raga-cli 0.0.0-stage → 0.9.0

package/bundled-skills/skills/peak/playbooks/runtime-forensics.md

@@ -0,0 +1,11 @@
+### Runtime forensics
+
+**You own the diagnosis. Instrument the live process, don't theorize from source.** The deliverable is a cited diagnosis, not a fix.
+
+1. Capture the live signal on the matching surface via the control skill: a CPU profile for a spinning process, a heap snapshot for a leak, a CDP trace for a visual glitch. A real artifact, not a guess.
+2. Reduce the artifact to the smoking gun: the function on the hot path, the retainer chain from the leaked object to a GC root, the loop firing without input. Parse large artifacts in a subagent (the **guard-the-context-window** principle skill), keep the reduced finding in the main thread. Pick each subagent's reasoning level and fast or standard speed for the job, per the Subagents section.
+3. Prove the mechanism before believing it. Inject instrumentation via CDP eval on the running process, or hotfix the live code without reloading, to confirm the hypothesis cheaply.
+4. Map the finding back to source: file, symbol, the line that allocates or schedules.
+5. Throughput checkpoint stays one line: `throughput checkpoint: n/a, read-only forensics`.
+
+**Reply:** the signal captured, the reduced finding, how you proved the mechanism, the source location, artifact paths. No fix unless asked. Hand back to Bug fix or Perf once the cause is known.

package/bundled-skills/skills/peak/playbooks/session-pickup.md

@@ -0,0 +1,11 @@
+### Session pickup
+
+**You own the resume point. Read the prior trail, don't redo it.**
+
+1. Locate the prior trail in the supplied state capsule, await context.history(), or project-scoped await memory.search(query). Read full entries with await memory.get(id). Earlier transcripts require an explicit supervisor-provided export. Preserve the original evidence and distinguish missing history from an empty result.
+2. Reconstruct operational state. The branch and worktree, what already landed (`git log`, `git diff` against the base), the open todos, the decisions made. The prior trail is authoritative input. Resist the bias to re-derive it.
+3. Diff done vs pending. Compare what shipped against what was planned, name the resume point, do not re-run the prior repro or redo completed work. A "let me verify from scratch" pass means you're treating the trail as untrustworthy when it's authoritative.
+4. Route the remaining work to the matching playbook and pick the verdict: continue the execution, ship a finished recommendation, ratify or override a prior conclusion, or postmortem a failed run. The pickup playbook ends here. The routed playbook owns the rest.
+5. Verify the inherited claims against the original goal on the real artifact (the **principle-prove-it-works** skill). A passing prior self-report is not the proof.
+
+**Reply:** where the prior agent stopped, what you inherited vs redid (ideally nothing redone), the resume point, and the outcome.

package/bundled-skills/skills/peak/playbooks/shipping.md

@@ -0,0 +1,12 @@
+### Shipping
+
+**You own what lands. Verify each PR independently, land only the verified run from the root, then keep your hands off the stack.**
+
+1. **Resolve the forge, then verify each PR independently before it lands.** GitHub CLI (`gh`) is the default. If `command -v origin` succeeds and Origin can resolve the repository, use `origin pr` for every PR operation. Otherwise stay on `gh` and record the fallback. Never require Graphite (`gt`). One verifier per PR, each a isolated raga child, runs the change the way a user would with the matching control skill (such as `control-ui` or `control-cli` from `cursor-team-kit`), parent against head, and posts `PASS`, `PASS+NOTES`, or `FAIL` on that PR. Only a verdict from an agent that did not write the code counts, because an author checks what it meant to build. Green CI and an approving bot review are not verdicts, because neither runs the change as a user would.
+2. **Land only the contiguous verified run rooted at the bottom.** Walk up from the lowest unmerged PR and stop at the first one without `PASS` or `PASS+NOTES`. The PRs below it are the run, and the top one is the ceiling. A verified PR above an unverified one waits, because it cannot land without the code below it. One whose branch neither holds nor needs that code can move onto trunk and land on its own. It needs a fresh verdict first, because its old one ran with that code beneath it and trunk does not have it.
+3. **Re-check that the verdicts still describe the code.** A verdict covers a patch, and a chain rebase moves every PR above it to a new SHA without failing any check. Compare `git patch-id` at the verdict SHA with the current head before trusting an older verdict. When the patches differ only in tests, docs, or lint config, a lane's result stands if what it ran builds the same at both SHAs apart from noise. Noise is a difference that also shows up between two builds of the verdict SHA, or an embedded commit SHA. Report each kind of noise with its files, so a reader can check the call. The change itself still gets fresh checks and a review, because the old verdict never saw it. A lane with no build output, such as a dev server, cannot show this, so rerun it, and re-verify anything else that drifted.
+4. **Land bottom-up, one PR at a time.** Always merge the run's lowest unmerged PR, with `origin pr merge <pr>` or `gh pr merge <pr> --squash`. On Origin the plain command also merges every PR below the one you name, and on GitHub a PR whose base is still its parent's branch merges into that branch, not trunk. Read the PR back for the result (`origin pr view <pr>` or `gh pr view <pr> --json state,mergedAt,baseRefName`), because a merge command can report success when a merge queue only queued the PR, and a requested auto-merge is not a merge. A landed PR reads merged with trunk as its base, not its parent's branch. Arm `--auto` (Origin merge-when-ready or GitHub auto-merge) only on a PR whose verdict covers its current head, because it lands whatever head is ready. After a merge lands, retarget the PR whose base was the merged branch to trunk (`origin pr edit <pr> --base <trunk>` or `gh pr edit <pr> --base <trunk>`), move its patch onto the merge result, and push with `--force-with-lease`. The verdict covers the patch, not its commits, so step 3's patch-id rule decides whether the verdict survives. Merge it once its checks pass and its verdict survives.
+5. **Leave the stack alone while it lands.** Step 4's move of the next PR is the only change to a branch or base, because a merge lands whatever its branch and base hold at that moment, verdict or not. On GitHub, watch the run with the queued watcher from `playbooks/babysit.md` (`scripts/watch-pr/watch-pr --queued-stack --stack-prs <bottom-to-top>`) under the in-process bounded loop until it reports `COMPLETE` at the ceiling. On Origin, re-read the current PR with `origin pr view <pr> --checks` each time `origin pr checks <pr> --watch` returns, until the ceiling reads merged. A base that moves to trunk when its parent merges is expected, not damage to repair. If a merge stalls, find out why before you change anything.
+6. **Stop at the ceiling.** Extending the run takes new verdicts, so it is a new pass through step 1.
+
+**Reply:** the verified run and its ceiling, each PR's verdict and who produced it, what landed and how you confirmed it, and what the next gap needs.

