Stacker

stacker

Skills overlap. Run a minimal eval over your installed skills and you find the uncomfortable truth: no tool wins on every axis. One scraper is faster, another renders JavaScript, a third survives bot walls. One reviewer catches security holes, another catches naming. The deltas are small and axis-specific.

Agents handle this badly in two symmetric ways:

• Load everything. Every overlapping skill goes into context "just in case." Context bloats, instructions conflict, the agent improvises.
• Pick "the best." A router chooses one winner and throws away every axis the losers were better at.

Stacker is the third option: run the top candidates on the same frozen task, score their outputs, merge them. Ensembling, applied to agent skills. The unit you combine is the output, never the context.

When to stack (and when not to)

Stacking costs tokens, time, and sometimes money. It is the paid upgrade, not the default. Stack only when ALL of these hold:

1. Output-critical — correctness or completeness matters more than latency and cost.
2. Real overlap — 2+ installed skills can each do the task end to end.
3. Complementary failures — evals or experience show they fail on different inputs.

Signal	Action
Routine task, one obvious tool	Route (single tool, escalate on failure)
One tool strictly dominates every axis	Route
Output feeds a decision, a publish, or money	Stack
Tools fail differently (JS vs static, bot walls, rate limits)	Stack
User says "double-check", "make sure", "cross-check"	Stack

If you only need a fallback ladder ("try cheap, escalate on failure"), that is routing, not stacking — keep the router skill you already have.

The stack workflow

1. Enumerate candidates. List the installed skills/tools that can genuinely complete this task end to end. Cap at 2–4. If you can't name what each one is better at, drop it.

2. Freeze the task spec. One paragraph: input, expected output shape, success criteria. Every candidate gets the identical spec — otherwise you're comparing answers to different questions.

3. Fan out in parallel. One isolated run per candidate (subagent, background job, or separate CLI invocation). Each branch loads ONLY its own skill's instructions. Nothing but the branch's output returns to the main context.

4. Score each output against the success criteria from step 2. Simple 3-axis rubric, 0–2 each:

   • Complete — covers everything the spec asked for?
   • Correct — spot-check 2–3 verifiable facts/behaviors?
   • Clean — usable as-is (format, noise, truncation)?

5. Merge with the strategy that fits the output type (table below).

6. Report with provenance. Deliver the stacked result, then one line per candidate: what it contributed, where it disagreed, its score. Disagreements between tools are signal — surface them, never average them away.

Merge strategies

Strategy	When	How
best-of	Outputs are interchangeable, one clearly scored highest	Take the winner, discard the rest
union	Outputs are lists/collections (search results, findings, extracted items)	Combine, dedupe by identity, keep the best-sourced duplicate
patch	One output is the best base but has gaps others fill	Take the winner, splice in the specific sections losers did better
consensus	Outputs assert facts that must be right	Keep facts 2+ tools agree on; flag single-source facts as unverified
synthesis	Outputs are prose/analysis with different strengths	Rewrite into one artifact, crediting which tool sourced which part

Default: union for collections, patch for documents, consensus for facts.

Common overlap families

Use these to enumerate candidates fast (step 1) — most skill collections grow the same overlaps:

family	typical candidates	default merge
Web access	scraper CLI, local browser, cloud browser, native fetch	patch
UI/UX review	design guidelines, frontend-ui engineering, a11y, web-perf	synthesis
Code review (TS/Rust/Python…)	language reviewer, generic code-quality, security reviewer	union + consensus severity
Test generation	TDD, e2e, regression-testing skills	union
Memory / context recall	session memory, knowledge graph, learned instincts	union, newest wins conflicts
Docs / research	web search, deep-research, vendor-docs lookup	consensus facts, synthesis prose

Worked example — web scraping

Task: extract full product data from a JS-heavy page, feeding a pricing decision (output-critical → stack).

