// Session wind-down action to capture learnings into issue graph and prepare for handoff via stigmergic signal tables.
[HINT] Download the complete skill directory including SKILL.md and all related files
| name | issues-beads-checkpoint |
| description | Session wind-down action to capture learnings into issue graph and prepare for handoff via stigmergic signal tables. |
| disable-model-invocation | true |
Symlink location: ~/.claude/skills/issues-beads-checkpoint/SKILL.md
Slash command: /issues:beads-checkpoint
Action prompt for session wind-down. Write the worker's state back to the DAG so subsequent workers can resume without information loss. Signal tables are the primary coordination mechanism; all checkpoint state flows through them.
If you need to refactor the issue graph structure during checkpoint (e.g., split epics, merge issues, restructure dependencies), use /issues:beads-evolve first, then complete the checkpoint workflow.
The checkpoint protocol follows eight steps, extending the convention specification's checkpoint write protocol with a buffer depletion check. Execute them in order for each issue that received active work this session. Post-checkpoint checks run between propagation and verification as best-effort notifications.
For each issue that was actively worked on (not just read), read the current notes and parse the signal table.
# Read current notes — field is ABSENT when unset, not null or empty
NOTES=$(bd show <id> --json | jq -r '.[0].notes // ""')
Extract the signal table from between the <!-- stigmergic-signals --> and <!-- /stigmergic-signals --> delimiters.
If no signal table exists, prepare a new one using defaults from /stigmergic-convention: cynefin=complicated, surprise=0.0, progress=not-started, escalation=none, planning-depth=standard.
Also read the issue's description and acceptance criteria to compare expectations against implementation experience. This comparison informs the surprise assessment in the next step.
Assess and update each signal field, setting the Updated column to today's date for any changed value.
Set the surprise score based on how much the actual work diverged from what the issue description led you to expect. Be honest in this assessment; it is the primary metric for downstream calibration and replanning triggers.
A reflexive surprise=0.0 on an issue where work encountered difficulties is misleading pheromone for the next worker.
Conversely, a complicated-domain rewrite of a skill that encounters zero divergence would itself be surprising as a 0.0.
Set the progress field to reflect where the work stands:
exploring: Complex-domain probe phase completed, implementation not yet started.implementing: Work is in progress but not yet ready for verification.verifying: Implementation is complete and verification is underway or passed.blocked: Work cannot proceed due to a dependency, missing information, or external blocker.Do not leave progress at not-started for any issue that received work.
If the implementation experience revealed a different domain than originally assessed, update the cynefin field. When cynefin changes, re-derive planning-depth using the default mapping unless a manual override is warranted.
The cynefin-to-planning-depth mapping: clear to shallow, complicated to standard, complex to deep, chaotic to probe.
When updating cynefin, compare the new value against the prior value read from the signal table in step 1. If the new value represents an upward shift in complexity, send an ntfy notification alerting System 5 that work originally planned as simpler may now require advisory coupling.
The complexity ordering is: clear < complicated < complex < chaotic. Only upward shifts trigger notification. Do not notify on downward reclassification (e.g. complex to complicated) or unchanged values.
This check is best-effort and must never block the checkpoint workflow. If the notification fails (network unavailable, ntfy unreachable), the checkpoint proceeds normally.
# Cynefin reclassification notification
# Only fires on upward complexity shifts
CYNEFIN_ORDER="clear complicated complex chaotic"
old_rank=$(echo "$CYNEFIN_ORDER" | tr ' ' '\n' | grep -n "^${OLD_CYNEFIN}$" | cut -d: -f1)
new_rank=$(echo "$CYNEFIN_ORDER" | tr ' ' '\n' | grep -n "^${NEW_CYNEFIN}$" | cut -d: -f1)
if [ "$new_rank" -gt "$old_rank" ] 2>/dev/null; then
REPO_NAME=$(basename "$(git rev-parse --show-toplevel 2>/dev/null || pwd)")
ntfy-send \
"Issue ${ISSUE_ID} reclassified from ${OLD_CYNEFIN} to ${NEW_CYNEFIN} in ${REPO_NAME}. May now require advisory coupling." \
"$(hostname -s)" \
-H "Title: Cynefin shift: ${ISSUE_ID}" \
-H "Priority: high" \
-H "Tags: warning,${REPO_NAME}" 2>/dev/null || true
fi
Where OLD_CYNEFIN is the value from the signal table read in step 1, NEW_CYNEFIN is the value just assessed, and ISSUE_ID is the current issue identifier.