package/bundled-skills/skills/peak/playbooks/trace-forensics.md

@@ -0,0 +1,14 @@
+### Trace forensics
+
+**You own the diagnosis from the artifact.**
+
+Distinct from **Runtime forensics**, which instruments the live process. Here the capture already exists. The artifact is a fixed dataset, read it, don't re-run it. Keep tooling generic: a DevTools or trace parser for cpuprofile and `.json.gz`, a text editor for a spindump, your heap tooling for a heapsnapshot.
+
+1. Identify the format and load it with the right tool. Parse large artifacts in a subagent (the **principle-guard-the-context-window** skill) and keep the reduced finding in the main thread. Pick each subagent's reasoning level and fast or standard speed for the job, per the Subagents section.
+2. Transform the raw artifact into a form you can query. Dump the trace or heap snapshot into sqlite, one row per sample, frame, or node. Reach the queryable shape before you read.
+3. Narrow to the cause. Query for the frames that hold the most time and walk the call tree to the hot path. For a leak, follow the retainer chain from the leaked object to a GC root. For a spindump, find the thread stuck on-CPU or blocked and its wait reason.
+4. Attribute to source. Map the hot frame to file, symbol, and line via the artifact's own symbols. A frame with no source mapping is not yet a diagnosis. Resolve the symbols, or say plainly the artifact does not carry them.
+5. If you have a paired capture, diff the before and after artifacts to confirm. Otherwise, mark the finding as the strongest hypothesis the artifact supports, not a confirmed cause.
+6. Hand back a cited diagnosis, no fix unless asked. Route to Bug fix or Perf issue once the cause is known. Throughput checkpoint stays one line: `throughput checkpoint: n/a, read-only forensics`.
+
+**Reply:** the artifact and format, the reduced finding, the source location, the artifact paths, and whether a paired capture confirmed it.

package/bundled-skills/skills/peak/playbooks/visual-parity.md

@@ -0,0 +1,11 @@
+### Visual parity
+
+**You own pixel-exact equivalence. The baseline is the spec. You do not touch it.** Equivalence is verified by image diff, not by eye.
+
+1. Establish the baseline first, before any migration: a visual regression harness that screenshots the current component across its states, plus the target when matching two implementations. No baseline, no parity claim. A blocking prerequisite, not a follow-up.
+2. Anti-shortcut clauses, stated and held: no harness modifications, no baseline tampering, no component restructuring to make a diff pass. If the baseline looks wrong, stop and ask, don't edit it.
+3. Migrate one component at a time. Parallelize across worktrees, one owner per component (the **separate-before-serializing-shared-state** principle skill). Shared primitives migrate first as a blocking phase.
+4. Verify each component against its baseline via image diff on the matching surface via the control skill. A nonzero diff is a fail. Investigate the pixel delta. the in-process bounded loop per component until the diff is zero.
+5. Run **Opening a PR** per component or per safe batch.
+
+**Reply:** components migrated, the diff result for each, the baseline harness location, what's left.

package/bundled-skills/skills/peak/playbooks/worktree-cleanup.md

@@ -0,0 +1,6 @@
+### Worktree cleanup
+
+1. Read the current instance's supervisor-owned workspace and session records. Run git worktree list --porcelain in the bound application repository. Match each path and base commit to its owner. Unknown ownership means keep it.
+2. Inspect dirty files, unpushed commits, accepted revisions, active jobs, and process liveness. A stopped chat or a merged branch alone is not proof that a descendant workspace is unused.
+3. Ask the supervisor to remove only its own stopped child workspace through the managed lifecycle. Never run git worktree remove --force or delete a user's unknown directory.
+4. Report what was removed and what remains, with owner and reason. Destructive removal of user data needs explicit operator approval.

