Salvage

/salvage

Extract learning from a session or piece of work before restarting. The insight: code is cheap now; learning is the asset.

Salvage is the bridge from Review back to Problem Space. When drift is detected, salvage captures what was learned so you can restart clean without losing understanding.

When to Use

Invoke /salvage when:

• Work is drifting - approach has changed direction multiple times
• Approach reversed 3+ times - you're oscillating, not converging
• Scope expanding while "done" keeps fuzzing - the finish line keeps moving
• You're protecting code you've invested in - afraid to throw it away
• Starting over feels right - but you don't want to lose what you learned

Do not use when: Work is on track and converging. Salvage is for extraction before restart, not routine reflection.

The Salvage Process

Step 1: Acknowledge the State

Before extracting, name what happened:

"This session/approach is being salvaged because [reason]. The original aim was [aim]. What actually happened was [reality]."

Be direct. No judgment—just clarity.

Step 2: Extract Five Things

Work through these extraction categories. Not everything will apply; extract what's present.

1. Model Shifts (What changed your understanding?)

• What assumptions were wrong?
• What did you learn about the problem that you didn't know before?
• What would you tell yourself at the start of this work?

"I thought X, but actually Y."

2. Guardrails (Constraints discovered the hard way)

• What boundaries should have been explicit from the start?
• What "don't do this" rules emerged?
• What edge cases bit you?

Format as explicit constraints:

Guardrail: [boundary]
Reason: [why this matters]
Trigger: [when to revisit this constraint]

3. Missing Context (What would have helped upfront?)

• What questions should have been asked at the start?
• What existing code/patterns should have been found first?
• What documentation was missing or would have prevented this?

"If I had known about [X], I would have [Y] instead."

4. Local Practices (Hard-won lessons worth encoding)

Local practices = practical wisdom, the kind you can only get from experience.

• What tribal knowledge did this work surface?
• What would help future-you (or future teammates) in similar situations?
• What patterns should be captured?

Good tribal knowledge is:

• Specific enough to be actionable
• General enough to apply beyond this exact case
• Non-obvious (not "write tests" but "this API silently returns 200 on auth failure")

5. What Worked (Don't lose the wins)

• What approaches or code fragments are worth keeping?
• What partial solutions could seed the restart?
• What tools or techniques proved useful?

Step 3: Package for Fresh Start

Synthesize the extraction into a restart kit:

## Salvage Summary

### Original Aim
[What we were trying to achieve]

### Why Salvaged
[Direct statement of what went wrong]

### Key Learnings
1. [Learning 1]
2. [Learning 2]
3. [Learning 3]

### New Guardrails
- [Guardrail 1]
- [Guardrail 2]

### Context for Restart
[What the next attempt should know before starting]

### Reusable Fragments
[Any code, patterns, or approaches worth keeping]

Step 4: Persist Learnings (if available)

If Open Horizons MCP is available:

1. Log to OH - Log tribal knowledge and guardrails to the graph
2. Update AGENTS.md - If learnings are project-wide, suggest additions

If no persistent storage is available, output the salvage summary for the user to capture manually.

Output Format

Always produce a salvage summary in this structure:

## Salvage Report

**Salvaged:** [date/session identifier]
**Reason:** [why this work is being salvaged]
**Original Aim:** [what we were trying to do]

### Learnings
[Numbered list of key insights]

### New Guardrails
[Explicit constraints with reason and trigger]

### Missing Context
[What would have helped]

### Local Practices
[Hard-won wisdom to encode]

### Reusable Fragments
[Code or patterns worth keeping]

### Fresh Start Recommendation
[How to approach this next time]

Examples

Example 1: Drifting Implementation

/salvage

The auth refactor has changed direction 4 times. Originally simplifying
the flow, now it's more complex than before.

Output:

## Salvage Report

**Salvaged:** Auth refactor session
**Reason:** Approach reversed 4 times, complexity increased not decreased
**Original Aim:** Simplify authentication flow

### Learnings
1. The complexity isn't in the auth code—it's in the 6 different OAuth providers
2. Each provider has subtly different token refresh behavior
3. The "simple" unified flow breaks provider-specific error handling

### New Guardrails
- Guardrail: Don't abstract OAuth providers into single interface
- Reason: Provider differences are semantic, not syntactic
- Trigger: Revisit if we drop to 2 providers or less

### Missing Context
- The original auth was complex for a reason (provider edge cases)
- Should have read the PR that introduced each provider

### Local Practices
OAuth providers aren't interchangeable. The abstraction leak is the feature.

### Fresh Start Recommendation
Instead of unifying the flow, document each provider's quirks explicitly.
Consider: is "simplify" the right aim? Maybe "make visible" is better.