The || true ensures a failed notification does not interrupt the checkpoint sequence.
After modifying signal values, reconstruct the full notes content. The signal table goes at the beginning of the notes field, followed by any checkpoint-context and escalation-context sections. Write the complete result back:
bd update <id> --notes "$RECONSTRUCTED_NOTES"
The --notes flag replaces the entire field.
All sections must be included in the reconstructed content: the signal table, the checkpoint-context section, and the escalation-context section (if any).
If the issue had no prior signal table, create one at the beginning of the notes field using defaults for any values not being set by this checkpoint, and append any pre-existing notes content after the closing delimiter.
Write a <!-- checkpoint-context --> section in the notes describing the current state.
This section uses replacement semantics: if a prior checkpoint-context exists, replace it entirely with the new content.
Each checkpoint context must be self-contained.
It synthesizes the current position, approach, and remaining work rather than referencing prior checkpoints.
Workers need only the current state to proceed.
Trajectory data is available through bd history <id> which provides native dolt version history per-issue.
<!-- checkpoint-context -->
## State estimate (YYYY-MM-DD)
What was done: [specific accomplishments this session]
What was learned: [key insights, unexpected findings, approach changes]
What remains: [concrete next steps for the next worker]
Downstream impact: [what dependent issues should know about]
<!-- /checkpoint-context -->
The state estimate must be sufficient for the next worker to orient without advisory input. If the issue is being left incomplete (progress=implementing or progress=blocked), this section is mandatory.
If the worker encountered genuine ambiguity that cannot be resolved from DAG context, create an escalation.
Set escalation to pending in the signal table.
Write the question in an <!-- escalation-context --> section.
The question must be precise enough for the human to answer without the worker's full context.
Reference specific alternatives and explain why the DAG does not contain enough information to choose.
<!-- escalation-context -->
## Pending (YYYY-MM-DD)
[Specific question with alternatives and rationale]
<!-- /escalation-context -->
If an escalation-context section already exists (from a prior resolved escalation), append the new question with a date separator to preserve the decision audit trail.
After setting the escalation, move on to other work rather than blocking. Escalation is orthogonal to progress; the issue remains in whatever progress state it was in.
If no escalation is needed, skip this step. Do not modify an existing resolved escalation unless new ambiguity has arisen.
Check whether context discovered during work affects dependent issues. If the worker learned something that changes assumptions for downstream work, update those issues' descriptions or notes to reflect the new understanding.
This is proactive pheromone propagation: rather than waiting for the next worker to discover discrepancies, the current worker corrects the trail.
# Check what depends on this issue
bd dep tree <id> --direction up
For each downstream issue affected by a discovery, update its notes or description:
bd update <downstream-id> --description "Updated: <incorporate discovery that affects this issue>"
Common propagation scenarios include an interface or API changing from what the downstream issue expects, a prerequisite proving harder than expected and changing the downstream scope, an assumption in the downstream description being invalidated, or a technical constraint being discovered that the downstream worker needs to know.
After propagating context, check whether accumulated surprise across the parent epic's children has exceeded the replanning threshold. This implements the MPC (model predictive control) decision rule: when the model (plan) diverges sufficiently from reality (measured surprise), replan rather than continuing on the original trajectory.
This check is best-effort and must never block the checkpoint workflow. If any command fails or the parent epic cannot be determined, skip silently and proceed to step 6 (buffer depletion check).
EPIC_ID=$(bd show <id> --json | jq -r '[.[0].dependencies[] | select(.dependency_type == "parent-child")] | .[0].id // empty')
If EPIC_ID is empty, the current issue has no parent epic.
Skip the remainder of this section and proceed to step 6.
Read all children of the epic and extract their surprise scores from the signal tables in their notes.