package/bundled-skills/skills/peak/references/architect/SKILL.md

@@ -0,0 +1,81 @@
+---
+name: architect
+description: "Sketch types, signatures, and module structure before code, then stay in the loop while implementation fills in. Use for /architect, 'architect this', 'design this', or non-trivial work where jumping to code would lock in the wrong shape."
+disable-model-invocation: true
+---
+
+Read [the native Python adapter](../raga-runtime/SKILL.md) before executing this workflow. Authoring targets candidate revisions. Children use run_child, not an external agent service.
+
+# Architect
+
+Sketch types, function signatures, class shapes, and module boundaries with `not implemented` bodies and pseudocode. Synthesize across multiple model perspectives, then fill in code against the chosen sketch.
+
+## Start
+
+Open a todolist with one entry per phase before starting.
+
+1. Ground
+2. Sketch
+3. Agree
+4. Implement
+5. Scrap
+
+## Phase A: Ground the problem
+
+Build a real mental model of every system the new code touches. Run the **how** skill over the relevant subsystems. Naming a file isn't grounding. Produce the architectural explanation `how` prescribes. If the design redefines ownership or layering, also run the **why** skill on the existing shape so the rationale becomes a constraint, not a guess.
+
+Skip Phase A only when the work is genuinely greenfield with no surrounding system to integrate.
+
+## Phase B: Sketch
+
+Run the **arena** skill with the design-sketch task and the Phase A grounding artifacts. Pass `references/runner-prompt.md` as each runner's prompt.
+
+Take the runners from the `architect runners` line in the `model-roles.json` rule, in place of the `arena runners` line. If no role is configured, inherit the parent model and report that the seats are not model-diverse. Alias and rejected entries follow the runner rules in the **arena** skill's Phase A.
+
+Design it twice. Require at least two structurally distinct candidates before synthesis, even when the first looks sufficient. This is the **exhaust-the-design-space** principle skill made concrete. Whole-shape alternatives, not point fixes inside one shape.
+
+Screen every candidate against [`references/design-red-flags.md`](references/design-red-flags.md) before synthesis. Assume the next contributor is an agent that sees only the files it opened, copies the nearest example, and takes the shortest path that compiles. Prefer the design where a change that looks right from one file is right for the whole repo.
+
+Compare viable candidates on interface depth. Prefer the design that hides more complexity behind a smaller, simpler public surface. A rich interface can keep call chains short by concentrating capability instead of scattering it across layers.
+
+Arena returns one synthesized design package. Its synthesis decision fills the rationale's "Synthesis decision" section.
+
+## Phase C: Agree (opt-in)
+
+Default: proceed directly to implementation with the synthesized design. No human checkpoint.
+
+Opt in to a checkpoint when the invoker explicitly asks: "/architect with checkpoint," "stop and show me before implementing," or similar. Then surface the synthesized design and pause for sign-off.
+
+The synthesis can ship as its own commit either way, as the "scaffold first" mode of the **foundational-thinking** principle skill. Planned and scoped breakage during fill-in is fine, per the **outcome-oriented-execution** principle skill. For adversarial pressure on the design before implementing, run the **interrogate** skill on the synthesized sketch.
+
+If the human pushes back on the shape (in a checkpoint or after the fact), treat that as Phase A evidence. Re-ground and re-run Phase B before writing more code.
+
+## Phase D: Implement against the sketch
+
+Replace `not implemented` bodies with code, pseudocode with logic. The synthesized sketch is the contract. Deviations from it are signal to surface, not friction to absorb silently. If a function needs a parameter the sketch didn't anticipate, ask whether the sketch was wrong, the requirement was missed, or the implementation is overreaching.
+
+## Phase E: Scrap when the architecture is wrong
+
+If implementation keeps producing friction the sketch can't absorb, throw the sketch out. Don't bolt fixes onto a wrong design, per the **redesign-from-first-principles** and **fix-root-causes** principle skills.
+
+The signal is a *pattern*, not single instances. Tells:
+
+- The same shape of workaround appearing repeatedly across unrelated code.
+- Multiple unrelated edge cases that all need special-case branches.
+- Types that need escape hatches (`any`, casts, optional fields always set in practice) to compile.
+- The "we need a lock" reflex when the sketch said the state wasn't shared.
+- Callers having to know the abstraction's internal rules to use it.
+- Two or more independent Phase D deviations of the same shape across the implementation.
+
+Use judgment. A few edge cases don't condemn an architecture. Some problems are legitimately complex. Complexity in the data is not complexity in the design.
+
+When you scrap:
+
+1. Re-run the **how** skill over what's been built.
+2. Redesign as if the new constraints had been day-one assumptions, per redesign-from-first-principles.
+3. Subtract before adding, per the **subtract-before-you-add** principle skill. The new sketch should be smaller than the old one before it grows.
+4. Return to Phase B and re-run arena.
+
+## Outputs
+
+The caller's usage is written first and the type sketch derived from it. One file with new types and signatures for small changes. Module map plus type definitions for larger work. The rationale ships alongside, shaped per `references/rationale-template.md`, including the usage sketch and the synthesis decision.

package/bundled-skills/skills/peak/references/architect/references/design-red-flags.md