Candidates:  firecrawl scrape (fast, clean markdown, fails on some SPAs)
             agent-browser   (renders JS reliably, noisier output)

Fan out:     firecrawl scrape <url>            → branch A
             agent-browser open <url> &&
             agent-browser get text body       → branch B

Score:       A: complete 1 (missing lazy-loaded specs), correct 2, clean 2
             B: complete 2, correct 2, clean 1 (nav noise)

Merge:       patch — A's clean markdown as base, B's rendered spec table
             spliced in.

Report:      stacked result + "specs table came from agent-browser only;
             prices identical in both (consensus ✓)"

Either tool alone ships a defect: A misses data, B ships noise. The stack ships neither.

Context discipline (the whole point)

The problem stacker exists to solve is agents loading every overlapping skill into one context. So:

• Never load candidate skills' full instructions into the main context. Each fan-out branch loads its own skill; the main context sees only outputs.
• The main context holds: the task spec, the candidate list, the scores, the merged result. That's it.
• 2–4 candidates max. A 6-tool stack is context bloat wearing a costume.

Persisting a stack — scripts/stack.sh

A stack you build more than twice deserves to become a skill of its own: ONE stacked skill that owns the routing, with the source skills unloaded from the agent's discovery path so they stop costing context every session. scripts/stack.sh does the mechanical half:

# unload sources from the discovery path, re-link them under the stacked skill
bash scripts/stack.sh stash --into ~/.agents/skills/web-access \
  firecrawl firecrawl-scrape firecrawl-search agent-browser

bash scripts/stack.sh status  --from ~/.agents/skills/web-access
bash scripts/stack.sh restore --from ~/.agents/skills/web-access firecrawl  # or no names = all

Nothing is deleted:

• A discovery entry that is a symlink (the skills-CLI norm) is unlinked; the canonical dir (e.g. ~/.agents/skills/<name>) stays put and keeps receiving upstream updates. The stacked skill reaches it via a new symlink at sources/<name> — updates flow through the link, no re-pull, no vendoring.
• A real directory is moved (not copied) to sources/<name>.
• sources/.manifest.tsv records every action, so restore is an exact undo.

Agent-managed skills are NEVER touched. Skills that ship with the host agent — Claude Code plugin skills (claude-in-chrome), Gemini CLI extensions, Cursor extensions, opencode plugins, Codex installs — belong to that agent's own installation. stash detects which agents are present (binary on PATH or config root on disk) and refuses to move anything inside their roots, whatever CLI is running the stack. Those still belong in the stacked skill's Routing table as rows — they cost no discovery entry, so there is nothing to stash. Extend the protected list with STACKER_PROTECTED=dir1:dir2.

Updating a stack needs no command: sources update through their symlinks the moment upstream does (npx skills update, git pull). To audit or grow a stack:

bash scripts/stack.sh doctor --from <stacked-dir>   # detected agents, broken links, upstream freshness
bash scripts/stack.sh stash  --into <stacked-dir> <new-source>   # idempotent — re-run to add sources

Discovery paths default to ~/.claude/skills; override with STACKER_AGENT_DIRS=dir1:dir2.

What a generated stacked skill MUST declare

After stashing, the stacked skill is the only one the agent loads — so it must carry what the stashed skills used to announce:

1. Routing — a decision table mapping task → source tool, plus the escalation path. This is what replaces "load them all".
2. Sources — where every candidate came from and how it stays fresh:

## Sources (stacked by stacker)

Unloaded from the discovery path; reachable via `sources/` symlinks.
Upstream updates flow through the links — no re-pull.

| source | reachable at | canonical |
|---|---|---|
| firecrawl     | sources/firecrawl     | ~/.agents/skills/firecrawl |
| agent-browser | sources/agent-browser | ~/.agents/skills/agent-browser |

Restore any source: `stack.sh restore --from <this dir> <name>`

A fan-out branch that needs a source's full instructions reads sources/<name>/SKILL.md — into ITS context, never the main one.

Anti-patterns

• Stacking cheap tasks. A weather lookup does not need an ensemble.
• Stacking dominated tools. If tool A beats tool B on every axis you care about, B adds cost and zero coverage.
• Merging by vibes. No frozen spec, no scores → you're just picking the longest output.
• Hiding disagreement. If two tools return different prices, that's a finding, not a rounding problem.