srgical is a local-first orchestration CLI for the workflow you have already been using manually:
- talk to an AI until the plan is sharp,
- write a five-file planning pack into the repo,
- repeatedly execute the next eligible step,
- force validation and handoff updates every time.
The current launch slice supports local codex, local claude, and local auggie installs through the same
.srgical/ workflow. srgical detects which supported tools are actually installed, keeps the planning pack
agent-neutral, and lets you choose the active agent for the current workspace session.
Install srgical, then make sure at least one supported local agent CLI is installed (codex, claude, or auggie).
npm install -g @launch11/srgical
srgical doctor --plan release-readinessCreate a named plan pack (required):
srgical init --plan release-readinessOpen studio and build context:
srgical studio --plan release-readinessInside studio:
- talk through scope/constraints and use
/readiness+/advice - inject repo files directly with
/read <path>(pressTabto autocomplete file paths) - run
/writeto generate the first grounded draft from transcript context - run
/reviewand/open all(or/open <path>) for human doc review - run
/confirm-plan(required before authored-plan refresh writes) - run
/writeagain when refreshing an authored plan - execute with
/preview,/run, or/auto 10(/stopto stop after current iteration)
CLI execution path:
srgical run-next --plan release-readiness --dry-run
srgical run-next --plan release-readiness
srgical run-next --plan release-readiness --auto --max-steps 10The reference system in G:\code\Launch11Projects\Writr\migrations-part-5 is strong because it does not just create a
plan. It creates momentum:
- a stable architecture file,
- a current-context handoff log,
- a step-by-step tracker,
- and a canonical execution handoff document that keeps execution disciplined.
srgical turns that from a repeated copy-paste ritual into a product.
This repo currently ships the foundation for:
srgical --versionPrints the installed version with release-note links instead of only echoing the semver.srgical doctorReports the active plan, plan readiness, execution state, auto-run state, and which supported agents are available locally, along with any cached AI advice for the selected plan.srgical aboutShows package details, release links, and the currently supported agent adapters.srgical changelogPoints straight at the installed version's release notes and the local packaged changelog.srgical initCreates a named local.srgical/plans/<id>/planning pack from built-in templates.--plan <id>is required.srgical studioOpens a full-screen planning studio where you can switch between named plans, inspect readiness with/readiness, inspect supported tools with/agents, refresh AI guidance with/advice, run human review with/reviewand/open, inject large file context via/read <path>, confirm with/confirm-planfor authored-plan refresh writes, and then trigger pack writes, single-step execution, or/auto.srgical run-nextReplays the generated execution handoff through the active agent, with--plan <id>for plan targeting,--dry-runfor safe preview,--agent <id>for a one-run override, and--autofor bounded multi-step execution.
codexSupported in the current launch slice for planning, pack writing, andrun-nextexecution.claudeSupported through the same adapter seam for planning, pack writing, and execution when the local Claude Code CLI is installed and available onPATH.augmentSupported through the same adapter seam by targeting the localauggiebinary for planning, pack writing, and execution when Augment CLI automation is available on the current machine.
If only one supported agent is installed, srgical can auto-select it for the workspace session. If more than one is
installed, you can keep the stored choice in the studio and still override a single execution with
srgical run-next --agent <id>.
The product should feel closer to a creative control room than a grey enterprise shell:
- dark graphite base
- hot coral and amber accents
- crisp cyan status treatment
- large, cinematic panel framing
- transcript-first layout instead of command soup
The first TUI pass already leans in that direction, and we can keep pushing it.
The first production channels are GitHub Packages, the public npm registry, and GitHub Releases for downloadable
release assets. Version intent stays in git, and GitHub Actions bumps package.json, writes CHANGELOG.md, pushes
that release commit back to main, publishes the GitHub-scoped package, publishes the npm org package, and creates a
GitHub Release with the built tarballs attached.
For a local production-style packaging check:
npm run release:packThe release bundle lands under .artifacts/release/. The broader distribution path, including standalone binaries and
wrapper package-manager installs, is documented in docs/distribution.md.
When that branch reaches main, the release workflow versions the package, publishes @launcheleven/srgical to
GitHub Packages, publishes @launch11/srgical to npm, and opens a matching GitHub Release entry with the packaged
artifacts.