@@ -0,0 +1,57 @@
+# Design red flags
+
+Screen every candidate before synthesis. A red flag is a reason to revise or reject the shape.
+
+## Shallow module
+
+A shallow module exposes a large interface while hiding little complexity. Judge depth by the capability and policy hidden behind the public surface relative to the size of that surface. Prefer a simple interface backed by substantial behavior.
+
+Do not confuse a deep module with a deep call chain. A deep call chain scatters understanding across layers. A deep module concentrates capability behind one interface.
+
+Look for these signs:
+
+- Callers coordinate several methods to complete one operation.
+- Public options expose internal stages or implementation choices.
+- Learning the interface does not save the caller from learning the implementation.
+
+## Information leakage
+
+Information leakage makes multiple modules depend on the same internal decision. A representation, policy, or protocol detail appears in more than one place, so changing it requires coordinated edits.
+
+Public re-exports of transport or wire types are leakage. Parse external data into domain types behind the interface. Keep storage schemas, framework objects, and protocol details private.
+
+## Temporal decomposition
+
+Temporal decomposition organizes modules by execution order instead of the knowledge they own. Separate load, validate, transform, and save stages often repeat one representation and its invariants across several boundaries.
+
+Group code around domain knowledge and ownership. Methods that run at different times can still belong to one module when they protect the same decisions.
+
+## Pass-through method
+
+A pass-through method forwards the same arguments to another method with the same shape. It adds a layer without hiding complexity.
+
+Remove it or move responsibility to the module that can complete the operation. Keep a forwarding boundary only when it adds policy, adaptation, or a distinct abstraction.
+
+## Split ownership
+
+More than one module writes the same state or keeps its own copy of it. An agent that edits one writer can't see the others, so their rules diverge.
+
+Give each piece of state one owner. Other modules read it or ask the owner to change it.
+
+## Two ways to do one task
+
+The design supports more than one way to do the same task. An agent copies whichever way it finds first, so every way keeps gaining callers.
+
+Keep one way. Move callers off the others and delete them in the same change.
+
+## Importable internals
+
+A caller can import a module's internals. An agent takes the shortest path that compiles, so it imports them directly and they become part of the interface.
+
+Make internals unreachable from outside the module, so an import from outside fails the build.
+
+## Hand-synced list
+
+Two or more places list the same items, and adding an item means editing every list. An agent that sees one list updates only that one.
+
+Keep one list and derive the others from it. If a list can't be derived, make the build fail when the lists disagree.

package/bundled-skills/skills/peak/references/architect/references/rationale-template.md

@@ -0,0 +1,35 @@
+# Rationale template
+
+The prose that ships alongside the type sketch. One page. Sentence-case headings, no boilerplate. Replace the italic notes with actual content.
+
+## Problem
+
+*One paragraph. What we're trying to do, and what about the existing system or constraints makes the shape non-obvious. If [Phase A](../SKILL.md#phase-a-ground-the-problem) surfaced constraints the design must honor (existing types to interop with, callers we can't break, invariants that crossed our boundary), name them here so the reader sees the same constraints you saw.*
+
+## Usage (caller's view)
+
+*Write this first, before the type sketch. Show the README or quickstart the consumer reads, plus two or three realistic call sites in their own code. What they import, what they call, what comes back. The type sketch in [Shape](#shape) is derived from this. The two must agree. When they diverge, reconcile the sketch to the usage, not the reverse. The caller's experience is the spec. The types serve it.*
+
+## Shape
+
+*The recommended architecture. Data structures first. Then how data flows through the signatures. Name the load-bearing decisions. State which invariants are encoded in types, where validation lives, and what the system deliberately does not do. Judge interface depth explicitly. State what complexity the public surface hides, what remains exposed to callers, and why the interface is no larger than needed. Cite the principle behind each decision (e.g., `per boundary-discipline`). Don't restate it.*
+
+## Synthesis decision
+
+*Filled in by [arena](../../arena/SKILL.md). Records which candidate became the base and why, what was adapted from each of the others, and what was rejected and why.*
+
+## Tradeoffs accepted
+
+*One bullet per tradeoff the chosen shape makes. Form: "we accept X in exchange for Y." Name anything a future reader might mistake for an oversight, including things that look like premature optimization or premature simplification.*
+
+## Alternatives considered
+
+*Required. Name at least one concrete alternative shape, with one line on why it lost. Judge each alternative on interface depth, not implementation simplicity alone. Name the complexity it exposes to callers and the complexity it hides. Two or three alternatives belong here when the design space had real contenders. One is fine when the constraints forced the answer, with the conclusion phrased as "this was the only viable shape because..." Avoid listing flavors of the same shape. This section covers design alternatives the chosen shape considered and rejected, not other runner candidates.*
+
+## Open questions and risks
+
+*Things you noticed during the sketch that the human needs to weigh in on, and risks worth flagging before implementation starts. Phrase as questions, not assertions, so the human's answer is the resolution rather than a comment.*
+
+## Next implementation step
+
+*The first thing to build against the sketch. One sentence. What you'd start writing immediately after synthesis (or after Phase D sign-off, if a checkpoint was opted into).*

package/bundled-skills/skills/peak/references/architect/references/runner-prompt.md