EPIC_JSON=$(bd show "$EPIC_ID" --json)
# Extract child IDs (issues linked to the epic via parent-child)
CHILDREN=$(echo "$EPIC_JSON" | jq -r '[.[0].dependents[] | select(.dependency_type == "parent-child")] | .[].id')
# Sum surprise scores across children
TOTAL_SURPRISE=0
CONTRIBUTING=0
for child_id in $CHILDREN; do
CHILD_NOTES=$(bd show "$child_id" --json | jq -r '.[0].notes // ""')
# Extract surprise value from signal table row: | surprise | <value> | <date> |
SURPRISE=$(echo "$CHILD_NOTES" | grep -oP 'surprise\s*\|\s*\K[0-9.]+' || echo "0.0")
if [ "$SURPRISE" != "0.0" ] && [ "$SURPRISE" != "0" ]; then
TOTAL_SURPRISE=$(echo "$TOTAL_SURPRISE + $SURPRISE" | bc)
CONTRIBUTING=$((CONTRIBUTING + 1))
fi
done
The replanning threshold default is 2.0, representing the cumulative surprise budget before the plan should be reconsidered.
When the CUE schema at ~/projects/sciexp/planning/schemas/stigmergic-workflow/schema.cue is updated with #ReplanningThreshold, this hardcoded value should be read from there instead.
REPLANNING_THRESHOLD=2.0
EXCEEDS=$(echo "$TOTAL_SURPRISE > $REPLANNING_THRESHOLD" | bc -l)
if [ "$EXCEEDS" -eq 1 ]; then
NTFY_TOPIC=$(hostname -s)
REPO_NAME=$(basename "$(git rev-parse --show-toplevel 2>/dev/null || pwd)")
ntfy-send \
"Epic ${EPIC_ID} in ${REPO_NAME}: accumulated surprise ${TOTAL_SURPRISE} exceeds threshold ${REPLANNING_THRESHOLD} (${CONTRIBUTING} issues contributing). Consider running /issues:beads-evolve to restructure the plan." \
"${NTFY_TOPIC}" \
-H "Title: Replanning needed: ${EPIC_ID}" \
-H "Priority: urgent" \
-H "Tags: rotating_light,${REPO_NAME}" 2>/dev/null || true
fi
When the notification fires, it means the epic's plan has drifted far enough from reality that continuing on the original trajectory is likely wasteful.
The recommended response is to run /issues:beads-evolve to restructure the plan based on what has been learned, then resume work from the revised issue graph.
After writing checkpoint state, check whether the current epic has ready work remaining. An epic with unclosed children but zero ready issues is stuck: all remaining work is blocked, and the next session will waste orient time discovering an empty queue. Detecting this at checkpoint time lets the current worker emit a notification before winding down.
Identify the current issue's parent epic:
EPIC_ID=$(bd show <id> --json | jq -r '[.[0].dependencies[] | select(.dependency_type == "parent-child")] | .[0].id // empty')
If the issue has no parent epic (standalone issue or the issue itself is an epic), skip this check.
Query the epic's children to determine total and closed counts:
EPIC_JSON=$(bd show "$EPIC_ID" --json)
TOTAL_CHILDREN=$(echo "$EPIC_JSON" | jq '[.[0].dependents[] | select(.dependency_type == "parent-child")] | length')
CLOSED_CHILDREN=$(echo "$EPIC_JSON" | jq '[.[0].dependents[] | select(.dependency_type == "parent-child") | select(.status == "closed")] | length')
Three outcomes are possible:
Epic complete (all children closed, TOTAL == CLOSED): no notification. Epic completion is handled by a separate notification (nix-iwo.1). Skip the rest of this step.
Normal state (unclosed children exist and some are ready): no notification. The buffer is healthy and the next worker has work to pick up.
Epic stuck (unclosed children exist but none are ready): send notification. All remaining work is blocked, and no forward progress is possible without unblocking action.
Check the ready count scoped to this epic:
READY_COUNT=$(bd ready --parent "$EPIC_ID" --json 2>/dev/null | jq 'length' 2>/dev/null || echo "0")
If ready issues remain or the epic is complete, no action is needed. If the buffer is depleted, send a notification:
if [ "$READY_COUNT" -eq 0 ] && [ "$CLOSED_CHILDREN" -lt "$TOTAL_CHILDREN" ]; then
NTFY_TOPIC=$(hostname -s)
REPO_NAME=$(basename "$(git rev-parse --show-toplevel 2>/dev/null || pwd)")
UNCLOSED=$((TOTAL_CHILDREN - CLOSED_CHILDREN))
ntfy-send \
"Epic ${EPIC_ID} in ${REPO_NAME} has ${UNCLOSED} unclosed issues but 0 are ready. All remaining work is blocked. Review dependencies to unblock progress." \
"${NTFY_TOPIC}" \
-H "Title: Buffer depleted: ${EPIC_ID}" \
-H "Priority: high" \
-H "Tags: warning,${REPO_NAME}" 2>/dev/null || true
fi
This check is best-effort.
