lossyrob-skills

Reusable Copilot CLI skills.

Install

If you previously installed the plugin directly from the repository, uninstall that copy first:

copilot plugin uninstall lossyrob-skills

Add this repo as a Copilot CLI plugin marketplace, then install the skills plugin from it:

copilot plugin marketplace add lossyrob/skills
copilot plugin marketplace browse lossyrob-skills
copilot plugin install lossyrob-skills@lossyrob-skills

Skills

council

Convene a bounded multi-model, multi-perspective deliberation to produce a recommendation with confidence, preserved dissent, and a transcript. Useful when a consequential decision benefits from independent perspectives, collaboration, and focused contest rather than another single-model critique loop.

Trigger phrases: "council", "panel", "deliberate", "multi-agent review", "advisory council", "get diverse perspectives", "discuss this with specialists"

Features:

• Gated, not ambient: use only when diversity, adversarial testing, or synthesis is likely to beat a single frontier model with enough reasoning time
• Adaptive primitives: starts with an isolated panel, then escalates only if needed into deliberate, debate, spar, red-team, or delphi modes
• Focus-conserving: holds a decision vector, classifies findings CORE/ADJACENT/PARKED, rewards sharper claims over disagreement-for-its-own-sake, and weights grounding over delivered confidence to prevent drift into tangential spikes
• Dynamic roster: accepts user- or driver-chosen models, personas, PAW SoT specialists, and perspective overlays; persona diversity is the guaranteed floor while model assignment is best-effort (it can silently fall back, so it is requested, not guaranteed, and never trusted via self-report)
• Dissent-preserving synthesis: returns a required-field packet (recommendation, confidence basis, decisive arguments with sources, minority report, open questions, reopen conditions, audit triggers, provenance manifest, coverage manifest, artifact paths); the main session always surfaces minority report and reopen conditions, and a faithfulness check guards rapporteur fidelity
• Context containment: a single contained council-runner subagent spawns members, runs the rounds in its own context, flushes the transcript turn-by-turn, and returns only a provenance-bearing synthesis; the driver reads synthesis by default and audits the transcript on request. Contained by default, with a light bounded context-request back to the driver only when the brief is thin -- say "closed council" to disable it, or ask for more check-ins when you want them
• Three execution modes: integrate, brief-back, and artifact-only

Requirements: Copilot CLI with the task subagent tool, nested subagent spawning (the runner spawns members as nested subagents), and at least one strong model. Multi-model councils need more than one available model; single-model councils can still use persona diversity, but should disclose the weaker diversity signal.

launch-copilot-terminal

Launch a new Windows Terminal tab running Copilot CLI with a requested title, tab color, and working directory. Supports a prompt-driven interactive session, an existing-session resume, and targeting either a separate window or the current Windows Terminal window. Useful for starting parallel Copilot sessions, focused worker windows, or opening a resumed session beside the current one.

Trigger phrases: "launch Copilot terminal", "open Copilot window", "start Copilot session", "spawn Copilot worker"

Features:

• Opens a Windows Terminal tab with a chosen title and tab color
• Two modes: prompt mode starts Copilot with copilot -i <prompt>; resume mode reattaches to an existing session with copilot --resume <id-or-name> (no prompt submitted)
• -Window {new|current} selects whether the tab opens in a separate window (wt -w -1, the default) or in the current Windows Terminal window (wt -w 0)
• Supports explicit working directories, extra Copilot CLI arguments, and prompt files
• Includes dry-run output for inspecting the generated launch command

Requirements: Windows, Windows Terminal (wt.exe), PowerShell 5.1+/7+, and Copilot CLI on PATH.

loop

Repeatedly run a check command or script on a configurable interval until a condition is met, a timeout is reached, or an actionable exit code is returned. Useful for waiting on services, polling PR status, retrying commands, and watching CI/review/mergeability gates before continuing.