@@ -0,0 +1,20 @@
+# Architect runner prompt
+
+The orchestrator passes this file through to every parallel candidate runner during Phase B and fills in the variable inputs around it: the task, the Phase A grounding artifacts, the isolated working directory, and the path to write outputs. The working directory is a git worktree when available, otherwise a per-runner subdirectory under the sketch dir. What matters is independence between candidates.
+
+You are producing one candidate design in architect's parallel exploration. Read the **architect** skill in full first. That's the workflow you're inside. Output a candidate design package: type sketch, function signatures, module map, and prose rationale shaped per [`rationale-template.md`](rationale-template.md).
+
+Apply the following discipline. The orchestrator compares candidates on these axes to pick a base.
+
+- Caller's usage first. Write the README-style usage and two or three real call sites before the types, then derive the type sketch from them. The usage is the spec. The two must agree, so reconcile the sketch to the usage, not the reverse.
+- Data structures first. Get the core types right and the code becomes obvious. Trace each dominant access pattern through the proposed structure. If the answer is "we'll add a map / index / cache later," the structure is wrong.
+- Interface depth. Compare the capability hidden behind the public surface relative to the size of that surface. Prefer a simple interface that pulls complexity into the callee, even when the implementation becomes less simple. Do not put transport or wire types on the public API. Parse into domain types behind the interface.
+- Shared state: if two actors might both write, ask "what happens?" If the answer isn't "nothing," default to per-actor state with a merge at the read boundary, per the **separate-before-serializing-shared-state** principle skill.
+- Make boundaries visible. `not implemented` errors for bodies, `// TODO` pseudocode for tricky logic, doc comments stating intent and invariants. A reader should trace data from input to output by reading types and signatures alone.
+- Encode invariants in types: hard-to-misuse types > runtime checks > prose comments, per the **encode-lessons-in-structure** principle skill.
+- Validate at boundaries, trust types inside, per the **boundary-discipline** principle skill. Business logic as pure functions. The shell stays thin.
+- Single source of truth per invariant. Derive instead of sync.
+- Idempotent state transitions where applicable, per the **make-operations-idempotent** principle skill. Ask what happens if the operation runs twice or crashes halfway.
+- Short call chains. If tracing the flow needs more than three files, flatten the hierarchy, per the **laziness-protocol** and **minimize-reader-load** principle skills.
+
+You are one of the parallel runners, each on a different model. Produce the best design your model can make. Don't hedge against the others. Differences between candidates are the signal used to pick a base and graft. Converging on a safe-looking middle defeats the exploration.

package/bundled-skills/skills/peak/references/arena/SKILL.md

@@ -0,0 +1,50 @@
+---
+name: arena
+description: "Spawn N parallel candidates at the same task, pick a base, graft the strongest parts of the losers into it. Use for /arena, 'arena this', 'throw it in the arena', or when one attempt at a non-trivial artifact would lock in the wrong shape."
+disable-model-invocation: true
+---
+
+Read [the native Python adapter](../raga-runtime/SKILL.md) before executing this workflow. Authoring targets candidate revisions. Children use run_child, not an external agent service.
+
+# Arena
+
+Open a todolist with one entry per phase (Frame, Fan out, Cross-judge, Pick, Graft, Verify) before launching anything.
+
+## Phase A: Frame
+
+Every candidate gets the same prompt, so the prompt is the contract.
+
+1. State the artifact each candidate is producing.
+2. Derive the rubric, 3-6 concrete gradeable criteria for what success looks like on *this* task. It is the picker's tool in Phase D. Candidates only see the task.
+3. Pick the runners. Use the `arena runners` line in `agents/model-roles.json`. If no role is configured, inherit the parent model and report that the seats are not model-diverse. An `auto` or `inherit-parent` entry in this line or the cross-judge line means the parent model, so omit `model` for it. If a configured model is rejected, report the unavailable selector and stop that seat. Compare actual configured provider/model families, not editor aliases. Report an unavailable configured selector instead of replacing it. An unconfigured role inherits the parent model. An unavailable model is a prerequisite failure. Spawn more runners when the arena covers multiple design directions. Use the same model N times when the work is generation-bound rather than judgment-sensitive. Pick each subagent's reasoning level and fast or standard speed for the job, per the Subagents section of the **peak** skill.
+4. Give each candidate its own output location (a supervisor-owned child workspace for code, or an independent returned artifact for a read-only design), per the **separate-before-serializing-shared-state** principle skill.
+
+## Phase B: Fan out
+
+Spawn all N subagents in one message with `asyncio.gather for independent child calls`, each with the task, the path to the shared grounding, its own output path, and instructions to produce the artifact and a short rationale naming the alternatives it considered and what it rejected.
+
+If a candidate produces no output, proceed with N-1 and note the dropout.
+
+## Phase C: Cross-judge
+
+After all Phase B candidates complete, choose one model from the `arena cross-judge pool` line in `agents/model-roles.json`. If the rule or that line is missing, choose from `None` and `None`. Prefer a different model family from the parent's. Spawn one readonly judge subagent on that model. It sees the rubric and the candidates by path label, scores each criterion, and recommends a base with rationale. It runs in parallel with your reading in Phase D. Don't spawn it while candidates are still writing.
+
+## Phase D: Pick a base
+
+Read every candidate end to end before picking. Score each against the rubric criterion by criterion, not on holistic feel, and compare with the cross-judge. Agreement on the base confirms the pick. Disagreement means one of you is biased or the rubric was ambiguous. Read both rationales before deciding.
+
+Pick the base a future maintainer can extend most easily without breaking invariants. When two feel tied, prefer the cleaner boundary or smaller API, per the Laziness Protocol.
+
+## Phase E: Graft
+
+Walk each losing candidate once more for what is worth porting into the base, usually one or two things per candidate, not most of it. Fold each graft in by hand, per the **redesign-from-first-principles** principle skill, so the result stays coherent under one mental model. Don't paste mechanically.
+
+If the candidates converge on one shape, that is strong agreement. Note it and ship the consensus shape. No graft is needed. If they wildly diverge, Phase A was under-specified. Reframe and re-run rather than averaging the divergence.
+
+## Phase F: Verify
+
+Verify the synthesized artifact like any other output, per the **prove-it-works** principle skill. If verification finds a problem the arena missed, either Phase A was wrong (re-frame and re-run) or a candidate caught it and you missed the graft (go back to Phase E). Don't paper over it.
+
+## Outputs
+
+One synthesized artifact, with a short synthesis note alongside naming the base and why, the cross-judge's verdict, each graft with its source candidate, what was rejected and why, any convergence or dropouts, and the verification result.