The repo keeps a base version line in package.json and CI computes the next patch version from matching git tags. For
example, a base version of 0.0.0 means releases flow as 0.0.1, 0.0.2, and so on. To move to a new minor line,
change the base version to something like 0.1.0.
Install srgical, then install at least one supported local agent CLI separately.
npm install
npm run build
node dist/index.js doctordoctor is the source of truth for local availability. If an agent CLI is missing, srgical reports it as missing
instead of pretending it can run that path anyway.
The package publishes in two install channels:
- GitHub Packages:
@launcheleven/srgical - npm public registry:
@launch11/srgical
For GitHub Packages installs, consumers need an .npmrc entry for @launcheleven plus a token before running:
npm install -g @launcheleven/srgicalFor npm installs, consumers can use:
npm install -g @launch11/srgicalnpm install
npm run build
node dist/index.js init --plan release-readiness
node dist/index.js doctor --plan release-readiness
node dist/index.js studio --plan release-readinessDuring development:
npm run dev -- studioTypical flow once a workspace has a pack:
node dist/index.js --version
node dist/index.js about
node dist/index.js doctor --plan release-readiness
node dist/index.js changelog
node dist/index.js init --plan release-readiness
node dist/index.js studio --plan release-readiness
node dist/index.js run-next --plan release-readiness --dry-run
node dist/index.js run-next --plan release-readiness
node dist/index.js run-next --plan release-readiness --auto --max-steps 10To override the active workspace agent for one execution only:
node dist/index.js run-next --plan release-readiness --agent codex
node dist/index.js run-next --plan release-readiness --agent claude
node dist/index.js run-next --plan release-readiness --agent augmentInside the studio, the footer is intentionally minimal:
PgUp/PgDnscrolls the transcript/agentschooses the current tool/helpshows the full command set/quitexits the studio
The composer is now multiline with an expanded six-line visible input area. Enter sends, while Shift+Enter,
Alt+Enter, or Ctrl+J inserts a newline when the terminal exposes those keys distinctly.
For large context dumps, type ===== on its own line, paste your block, close with another =====, then press
Enter once to submit the full payload. While that delimiter block is open, Enter adds newlines instead of sending.
Tab and Shift+Tab now cycle file-path completions for /read, /open, and /workspace.
Planner replies, /write, and /run now stream model output into the transcript while the underlying CLI tool is
still running, so users can see progress live instead of waiting for one final blob.
The studio can also ask the active agent for an AI assessment of the current planning state. Run /advice to cache a
plain-English summary of:
- the problem statement the agent believes you are solving,
- whether the current plan state is clear or still fuzzy,
- what research or repo truth still needs to be gathered,
- and the best next move right now.
Before /write, a human must review and confirm:
- use
/reviewto get the checklist and file targets - use
/open allto open the planning docs in VS Code - use
/confirm-planto approve authored-plan refresh writes
Claude support is real, but it is not treated as interchangeable with Codex. The current non-interactive Claude path
uses plan mode for planner replies and acceptEdits with allowlisted local tools for pack-writing and execution.
If the Claude CLI is not installed locally, doctor, the studio, and run-next --agent claude all report that
honestly instead of falling back to a fake Claude path.
Augment support is wired to the documented auggie automation flags: --print, --quiet, --instruction-file,
--workspace-root, --rules, --allow-indexing, --wait-for-indexing, --max-turns, and --ask for planner-only
runs.
The defaults deliberately force workspace indexing and append srgical-specific Augment rules so the agent stays biased toward incremental planning, validated execution, and clear next-step handoffs. Session history is also left on so workspace iterations can accumulate inside Auggie instead of being treated as throwaway runs.
That means successful Augment execution still depends on a real Augment CLI install, an authenticated session such as
auggie login or AUGMENT_SESSION_AUTH, and whatever automation entitlements or local permission policies your
account requires.
- deepen the studio experience without weakening the terminal-first workflow
- keep multi-agent docs and validation honest as Claude and Augment runtime behavior get more live coverage
- expand release outputs from npm tarballs into standalone binaries and wrapper package-manager installers