Retrospectives

Before proceeding, check for an active session and determine which evidence source to use:

REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
session_check=$(bash "$REPO_ROOT/.agents/bin/swain-session-check.sh" 2>/dev/null)
session_status=$(echo "$session_check" | jq -r '.status // "none"')

if [ "$session_status" = "active" ]; then
  echo "Using active session evidence"
else
  echo "No active session — falling back to git log evidence"
  git_evidence=$(git log --oneline -20 2>/dev/null)
  echo "Recent commits:"
  echo "$git_evidence"
fi

If no active session, proceed with the retro using git log as the primary evidence source. Produce a standalone retro document in docs/swain-retro/ rather than embedding in an artifact. Adapt reflection questions to the unscoped context.

Captures learnings at natural completion points and persists them for future use. This skill is both auto-triggered (EPIC terminal transition hook in swain-design) and manually invocable via /swain-retro.

Output modes

Scope	Output	Rationale
EPIC-scoped (auto or explicit)	## Retrospective section appended to the EPIC artifact	The EPIC already contains lifecycle, success criteria, and child specs — it's the single source of truth for "what we shipped and what we learned"
Cross-epic / time-based (manual)	Standalone retro doc in docs/swain-retro/	No single artifact owns the scope — a dedicated doc is required

Invocation modes

Mode	Trigger	Context source	Output	Interactive?
Auto	EPIC transitions to terminal state (called by swain-design)	The EPIC and its child artifacts	Embedded in EPIC	No — fully automated
Interactive	EPIC transitions to terminal state during a live session	The EPIC and its child artifacts	Embedded in EPIC	Yes — reflection questions offered
Manual	User runs /swain-retro or /swain retro	Recent work — git log, closed tasks, transitioned artifacts	Standalone retro doc (required)	Yes
Scoped	/swain-retro EPIC-NNN or /swain-retro SPEC-NNN	Specific artifact and its related work	Embedded in EPIC (if EPIC-scoped) or standalone	Yes

Terminal states that trigger auto-retro: Complete, Abandoned, Superseded. The retro content adapts to the terminal state — an Abandoned EPIC's retro focuses on why work stopped and what was learned, not on success criteria.

Step 1 — Gather context

Collect evidence of what happened during the work period.

For EPIC-scoped retros (auto or scoped)

REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
# Get the EPIC and its children
bash "$REPO_ROOT/.agents/bin/chart.sh" deps <EPIC-ID>

# Session log — the primary evidence source for retros (ADR-015)
# Contains decisions, pivots, rationale, and operator feedback
cat .agents/session.json 2>/dev/null | grep -i "<EPIC-ID>\|<SPEC-IDs>"

Also read:

• The EPIC's lifecycle table (dates, duration)
• Child SPECs' verification tables (what was proven)
• Any ADRs created during the work
• Git log for commits between EPIC activation and completion dates

Note (ADR-015): Do not use tickets (tk / .tickets/) as retro evidence. Tickets are ephemeral execution scaffolding — they record task status, not decisions or rationale. The session log (.agents/session.json JSONL) captures the actual conversation: what was tried, what pivoted, why, and what the operator said. Build the retro narrative from session logs and git history.

For manual (unscoped) retros

REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
# Recent git activity
git log --oneline --since="1 week ago" --no-merges

# Session log — primary evidence source (ADR-015)
cat .agents/session.json 2>/dev/null | tail -100

# Recently transitioned artifacts
bash "$REPO_ROOT/.agents/bin/chart.sh" status 2>/dev/null

Also check:

• Existing memory files for context on prior patterns
• Previous retro docs in docs/swain-retro/ for recurring themes

Step 2 — Generate or prompt reflection

Auto mode (non-interactive)

When invoked by swain-design during a non-interactive EPIC terminal transition (e.g., dispatched agent, batch processing), generate the retro content automatically from the gathered context:

1. Synthesize what was accomplished, what changed from the original scope, and what patterns are visible in the commit/task history
2. For Abandoned or Superseded EPICs, focus on why the work stopped and what was learned
3. Proceed directly to Step 3 (memory extraction) and Step 4 (write output)

Interactive mode

When the user is present (live session, manual invocation), present a summary and offer reflection:

Summary format

Retro scope: {EPIC-NNN title / "recent work"} Period: {start date} — {end date} Artifacts completed: {list} Tasks closed: {count} Key commits: {notable commits}

Reflection questions

Ask these one at a time, waiting for user response between each:

1. What went well? What patterns or approaches worked effectively that we should repeat?
2. What was surprising? Anything unexpected — blockers, shortcuts, scope changes?
3. What would you change? If you could redo this work, what would you do differently?
4. What patterns emerged? Any recurring themes across tasks — tooling friction, design gaps, communication patterns?