package/bundled-skills/skills/peak/references/automate-me/SKILL.md

@@ -0,0 +1,106 @@
+---
+name: automate-me
+description: "Use for \"automate me\", \"create/update/refresh my -mode skill\", \"turn/capture my preferences or working style into a skill\", or wanting agents to follow how the user works. Drafts or revises a personal -mode skill via create-skill + unslop, optionally pulling fresh evidence from recent transcripts."
+disable-model-invocation: true
+---
+
+Read [the native Python adapter](../raga-runtime/SKILL.md) before executing this workflow. Authoring targets candidate revisions. Children use run_child, not an external agent service.
+
+# Automate me
+
+A guided flow for turning the user's working conventions into a skill agents will follow. The output is one `-mode` skill tailored to them (e.g. `jay-mode`, `priya-mode`).
+
+This skill orchestrates three others: an inline mining pass (see step 1), the bundled `create-skill` skill (authoring), and the **unslop** skill (prose discipline). It sequences them. It doesn't replace them.
+
+## Flow
+
+### 0. Check for an existing skill
+
+Look recursively for `skills/**/*-mode/SKILL.md` and `skills/*-mode/SKILL.md` matching the user's handle. Mode skills can live in a personal category directory (`skills/<handle>/`), not only at the top level. If one exists, confirm intent with `ask the user in the final response` (unless they already said "update my skill" or similar):
+
+- Update the existing skill (default for repeat runs)
+- Start fresh (rare, ask why before doing it)
+
+Update mode changes the rest of the flow:
+- Step 1 mines only history since the skill was last edited (`git log -1 --format=%cI <path>`).
+- Step 2 asks what's changed or missing, not what to capture from zero.
+- Step 4 edits only the supervisor-created candidate. Preserve sections the user hasn't contradicted. Revise ones with new evidence. Add new sections only for genuinely new rules.
+
+### 1. Mine their history
+
+Locate the active workspace's transcripts before fanning out. Use await context.history() for this session. For previous work, use await memory.search(query), then await memory.get(id). Only mine earlier transcripts explicitly supplied by the supervisor.
+
+Survey recent agent conversations within that scope for recurring patterns. Run multiple parallel subagents across slices of history (e.g. last 2-4 weeks, split into 3 slices so each has enough material). Pick each subagent's reasoning level and fast or standard speed for the job, per the Subagents section of the **peak** skill. Each slice mining subagent reads transcripts from the workspace-scoped path the parent provides, looks for the signals below, and returns a short structured list of patterns it saw with evidence pointers. Default signals worth hunting:
+
+- Response preferences (length, tone, format, "dumb it down" corrections)
+- Delegation habits (subagents, models, specialized workflows, parallelism)
+- Verification posture (what "done" means, unit tests vs live repro, reviewers)
+- Code and prose discipline (style, principles cited, lint/format tools)
+- Process conventions (worktrees, commits, PRs, review/merge tooling)
+- Meta preferences (fixing skills mid-task, proposing new ones)
+
+Cross-check across slices before elevating a signal. Patterns seen in 2+ slices are high-confidence. Lone signals are weak and usually get dropped.
+
+### 2. Ask the user directly
+
+Mining misses intent that hasn't come up yet. Ask a plain-language question in the final response. There is no structured question tool.
+
+Ask one or two questions in ordinary language. The user can answer with their own words and select more than one category.
+
+Don't dump 20 questions.
+
+### 3. Cluster findings
+
+Group the combined signals into sections. Common ones (use only what applies):
+
+- **Response style**: length, tone, format.
+- **Autonomy**: how much to do without asking, MCP tool use.
+- **Understand first**: which skills to reach for when scoping or investigating a change.
+- **Subagents**: default, parallelism, model-to-task, specialized workflows.
+- **Prose / code discipline**: principles, lint tools, style guides.
+- **Review and verify**: repro posture, verification skills, live-testing tools.
+- **Process**: git worktrees, commits, PRs, review/merge tooling.
+- **Skills**: skill-authoring habits, fix-the-skill-first, proposing new skills.
+
+The **peak** skill shows the shape. Read it for granularity. Don't copy its content. The user's rules are not the same as peak's.
+
+### 4. Draft the skill
+
+Use the bundled `create-skill` skill skill to author the skill. Placement:
+
+- Path: preserve an existing mode skill's category. For a new mode, use `skills/<handle>/<handle>-mode/SKILL.md` when the repo has an established personal category for that handle. Otherwise default to `skills/<handle>-mode/SKILL.md` in the project (or `skills/<handle>-mode/` if the user prefers a personal skill).
+- Handle: the user's first name or chosen identifier.
+- Frontmatter `description`: trigger on their name + `/<handle>-mode` + "work in their style", not on generic keywords like "write code" or "review PR".
+- Frontmatter formatting: follow `create-skill`'s YAML rules. Keep `description` as one YAML scalar. Quote it or use `description: >-` with indented continuation lines when punctuation or wrapping requires it.
+- Frontmatter `disable-model-invocation: true` by default. Opt out only if the user explicitly wants their mode to apply on every turn.
+
+### 5. Iterate on prose
+
+Apply the **unslop** skill and `create-skill`'s writing guidelines to every line.
+
+Show the draft to the user and take feedback. Expect multiple iterations. Cut ruthlessly. A mode skill is not a manual.
+
+### 6. Land it
+
+Write the skill through instance.candidate_write and commit through instance.candidate_commit. Return the candidate and evidence for explicit operator promotion. Do not open an application PR for an instance-only proposal.
+
+## Guardrails
+
+- **Don't overfit to one conversation.** A preference stated once and contradicted another time is noise. Require multiple instances before codifying it.
+- **Don't be clever.** Restating other skills' contents, inventing metaphors, or writing "poetic" prose for an agent reader is cost without benefit. Keep it operational.
+- **Reference, don't inline.** Other skills the user relies on should appear as path references, not pasted excerpts. Same for any principle docs they maintain elsewhere.
+- **Keep sections minimal.** Only add a section if the user has a specific, non-default rule there. "Communicate clearly" is not a section. "Short paragraphs. Tables when comparing options. Bullets only when items are genuinely parallel." is.
+- **Name conventions generic.** Use "the user" or "the human" in imperatives, not the author's first name.
+- **Don't force symmetry.** If a user has no process rules worth writing down, skip the Process section entirely.
+
+## Evaluation
+
+A `-mode` skill is subjective output. A `create-skill`-style test/iterate benchmark loop isn't useful here. Vibe-check with the user: does it read like them? Did it miss anything? Then ship.
+
+Run a description-optimization loop only if the skill's trigger accuracy turns out to be a problem in practice.
+
+## When not to use
+
+- User wants a task-specific skill (not working conventions): `create-skill` alone, no mining required.
+- User wants to capture one narrow workflow (e.g. "how I write commit messages"). That's a regular skill, not a mode skill.
+