Example 2: Scope Creep

/salvage

Started as "add dark mode toggle." Now touching 14 files and the
theme system doesn't exist yet.

Output:

## Salvage Report

**Salvaged:** Dark mode implementation
**Reason:** Scope expanded from toggle to theme system
**Original Aim:** Add dark mode toggle to settings

### Learnings
1. No theme system exists—colors are hardcoded across components
2. A toggle without infrastructure is meaningless
3. This is actually two tasks: (1) build theme system, (2) add toggle

### New Guardrails
- Guardrail: UI feature requests need infrastructure check first
- Reason: "Add X" often implies "build system for X"
- Trigger: Any request that touches visual consistency

### Missing Context
- Should have grepped for color usage before starting
- No design tokens or CSS variables in codebase

### Local Practices
"Add [feature]" is not the same as "build [feature]." Check if the
infrastructure exists before estimating.

### Reusable Fragments
- Color mapping I started: styles/colors.js (incomplete but useful)
- Component audit list: 14 files that hardcode colors

### Fresh Start Recommendation
1. First: Create theme system (CSS variables, ThemeContext)
2. Then: Add dark mode as first theme variant
3. The toggle is the easy part—do it last

Session Persistence

This skill can persist context to .oh/<session>.md for use by subsequent skills.

If session name provided (/salvage auth-refactor):

• Reads/writes .oh/auth-refactor.md directly

If no session name provided (/salvage):

• After producing the salvage report, offer to save it:

  "Save to session? [suggested-name] [custom] [skip]"

• Suggest a name based on git branch or the work being salvaged

Reading: Check for existing session file. Read everything—Aim, Problem Statement, Problem Space, Solution Space, Execute, Review—to understand what was attempted and what happened.

Writing: After producing the salvage report:

## Salvage
**Updated:** <timestamp>
**Outcome:** [extracted learnings, ready for restart]

[salvage report: learnings, guardrails, context for fresh start]

The salvage section is the capstone—it captures what was learned before the session ends or restarts.

Adaptive Enhancement

Base Skill (prompt only)

Works anywhere. Produces salvage summary for manual capture. No persistence.

With .oh/ session file

• Reads .oh/<session>.md for full session context — aim and guardrails frame what counts as "off-track"
• Writes salvage report to the session file; the report seeds the next session
• After salvage, offer to compact the session file: remove stale planning artifacts, keep settled decisions as brief anchors

With Open Horizons MCP

• Queries related past decisions before salvaging
• Logs learnings to graph database
• Creates dive pack for restart session
• Session file serves as local cache

With RNA MCP (repo-native-alignment)

Before extracting learnings: call search with the failure domain, include_artifacts: true, and the active phase. Ask: was this predictable from the corpus? Three outcomes change what gets written:

• Relevant entry existed and applied → failure was predictable. The learning is about process (corpus not consulted), not domain. Write that metis entry, not a domain one.
• Relevant entry existed but was insufficient → update or strengthen existing entries rather than creating near-duplicates.
• Nothing found → genuinely new territory. Write fresh metis with confidence.

After extracting:

• Write human-approved learnings to .oh/metis/ (new entries only — no duplicates)
• Write human-approved hard constraints to .oh/guardrails/
• outcome_progress to record what was accomplished toward the outcome before restarting

Meta-signal: If salvage repeatedly surfaces similar learnings, the corpus has the knowledge but it's not being consulted. That pattern warrants /distill.

Position in Framework

Comes after: /review (when drift is detected) or /execute (when you recognize thrashing). Leads to: /aim or /problem-space with fresh understanding—you climb back up. This is the feedback loop: Salvage is how learning flows back to the top of the framework.

Leads To

After salvage, typically:

• /aim - Reclarify what we're actually trying to achieve
• /problem-space - Fresh start with new understanding
• Create new task/issue with salvage context attached

Remember: Salvage is not failure. It's learning made explicit. The only failure is losing what you learned.