Failures in any command (network errors, unexpected bd output, jq parsing) must not block the checkpoint workflow.
The || true and || echo "0" guards ensure the surrounding steps proceed regardless.
Before committing, verify graph integrity:
# Must be zero — cycles corrupt the graph
bd dep cycles
# Check for structural issues
bd lint
# Confirm ready queue makes sense
bd ready | head -5
If cycles are detected, identify and break them:
bd dep cycles --verbose
bd dep remove <upstream> <downstream>
For each issue worked on, verify its relationship to its parent epic.
Run bd show <id> --json and confirm that the dependency linking it to its epic uses type parent-child.
If it uses child-of, parent, blocks, or any other type for what should be containment, fix it:
bd dep remove <child-id> <epic-id>
bd dep add <child-id> <epic-id> --type parent-child
For each active epic, verify that bd epic status <epic-id> shows the expected child count.
If the count is zero or lower than expected, investigate for mistyped containment dependencies.
When creating issues during checkpoint (discovered work, follow-ups), always specify the parent epic:
# Preferred: parent at creation time
bd create "Discovered: ..." -t bug -p 2 --parent <relevant-epic-id>
bd dep add <new-id> <current-id> --type discovered-from
# If parent was omitted at creation, add it explicitly
bd dep add <new-id> <relevant-epic-id> --type parent-child
Do not rely on blocks relationships for containment.
An issue connected to an epic via blocks is a sequencing constraint, not containment, and will not appear in bd epic status child counts.
For each issue that was actively worked on, verify its signal table reflects actual work state rather than stale defaults.
An issue that received implementation work but still shows progress=not-started indicates an incomplete checkpoint.
A surprise=0.0 on an issue where work encountered difficulties is misleading pheromone.
If cynefin classification changed during work, verify the signal table reflects it and that planning-depth was re-derived.
For issues being left incomplete (progress=implementing or progress=blocked), verify that a checkpoint-context section exists with a self-contained state estimate.
Push beads state to the dolt remote for backup:
bd dolt commit -m "checkpoint: <id>"
bd dolt push
The commit message should reference the specific issue being checkpointed. For multi-issue checkpoints, name the primary issues or summarize by category.
After completing the per-issue checkpoint steps above, run session-level diagnostics to verify overall state.
# Current state (compare to session start)
bd status
# Epic progress change
bd epic status
For additional session summary context:
# What's ready for the next session
bd ready | head -5
# Stale issues that may need attention
bd stale
Determine session scope by counting issues touched:
# Count issues modified this session (from date-filtered list)
bd list --updated-after <session-start-date> --json | jq length
Tailor summary depth to session scope.
Small sessions (1-3 issues touched): full detail on each change with complete before/after state, full descriptions for modified issues, complete dependency changes with rationale.
Medium sessions (4-10 issues touched): summary table of changes by category (closed with completion notes, updated with what changed, created with traceability, dependency additions/removals), highlight key closures and discoveries.
Large sessions (10+ issues touched): epic-level aggregation of changes, statistics only ("8 closed, 3 created, 12 dependencies modified"), only critical highlights (blockers discovered, scope corrections, priority changes), per-epic progress deltas ("Domain layer: 42% to 58%, +5 closed").
Consider what was learned this session that the issue graph does not yet reflect.
Scope changes: Was the work larger or smaller than the issue described? Did implementation reveal unanticipated complexity? Should the issue description be updated to reflect actual scope?
Discovered work: Were bugs found during implementation? Was technical debt identified that should be tracked? Are there follow-up enhancements worth capturing?
Dependency corrections: Were any listed blockers not actually blocking? Were hidden prerequisites discovered? Can work that was assumed sequential actually proceed in parallel?
Priority shifts: Did this work reveal that other issues are more or less critical than assigned? Were any P3 items actually critical blockers? Were any P1 items actually low-impact?