package/bundled-skills/skills/peak/references/benchmark-checklist/SKILL.md

@@ -0,0 +1,41 @@
+---
+name: benchmark-checklist
+description: "Vet a perf measurement (limiter, tuning, limits, errors, repeatability, relevance, and whether the work happened) before you report or act on it. Use when you run a benchmark or report a speedup or regression you measured."
+disable-model-invocation: true
+---
+
+Read [the native Python adapter](../raga-runtime/SKILL.md) before executing this workflow. Authoring targets candidate revisions. Children use run_child, not an external agent service.
+
+# Benchmark checklist
+
+Use this when you produce a performance number: a PR's before and after, a regression claim, a hillclimb harness, or a library or config choice. Answer each question with evidence from a run, not a guess about the code.
+
+For a quick ballpark the user asked for, one run is enough. Still check questions 4 and 7, and say that it is one run. Skip the rest unless that run looks wrong. A choice between options is never a ballpark.
+
+## Before you run anything
+
+- Write down the claim you expect to make, in the words you would ship ("export is 30% faster at p50 on the 60k-row dataset"). The questions test that sentence.
+- Read the measurement script. Note what it times, counts, and ignores.
+- Check the load average with `uptime` and the core count with `nproc`. If the machine is busy, find out what is running. If you cannot stop it, interleave the sides so both see the same noise, and say so in the report.
+
+## The questions
+
+1. **Why not double?** Name the limiter. Profile in a run you do not report, because profilers and tracers slow the work down. Use CPU per process (`top`, `pidstat`), a runtime profiler (`node --cpu-prof`, `py-spy`, `perf`), I/O wait, and syscall counts (`strace -c` on Linux), then map the hot spot to source. Watch the load generator too: if it saturates first, you measured the load generator. If a change did not move the number, the limiter explains why, so find it before you call the change useless.
+2. **Was it tuned?** Run every side the way production runs it: release builds, production flags and env, batching and transaction settings, connection pools, caches as warm or cold as production sees them, and the same versions and data. If one side runs on defaults, you compared configurations, not implementations. A limiter that is a setting (a commit per row, a debug build, a missing index) means that side is untuned. Tune it and measure again before you pick a winner. If you cannot tune it, do not pick a winner from that run. When the user is choosing what to adopt, narrowing the claim to the code as it ships today does not fix this, because they adopt the option, not today's settings.
+3. **Did it break limits?** Do the arithmetic. Compare bytes per second with disk and network bandwidth, operations per second times cost per operation with your cores, and the time saved with the time the changed piece took. Removing a piece that takes 10% of the run can make the run at most about 11% faster. A result past a limit measured something other than the work, such as a cache, a no-op, or a bug.
+4. **Did it error?** Count failures and non-success responses, and check that outputs are correct, not just present. Rejections are often fast, and timeouts and retries are slow. If the script does not count errors, add the count.
+5. **Does it reproduce?** Run each side at least 5 times, alternating (A, B, A, B, and so on) so warmup, lazy initialization, caches, and drift favor neither. Report the median and range. A gap smaller than the run-to-run variation is no measurable difference. When the call is close, use a rank-sum test or the harness's own statistics.
+6. **Does it matter?** Next to any micro result, measure the end-to-end path a user waits on, with realistic data sizes and concurrency, and report the micro result as a share of the whole. A helper that takes 1% of a request can make the request at most 1% faster, however fast the helper gets.
+7. **Did it even happen?** Confirm the work ran inside the timed region: the request reached the server, the rows were written, the bytes were read, and the code used the result. Lazy code (generators nobody iterates, promises nobody awaits, results the JIT can discard) and timeouts produce numbers for work that never happened.
+
+## Report
+
+- Lead with the verdict: faster, slower, no measurable difference, or inconclusive.
+- Give the number with its unit, the run count, the range, and the limiter. For example, "p50 41 ms → 33 ms, median of 7 runs per side, range 32 to 35 ms after, bound by JSON parsing on one core."
+- Call the verdict inconclusive when you claim a difference but cannot name the limiter, when a side ran untuned, or when you could not check questions 4 and 7. Name the gap.
+- Keep a PR body to one primary number, per the **Opening a PR** playbook. Put the runs, range, and limiter evidence in a linked artifact or notes file.
+
+## Other perf material
+
+- Vet the **Perf issue** playbook's baseline before the playbook plans from it, and every number after that.
+- Vet the **Hillclimb** playbook's harness before it is frozen. The frozen harness then prints error and work counts, so each keep-or-revert checks questions 4 and 7 for free.