Trigger phrases: "loop", "wait until", "poll", "watch", "retry", "check every N minutes", "wait for CI", "wait for approval"

Features:

• Cross-platform runners for Bash (scripts/loop.sh) and PowerShell, where agent workflows default to detached Start-LoopDetached.ps1 with attached loop.ps1 reserved for short-lived/debug cases
• Fixed interval, timeout, max tries, exponential backoff, jitter, and stability windows
• Retry-vs-stop exit-code routing so the agent can fix actionable states and restart the loop
• Optional success action, acknowledgement command, retry hook, singleton lock, quiet mode, and dry-run output
• PowerShell persistent state files (last-result.json, heartbeat.json, immutable event files) plus detached launch/status/wait helpers with PID start-time validation; the quiet waiter lets the agent sleep until detached state becomes actionable or final and can be backgrounded in the host CLI while preserving automatic wakeup
• GitHub PR readiness helpers for approval, checks, merge conflicts, closed PRs, and merge race protection

Requirements: Bash for macOS/Linux/Git Bash workflows or PowerShell 5.1+/7+. GitHub PR polling requires gh.

session-branch

Branch the current Copilot CLI session, creating a new session that inherits conversation history up to the current point while preserving the original session intact. Useful for experimentation or parallel development without losing your place.

Trigger phrases: "branch", "branch session", "fork session", "create a branch from here". Append "launch in terminal" or "open in a new tab" to open the branched session in a new Windows Terminal tab beside the current one (Windows only).

Features:

• Copies full session state (events, workspace config) into a temporary staging directory and atomically renames into place; failed branches never leave a half-branched session on disk
• YAML-aware workspace.yaml rewriter handles block-scalar names (name: |-, folded >, etc.) correctly — the original block body is fully replaced and the branch title is derived from the reconstructed content
• Post-rewrite validation refuses duplicate top-level keys and orphan indented lines, so corruption in the source is caught before the branched session is committed to disk
• Assigns each branch a unique Copilot CLI resume title like Branch: <title> [<id>] and tracks lineage via branch_of / branch_note in workspace.yaml
• Includes Bash and Windows PowerShell branching workflows; branch logic ships as scripts/branch_session.py with a unittest test suite
• Removes stale in-use locks from the branched session; resets checkpoints and rewind snapshots for a clean slate
• Optional "launch in terminal" mode (Windows-only) opens the branched session in a new Windows Terminal tab inside the current window via the launch-copilot-terminal helper
• Optional truncation ("branch from N turns ago") and optional git worktree integration

Requirements: Python 3. On Windows, the skill validates Python candidates and prefers the Python launcher (py -3) or a real python.exe before falling back to python3, avoiding unusable Windows Store app-execution aliases. The launch-in-terminal mode additionally requires Windows, Windows Terminal (wt.exe), and the launch-copilot-terminal skill.

odt-convert

Convert ODT (OpenDocument Text) files to Markdown with full comment and embedded object extraction.

Trigger phrases: "convert odt", "extract odt comments", "odt to markdown", or when working with .odt files

Features:

• Document body conversion via pandoc with --wrap=none
• Threaded comment extraction with anchor text and reply grouping
• Inline image extraction (fixes pandoc []{.image} placeholder failures)
• Visio diagram extraction (.vsdx) with PNG preview generation
• All media output to a <name>-embedded/ subdirectory

Requirements: pandoc, Python 3. Optional: olefile (Visio), libreoffice (EMF→PNG).

paw-pr-lifecycle

Operate PAW implementer and reviewer GitHub PR lifecycle loops on top of the loop skill: PR discovery for reviewers, review-response and merge-readiness sentries for implementers, marker-driven handoff between roles, and re-review requests after substantive post-approval changes. Opinionated for PAW workflow sessions and their 🐾 PAW … comment/review markers.

Trigger phrases: "PAW PR lifecycle", "PAW implementer loop", "PAW reviewer loop", "watch this PR until merge", "wait for PAW approval", "PAW PR sentry"

Features:

• Mode-based: Implementation → Review Response → PR Sentry for the implementer; PR Discovery → Review → Follow-up Sentry for the reviewer
• Marker contract for the three 🐾 PAW … events; a +1 review may include non-blocking notes and the implementer is required to read the body before transitioning to PR Sentry
• Canonical check scripts for implementer/reviewer lifecycle loops (impl-review-response-check.ps1, impl-merge-sentry-check.ps1, review-pr-discovery-check.ps1, review-addressed-check.ps1, review-azdo-addressed-check.ps1, plus the shared github-loop-common.ps1) with GitHub rate-limit/transient-error routing through the loop skill's retry/stop exit codes
• Get-LoopScriptPaths.ps1 resolves the sibling loop skill automatically (checked-out repo → default plugin install → bare-skills install → recursive ~/.copilot fallback)
• Multi-account gh support: the loop scripts assert the requested <gh-user> is authenticated and pin API calls to that account's token