If reflection revealed priority mismatches, update before committing:
# Issue proved more critical than marked
bd update <id> --priority 1
# Issue proved less critical than marked
bd update <id> --priority 3
# Add context for the change
bd comments add <id> "Priority adjusted: discovered this blocks X during implementation"
For each learning identified in the reflection, execute the appropriate update.
Scope and understanding changes:
bd update <issue-id> --description "Revised: <what changed and why>"
Discovered issues:
# Create with traceability and epic parentage
bd create "<title>" -t <bug|task|feature> -p <priority> --parent <relevant-epic-id>
bd dep add <new-id> <current-issue-id> --type discovered-from
Dependency corrections:
# Remove false dependencies
bd dep remove <not-actually-blocking> <issue-id>
# Add discovered dependencies
bd dep add <actual-blocker> <issue-id>
Completed work:
# Close and see what was unblocked
bd close <issue-id> --reason "Implemented in commit $(git rev-parse --short HEAD). <notable learnings>" --suggest-next
When closing issues, show the cascade effect:
# See what completing this unblocked
bd close <id> --reason "..." --suggest-next
# Show full unblock chain
bd dep tree <closed-id> --direction up
Extract structured unblock data:
bd dep tree <closed-id> --direction up --json | jq '{
direct: [.children[] | .id],
direct_count: [.children[]] | length,
total_downstream: [.. | .id? // empty] | unique | length
}'
Present the cascade: "Closing nix-50f.2 unblocked 3 issues (nix-50f.1, nix-l2a.2, nix-l2a.5). These 3 issues unblock 7 more downstream. Total downstream impact: 10 issues now closer to ready."
For medium and large sessions, show epic-level progress change:
bd epic status
Present as progress deltas: "Domain layer: 42% to 58% (closed 5 issues). Infrastructure: 20% to 35% (closed 3 issues, created 2). Frontend: unchanged (0 issues touched)."
Prepare for the next session to pick up cleanly.
If work is complete:
# Close and immediately see what was unblocked
bd close <issue-id> --reason "Implemented in..." --suggest-next
Review unblocked issues and check whether their descriptions need updating given what was learned.
Do not check or report epic closure eligibility; see the closure policy in /issues:beads-prime.
If work is incomplete, ensure the checkpoint-context section written in step 3 is self-contained.
The signal table's progress field should be implementing or blocked as appropriate.
Use bd update <id> --claim to atomically claim an issue when resuming work in a future session.
Produce a prose summary for the next session:
Session summary:
Completed:
- <issue-id>: <brief description of what was implemented>
Discovered:
- <issue-id>: <why this was created, what triggered discovery>
Learned:
- <key insight that changed understanding>
- <assumption that proved incorrect>
Next:
- Recommended starting point: <issue-id> (<why this is the logical next step>)
- Alternative entry points: <issue-id>, <issue-id>
Warnings:
- <blockers discovered but not resolved>
- <scope risks identified>
This mirrors beads-orient's synthesis section but for session end.
The next session can use this summary plus /session-orient (or /issues:beads-orient in beads-only repos) to quickly resume.
Before ending the session, ensure the next agent can pick up cleanly via /session-orient (or /issues:beads-orient in beads-only repos).
# Quick check of top ready issues
bd ready | head -5
# Review the top candidate's description
bd show <top-candidate-id>
If the description is stale or incomplete based on what was learned this session:
bd update <top-recommendation-id> --description "Updated: <incorporate session learnings that affect this issue>"
The issue graph becomes the handoff.
When descriptions are accurate and signal tables reflect actual state, /session-orient (or /issues:beads-orient in beads-only repos) in the next session will find actionable, up-to-date work.
Provide the user a summary scaled to session scope.
Small sessions: full detail with issues closed (completion notes), updated (what changed), created (traceability), dependencies modified (rationale), unblock cascade (what is now ready).
Medium sessions: category summary with counts, key items named, epic progress deltas, next session recommendation.
Large sessions: executive summary with statistics, epic progress deltas, only critical items, next session recommendation with rationale.
The next session can run /session-orient (or /issues:beads-orient in beads-only repos) to see the updated project state.
Reference docs (read only if deeper patterns needed):
/stigmergic-convention -- signal table schema, field definitions, and update protocol/issues:beads-evolve -- comprehensive adaptation patterns/issues:beads -- comprehensive reference for all beads workflows and commands