Execution Tracking

Abstraction layer for agent execution tracking. Other skills (e.g., swain-design) express intent using abstract terms; this skill translates that intent into concrete CLI commands.

Before first use: Read skills/swain-do/references/tk-cheatsheet.md for complete command syntax, flags, ID formats, and anti-patterns.

Artifact handoff protocol

This skill receives handoffs from swain-design based on a four-tier tracking model:

Tier	Artifacts	This skill's role
Implementation	SPEC	Create a tracked implementation plan and task breakdown before any code is written
Coordination	EPIC, VISION, JOURNEY	Do not track directly — swain-design decomposes these into children first, then hands off the children
Research	SPIKE	Create a tracked plan when the research is complex enough to benefit from task breakdown
Reference	ADR, PERSONA, RUNBOOK	No tracking expected

If invoked directly on a coordination-tier artifact (EPIC, VISION, JOURNEY) without prior decomposition, defer to swain-design to create child SPECs first, then create plans for those children.

Term mapping

Other skills use these abstract terms. This skill maps them to the current backend (

tk

Abstract term	Meaning	tk command
implementation plan	Top-level container grouping all tasks for a spec artifact	`tk create "Title" -t epic --external-ref <SPEC-ID>`
task	An individual unit of work within a plan	`tk create "Title" -t task --parent <epic-id>`
origin ref	Immutable link from a plan to the spec that seeded it	`--external-ref <ID>` flag on epic creation
spec tag	Mutable tag linking a task to every spec it affects	`--tags spec:<ID>` on create
dependency	Ordering constraint between tasks	`tk dep <child> <parent>` (child depends on parent)
ready work	Unblocked tasks available for pickup	`tk ready`
claim	Atomically take ownership of a task	`tk claim <id>`
complete	Mark a task as done	`tk add-note <id> "reason"` then `tk close <id>`
abandon	Close a task that will not be completed	`tk add-note <id> "Abandoned: <why>"` then `tk close <id>`
escalate	Abandon + invoke swain-design to update upstream artifacts	Abandon, then invoke swain-design skill

Configuration and bootstrap

Config stored in

.agents/execution-tracking.vars.json

(created on first run). Read references/configuration.md for first-run setup questions, config keys, and the 6-step bootstrap workflow.

Statuses

tk accepts exactly three status values:

open

in_progress

closed

. Use the

status

command to set arbitrary statuses, but the dependency graph (

ready

blocked

) only evaluates these three.

To express abandonment, use

tk add-note <id> "Abandoned: ..."

then

tk close <id>

— see Escalation.

Operating rules

Always include
--description
(or
```
-d
```
) when creating issues — a title alone loses the "why" behind a task. Future agents (or your future self) picking up this work need enough context to act without re-researching.
Create/update tasks at the start of work, after each major milestone, and before final response — this keeps the tracker useful as a live dashboard rather than a post-hoc record.
Keep task titles short and action-oriented — they appear in
```
tk ready
```
output, tree views, and notifications where space is limited.
Store handoff notes using
```
tk add-note <id> "context"
```
rather than ephemeral chat context — chat history is lost between sessions, but task notes persist and are visible to any agent or observer.
Include references to related artifact IDs in tags (e.g.,
```
spec:SPEC-003
```
) — this enables querying all work touching a given spec.
Prefix abandonment reasons with
Abandoned:
when closing incomplete tasks — this convention makes abandoned work findable so nothing silently disappears.
Use
ticket-query
for structured output — when you need JSON for programmatic use, pipe through
```
ticket-query
```
(available in the vendored
```
bin/
```
directory) instead of parsing human-readable output. Example:
```
ticket-query '.status == "open"'
```

TDD enforcement

Strict RED-GREEN-REFACTOR with anti-rationalization safeguards and completion verification. Read references/tdd-enforcement.md for the anti-rationalization table, task ordering rules, and evidence requirements.

Spec lineage tagging

Use

--external-ref SPEC-NNN

on plan epics (immutable origin) and

--tags spec:SPEC-NNN

on child tasks (mutable). Query:

ticket-query '.tags and (.tags | contains("spec:SPEC-003"))'

. Cross-plan links:

tk link <task-a> <task-b>

Escalation

When work cannot proceed as designed, abandon tasks and escalate to swain-design. Read references/escalation.md for the triage table, abandonment commands, escalation workflow, and cross-spec handling.

"What's next?" flow

Run

tk ready

for unblocked tasks and

ticket-query '.status == "in_progress"'

for in-flight work. If

.tickets/

is empty or missing, defer to

bash "$(find "$(git rev-parse --show-toplevel 2>/dev/null || pwd)" -path '*/swain-design/scripts/chart.sh' -print -quit 2>/dev/null)" ready

for artifact-level guidance.

Context on claim

When claiming a task tagged with

spec:<ID>

, show the Vision ancestry breadcrumb to provide strategic context. Run

bash "$(find "$(git rev-parse --show-toplevel 2>/dev/null || pwd)" -path '*/swain-design/scripts/chart.sh' -print -quit 2>/dev/null)" scope <SPEC-ID> 2>/dev/null | head -5

to display the parent chain. This tells the agent/operator how the current task connects to project strategy.

Artifact/tk reconciliation

When specwatch detects mismatches (

TK_SYNC

TK_ORPHAN

.agents/specwatch.log

), read references/reconciliation.md for the mismatch types, resolution commands, and reconciliation workflow.