Requirements: PowerShell 7+ on any OS, GitHub CLI authenticated against github.com, and the sibling loop skill (≥ 0.1.12).

backlog-orchestrator

Drive a backlog of GitHub issues to PRs autonomously and sequentially. The loaded session becomes an orchestrator that triages issues into S/M/L tiers, spawns PAW implementer (and optional PAW Review) worker terminals that coordinate over telex instead of GitHub-comment polling, gates each PR through a preference/human-floor merge review, and auto-merges or routes to human review.

Trigger phrases: "work through a backlog of issues autonomously", "run an autonomous issue-fixing pipeline", "orchestrate PAW sessions across many issues", "drive these issues to PRs"

Features:

• Four-phase model: telex station setup → interactive triage (S/M/L sizing + per-tier config) → sequential per-issue execution → merge gate + advance
• Spawns an implementer (paw-lite, loaded as a skill) and an optional reviewer (launched as the PAW-Review agent with autonomous review submission) in their own terminals via launch-copilot-terminal; they run the review handshake over telex (review-ready → review-posted → re-review → 🐾 +1), not GitHub-comment polling
• Last-line merge gate: an Opus subagent detects high-spread preference forks the builder should own (a filtered work-geometry lens, not a correctness re-review), tuned by a per-issue care-knob — auto-merges clear/low-spread PRs and routes preference-debt / constitution / human-floor PRs to human review
• Deferred-work tracking: every carry-forward item is harvested at the gate (field report + diff markers) and driven to a terminal disposition (filed / folded / skipped / done / moot) — the run is not complete while any item is open
• Deferred human-review holds: a PR routed to you stays live — the implementer's merge sentry keeps it mergeable (repairing CI/conflicts) until you merge, then reports back for stand-down
• Field reports on each issue + a run ledger; a final report bubbles up pivots, preference debt, no-auto-merge decisions, deferred work, and learnings, plus a process-feedback → skill-improvement loop
• Durable telex backend pinning so messages survive holder restarts and wake idle worker sessions

Requirements: Windows + Windows Terminal (for launch-copilot-terminal), Copilot CLI on PATH, telex on PATH, GitHub CLI authenticated for the target repo, and the installed skills launch-copilot-terminal, paw-pr-lifecycle, loop, spar, plus the PAW workflow skills (paw-lite / paw-review-workflow) and the PAW-Review custom agent.

spar

Get a sharp second opinion from a different model before committing to a consequential decision. A pairing-style critique skill run as sparring rounds: it keeps the structure of pair programming (shared goal, two perspectives per decision, the driver holds the pen) but drops the ego-protecting hedging that turns review into a rubber stamp, since there is no human ego to protect between agents.

Trigger phrases: "pair", "spar", "rubber duck", "get a second opinion", "discuss amongst yourselves", "consult another model before committing"

Features:

• Gated, not ambient: open an episode only for decisions that are both consequential and uncertain (load-bearing design/contract, plan invalidation, non-trivial fork, boundary change, repeated failure) — never for routine coding or "review this diff"
• Cross-model by design: the pair runs as a different-model rubber-duck subagent (default Opus 4.8 when the driver is GPT-5.5, and vice versa); operators can pin a model
• Dialogic depth: gate the entrance, not the depth — run several short rounds while they keep resolving the same decision, with a context packet to offset the subagent's missing history
• Anti-sycophancy: the pair attacks assumptions and surfaces failure modes rather than validating; explicit closure (revise the plan or record why the critique does not apply)
• Two modes: integrate (fold the outcome in and proceed) or brief-back (bring the operator a brief of gaps and recommendations)
• Caller-customizable: prompts or operators can add domain triggers, pin a model, or force a mode

Requirements: Copilot CLI with the task (subagent) tool and a second model available for the pairing subagent. Uses the rubber-duck agent type when available and falls back to a general-purpose subagent with the sparring role in its prompt; if the named default pairing model is unavailable, any model different from the driver works.

License

MIT