Adapt follow-up questions based on user responses. If the user gives brief answers, probe deeper. If they're expansive, move on.

Step 3 — Distill learnings

After the reflection conversation, persist the learnings — but where they go depends on whether this is the swain project itself or a consumer project that uses swain.

Detect context

# Check if the current repo IS swain (the tool itself)
REPO_ROOT=$(git rev-parse --show-toplevel 2>/dev/null || pwd)
if git remote get-url origin 2>/dev/null | grep -q "cristoslc/swain"; then
  echo "SWAIN_REPO"
else
  echo "CONSUMER_PROJECT"
fi

In consumer projects: write to memory

Learnings from consumer projects go into Claude memory files because they represent operator preferences and project-specific patterns that should persist across sessions.

Feedback memories

For behavioral patterns and process learnings that should guide future agent behavior:

---
name: retro-{topic}
description: {one-line description of the learning}
type: feedback
---

{The pattern or rule}

**Why:** {User's explanation from the retro}
**How to apply:** {When this guidance kicks in}

Write to the project memory directory:

~/.claude/projects/<project-slug>/memory/feedback_retro_{topic}.md

The project slug is the project path with slashes replaced by dashes (e.g., /Users/cristos/Documents/code/myapp → -Users-cristos-Documents-code-myapp). These files live in Claude's memory system (not swain's .agents/ state), which is intentional — retro learnings persist across all Claude Code sessions for this project.

Update MEMORY.md index.

Project memories

For context about ongoing work patterns, team dynamics, or project-specific learnings:

---
name: retro-{topic}
description: {one-line description}
type: project
---

{The fact or observation}

**Why:** {Context from the retro}
**How to apply:** {How this shapes future suggestions}

Rules for memory creation

• Only create memories the user has explicitly validated during the reflection
• Merge with existing memories when the learning extends a prior pattern
• Use absolute dates (from the retro context), not relative
• Maximum 3-5 memory files per retro — distill, don't dump

In the swain repo: propose artifacts, not memory

When the retro is running inside swain itself, learnings that imply behavioral changes, new capabilities, or process fixes are not memory — they are work items. Swain's behavior is encoded in skills, specs, and ADRs, not in operator memory files.

Classify each learning and route it to the right output:

Learning type	Route	Example
A skill needs new behavior or guardrails	SPEC candidate — add to the retro doc's "SPEC candidates" section	"swain-release should refuse to run in worktrees"
A cross-cutting process change	ADR candidate — note in retro doc	"Dispatched agents should never use reset --hard"
A new capability or initiative	EPIC candidate — note in retro doc	"Session logs need command-level forensics"
A bug in an existing skill	GitHub issue via gh issue create	"swain-sync doesn't preserve symlinks"
Operator preference or project context	Memory (same as consumer)	"Operator prefers bundled PRs for refactors"

For artifact candidates, add a ## SPEC candidates section (or ## ADR candidates, ## EPIC candidates) to the retro doc output. Each entry should include enough context to draft the artifact later:

## SPEC candidates

1. **{Title}** — {one-line description of what needs to change and why}
2. ...

Do not create the specs/epics/ADRs during the retro — that is swain-design's job. The retro captures what was learned; the operator decides what to build next. Only write to memory for learnings that are genuinely about operator preferences or project context (the last row in the table above).

Step 4 — Write output

Output destination depends on scope — see Output modes at the top.

EPIC-scoped: embed in the EPIC artifact

Append a ## Retrospective section to the EPIC markdown file, before the ## Lifecycle table. This keeps the EPIC as the single source of truth.

## Retrospective

**Terminal state:** {Complete | Abandoned | Superseded}
**Period:** {activation date} — {terminal date}
**Related artifacts:** {SPEC-NNN}, {SPEC-NNN}, ...

### Summary

{What was accomplished — or for Abandoned/Superseded, what was learned and why work stopped}

### Reflection

{Synthesized findings — from auto-generation or interactive Q&A}

### Learnings captured

<!-- In consumer projects, this table lists memory files. In the swain repo, it lists
     artifact candidates (SPECs, ADRs, issues) plus any memory files for operator preferences. -->

| Item | Type | Summary |
|------|------|---------|
| ... | memory / SPEC candidate / issue | ... |

Hyperlink the artifact IDs in Related artifacts using Step 4.5.