package/bundled-skills/skills/peak/references/blast-radius/SKILL.md

@@ -0,0 +1,52 @@
+---
+name: blast-radius
+description: "Find what a change could break somewhere else before it ships, beyond the diff, and prove the one fact it's safe because of by running real code instead of writing it up. Use for 'blast radius of X', 'what could this break', or reviewing a small diff you don't trust."
+disable-model-invocation: true
+---
+
+Read [the native Python adapter](../raga-runtime/SKILL.md) before executing this workflow. Authoring targets candidate revisions. Children use run_child, not an external agent service.
+
+# Blast radius
+
+Find what a change breaks somewhere else, before it ships. Use for "blast radius of X", "what could this break", or reviewing a small diff you don't trust yet.
+
+Companion to `how` and `why`. `how` tells you what the code does. `why` tells you why it's shaped that way. Blast radius tells you what it breaks somewhere else.
+
+Listing the callers is not the job. The agent can grep those in a second. The job is the breakage grep won't show you.
+
+## Don't trust your own writeup
+
+A blast-radius writeup that sounds right is worthless. It reads as convincing whether or not it's true. So don't hand back the writeup. Find the one or two facts the whole thing depends on and prove them by running code.
+
+### How sure are you
+
+For each fact the change's safety depends on, get it as far down this list as is cheap, and say where it stopped.
+
+1. You said so. Worthless on its own.
+2. You pointed at the line. A real `file:line`, or the library's own source.
+3. You showed the bad case can't happen. You walked the failure step by step and it doesn't reach.
+4. You ran it. A script or test that calls the real code and fails loud if you're wrong.
+5. You reproduced it in the running app.
+
+Step 4 is usually one small script that imports the same library the app ships and calls the exact function you're worried about.
+
+## Steps
+
+1. Read the change. The diff, the symbols it adds, changes, and deletes, and what it now does differently, including the part the diff doesn't spell out. Use `why` step 2 to pull the PR and commits.
+2. Find the one fact it's safe because of. Most changes that look risky are safe because of a single fact, like "this call only drops already-dead cache entries and does nothing else". Find that fact. If it holds, most risky cases are cleared at once. Spend your time here, not on a long list of maybes.
+3. Look where grep stops. Read the source of the library you call, and check its pinned version and any local patch. Work out when things run: microtasks, unmount and teardown, Solid versus React. Follow what a symbol search misses: the JSON an API returns, a DB column, a wire format, another language reading the same bytes, a feature flag, code three hops downstream.
+4. Be honest about each risk. Give it a real chance of happening and a real cost if it does. Keep the risks you confirmed. List the ones you checked and cleared separately. Same rules as `why`. Cite a real `file:line`, a search that finds nothing is still an answer, and never make up a caller or an API.
+5. Prove the one fact. Write a script or test that runs the real code, run it, and paste what happened.
+6. For a big or wide change, run it as an `arena`. Ask more than one model the same question and merge the answers. Different models catch different real bugs.
+
+## What to hand back
+
+- **What it does.** What changed, including the part that isn't obvious.
+- **The one fact it's safe because of.** State it, say which step you got it to, and show the proof. If you couldn't prove it, write unproven.
+- **Risks.** Each names how it breaks, the `file:line`, how likely and how bad, and how to check. Paste the proof for the ones that matter.
+- **Cleared.** What you checked and why it's fine.
+- **Before you merge.** The cheapest test or repro that catches the real bug, including the script you wrote.
+
+Write it through `unslop`, cite real code, and strip anything private before it goes anywhere public.
+
+**Reply:** the writeup above, with the one safety fact either proven or marked unproven.

package/bundled-skills/skills/peak/references/bro/SKILL.md

@@ -0,0 +1,9 @@
+---
+name: bro
+description: Restate the last message in plain human language, with no jargon.
+disable-model-invocation: true
+---
+
+Read [the native Python adapter](../raga-runtime/SKILL.md) before executing this workflow. Authoring targets candidate revisions. Children use run_child, not an external agent service.
+
+Restate your last message. Stop using jargon and speak coherently. State it more simply and concisely, like one human talking to another.