coverage for AC2/AC3, then implement scoped fix and rerun targeted suite.

Open Items

• To decide: exact response-surface format for conflict-resolution summary in user-facing output.
• To do: AC2 + AC3 implementation/verification; post substantial progress checkpoint with test evidence.
• Blocked by: final landing path is stacked (#106 depends on base branch flow before merge-to-main readiness).