Cross-epic / time-based: standalone retro doc (required)

For manual retros not scoped to a single EPIC, a standalone doc is required — no single artifact owns the scope.

mkdir -p docs/swain-retro

File: docs/swain-retro/YYYY-MM-DD-{topic-slug}.md

---
title: "Retro: {title}"
artifact: RETRO-{YYYY-MM-DD}-{topic-slug}
track: standing
status: Active
created: {YYYY-MM-DD}
last-updated: {YYYY-MM-DD}
scope: "{description of what's covered}"
period: "{start} — {end}"
linked-artifacts:
  - {ARTIFACT-ID-1}
  - {ARTIFACT-ID-2}
---

# Retro: {title}

## Summary

{Brief description of what was completed across the scope}

## Artifacts

| Artifact | Title | Outcome |
|----------|-------|---------|
| ... | ... | Complete/Abandoned/... |

## Reflection

### What went well
{User's responses, synthesized}

### What was surprising
{User's responses, synthesized}

### What would change
{User's responses, synthesized}

### Patterns observed
{User's responses, synthesized}

## Learnings captured

<!-- In consumer projects, this table lists memory files. In the swain repo, it lists
     artifact candidates (SPECs, ADRs, issues) plus any memory files for operator preferences. -->

| Item | Type | Summary |
|------|------|---------|
| ... | memory / SPEC candidate / issue | ... |

Step 4.5 — Hyperlink artifact references

After writing the retro output (standalone doc or embedded EPIC section), scan all body text for bare artifact ID references matching (SPEC|EPIC|INITIATIVE|VISION|SPIKE|ADR|PERSONA|RUNBOOK|DESIGN|JOURNEY|TRAIN)-[0-9]+. For each bare ID not already inside a markdown link or code fence, resolve and replace:

REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
bash "$REPO_ROOT/.agents/bin/resolve-artifact-link.sh" <ARTIFACT-ID> <RETRO-FILE>

Replace bare IDs with [ARTIFACT-ID](relative-path). If the script returns non-zero or empty output (artifact not found), leave the bare ID as-is. Frontmatter related-artifacts values stay as plain IDs (YAML compatibility).

Context-rich artifact references

When referencing artifacts in retro output (child specs, related artifacts, linked work), use context lines:

REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
bash "$REPO_ROOT/.agents/bin/artifact-context.sh" <ID> 2>/dev/null

Fall back to <ID> — <title> if the utility is unavailable.

Step 4.7 — README drift check (SPEC-210)

After reflection and before closing the retro, check if the README still matches the project.

REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
[ -f "$REPO_ROOT/README.md" ] && echo "has_readme" || echo "no_readme"

If README.md exists:

1. Read README.md and extract claims about what the project does and how it works.
2. Compare claims against artifacts that changed during the retro scope (from Step 1 context).
3. Surface drift findings:
   • New features the README omits — the epic shipped something the README does not mention.
   • Stale promises — the epic dropped or replaced something the README still claims.
   • Changed behavior — the epic changed how something works but the README still shows the old way.

Show findings to the operator. They can fix the README, fix the artifact, or defer.

In auto mode, add drift findings to the ## Reflection section. In interactive mode, show them after the reflection questions and before writing output.

If no drift exists, skip — do not add a "no drift" note.

Deferred findings go in a ### README drift subsection of the retro output so they stay visible.

Step 5 — Update session bookmark

REPO_ROOT="$(git rev-parse --show-toplevel 2>/dev/null || pwd)"
bash "$REPO_ROOT/.agents/bin/swain-bookmark.sh" "Completed retro for {scope} — {N} learnings captured"

Integration with swain-design

swain-design orchestrates this skill when an EPIC transitions to any terminal state (Complete, Abandoned, Superseded):

1. swain-design completes the phase transition (move, status update, commit, hash stamp)
2. swain-design invokes swain-retro with the EPIC ID and terminal state
3. swain-retro gathers context, generates/prompts reflection, extracts memories, and embeds the ## Retrospective section in the EPIC
4. swain-design commits the retro content as part of (or immediately after) the transition

Interactive detection: If the session is interactive (user is present and responding), swain-retro offers the reflection questions. If non-interactive (dispatched agent, batch), it runs fully automated.

This is best-effort — if swain-retro is not available, the EPIC transition still succeeds without a retro section.

Referencing prior retros

When running a new retro, scan both EPIC artifacts (grep for ## Retrospective sections) and docs/swain-retro/ for prior retros. If patterns recur across multiple retros, call them out explicitly — recurring themes are the most valuable learnings.

# Check standalone retro docs
ls docs/swain-retro/*.md 2>/dev/null | head -10

# Check embedded retros in EPICs
grep -rl "## Retrospective" docs/epic/ 2>/dev/null | head -10