Session bookmark

After state-changing operations, update the bookmark:

bash "$(find . .claude .agents -path '*/swain-session/scripts/swain-bookmark.sh' -print -quit 2>/dev/null)" "<action> <task-description>"

Superpowers skill chaining

When superpowers is installed, swain-do must invoke these skills at the right moments — do not skip them or inline the work:

Before writing code for any task: Invoke the
```
test-driven-development
```
skill. Write a failing test first (RED), then make it pass (GREEN), then refactor. This applies to every task, not just the first one.
When dispatching parallel work: Invoke
```
subagent-driven-development
```
(if subagents are available and tasks are independent) or
```
executing-plans
```
(if serial). Read references/execution-strategy.md for the decision tree.
Before claiming any task or plan is complete: Invoke
```
verification-before-completion
```
. Run the verification commands, read the output, and only then assert success. No completion claims without fresh evidence.

Detection:

ls .agents/skills/test-driven-development/SKILL.md .claude/skills/test-driven-development/SKILL.md 2>/dev/null

— if at least one path exists, superpowers is available. Cache the result for the session.

When superpowers is NOT installed, swain-do uses its built-in TDD enforcement (see references/tdd-enforcement.md) and serial execution.

Plan ingestion (superpowers integration)

When a superpowers plan file exists, use the ingestion script (

skills/swain-do/scripts/ingest-plan.py

) instead of manual task decomposition. Read references/plan-ingestion.md for usage, format requirements, and when NOT to use it.

Execution strategy

Selects serial vs. subagent-driven execution based on superpowers availability and task complexity. Read references/execution-strategy.md for the decision tree, detection commands, and worktree-artifact mapping.

Pre-plan implementation detection

Before creating a plan for a SPEC, check whether it is already implemented. Perform all three steps:

Step 1 — Check git history for the spec ID:

bash

git log --oneline --all | grep -i "<SPEC-ID>"

Matching commits are a strong signal.

Step 2 — Check whether deliverable files exist: Read the spec artifact to identify described deliverables (files, scripts, configs). Check whether those files exist using

ls

Glob

Step 3 — Re-run tests fresh (REQUIRED if tests are mentioned in the spec): Do NOT trust any prior claim or cached result. Run the tests now and read the actual output. This is the critical evidence gate.

Decision:

2 or more signals confirm implementation → offer the retroactive-close path (below)
1 signal → proceed with normal plan creation; note the signal in the first task's description
0 signals → proceed normally

Retroactive-close path:

Do NOT create a full task decomposition.

Create a single tracking task:

tk create "Retroactive verification: <SPEC-ID>" -t task --external-ref <SPEC-ID>

Claim it:
```
tk claim <id>
```
Run
```
verification-before-completion
```
(if superpowers installed) or re-run the spec's tests manually.
If verification passes: add a note with the evidence, close the task, then invoke swain-design to transition the spec to Complete.
If verification fails: fall back to normal plan creation — the prior implementation was incomplete.

Anti-rationalization safeguard (CRITICAL): The agent MUST re-run tests during detection — it cannot assume prior passing results are current. "Tests passed before" or "I implemented this earlier" is NOT evidence. Fresh test execution output is evidence.

Worktree isolation preamble

Before any implementation or execution operation (plan creation, task claim, code writing, execution handoff), run this detection:

bash

GIT_COMMON=$(git rev-parse --git-common-dir 2>/dev/null)
GIT_DIR=$(git rev-parse --git-dir 2>/dev/null)
[ "$GIT_COMMON" != "$GIT_DIR" ] && IN_WORKTREE=yes || IN_WORKTREE=no

Read-only operations (

tk ready

tk show

, status checks, task queries) skip this check entirely — proceed in the current context.

If
IN_WORKTREE=yes
: already isolated. Proceed normally.

If
IN_WORKTREE=no
(main worktree) and the operation is implementation or execution:

Use the
```
EnterWorktree
```
tool to create an isolated worktree. This is the only mechanism that actually changes the agent's working directory — manual
```
git worktree add
```
+
```
cd
```
does not persist across tool calls.

After entering, re-run tab naming to reflect the new branch:

bash

REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
bash "$REPO_ROOT/skills/swain-session/scripts/swain-tab-name.sh" --path "$(pwd)" --auto

If EnterWorktree
fails — stop. Surface the error to the operator. Do not begin implementation work.

Note: swain-session auto-enters a worktree at startup (Step 1.5), so this preamble is a fallback for sessions that skipped isolation or where the operator exited the worktree mid-session.

When all tasks in the plan complete, or when the operator requests, call

ExitWorktree

to return to the main checkout.

Fallback

tk

cannot be found or is unavailable:

Log the failure reason.
Fall back to a neutral text task ledger (JSONL or Markdown checklist) in the working directory.
Use the same status model (
```
open
```
,
```
in_progress
```
,
```
blocked
```
,
```
closed
```
) and keep updates externally visible.

swain-do

NPX Install

Tags

SKILL.md Content

Execution Tracking

Artifact handoff protocol

Term mapping

Configuration and bootstrap

Statuses

Operating rules

TDD enforcement

Spec lineage tagging

Escalation

"What's next?" flow

Context on claim

Artifact/tk reconciliation

Session bookmark

Superpowers skill chaining

Plan ingestion (superpowers integration)

Execution strategy

Pre-plan implementation detection

Worktree isolation preamble

Fallback