gotta 0.38.0

Session-rooted CLI for remote discovery, evidence acquisition, durable operator memory, and actor-coordinated workflows

Gotta

gotta is a session-rooted CLI for remote discovery, evidence acquisition, durable operator memory, and actor-coordinated workflows.

The point is broader than retrieving one more document. gotta is for terminal work that would otherwise die in scrollback, browser tabs, or compacted model context. It externalizes enough evidence, state, provenance, and follow-up structure that a session can be reopened and continued from grounded context instead of reconstructed from memory.

At a glance:

• gotta search routes plain-text remote discovery into provider-native search surfaces
• gotta read materializes local or remote evidence into a durable session graph
• gotta session turns that graph into manifest, timeline, graph, leads, analyze, and scan views
• gotta notes, gotta logs, gotta oops, and gotta todo are live CLI surfaces over canonical JSONL state
• gotta actor coordinates sibling actors inside one shared session without flattening everything into one transcript

The public surface stays small:

gotta <plugin> ...

gotta ... is the canonical operator path. It binds or discovers the right session for the current context, hydrates the runtime environment, and dispatches the requested plugin with that session active. You do not need to cd into the session root first.

Why gotta

The name is a modal verb of necessity. Every subcommand reads as a natural English sentence expressing what needs to happen:

• gotta read — I have to read this
• gotta oops — I have to inspect or record this friction
• gotta session analyze — I have to analyze the session
• gotta want — I have to express what I want
• gotta todo append — I have to track this work

That composability is not accidental. Alternative names work for acquisition commands but break down everywhere else. glean read makes sense; glean oops does not. trace jira search reads well; trace want does not. Only a modal verb of necessity composes across the entire surface: evidence acquisition, friction capture, intent declaration, actor coordination, and session synthesis.

The semantics are also exact. The tool's thesis is that evidence must be externalized, context must survive compression, and friction must be recorded. "Gotta" is that necessity made literal.

Continuity Over Context Windows

Actors work in waves. They pull a few strong anchors, expand into adjacent evidence, synthesize the resulting web, and then eventually hit compression pressure. Context windows compact. Terminal history scrolls away. Thin summaries preserve headlines but lose the path that made those headlines trustworthy.

gotta is shaped around that boundary.

It externalizes the parts of the working context that should survive:

• the session itself as a durable working root
• canonical task, log, and friction state
• materialized evidence with native reopen handles
• synthesis surfaces such as manifest, timeline, graph, leads, analyze, and scan

That is why gotta is session-rooted rather than request-rooted. A request is ephemeral. A session can be reopened, inspected, extended, handed off, or compacted and rehydrated without losing the shape of the work.

The choice of medium is equally deliberate: native CLI working surfaces are one of the strongest places for serious actor work. They provide room to inspect, correlate, transform, act on, and resume real evidence. gotta leans into that by treating retrieval, memory, and action as part of the same working surface.

Why It Is Shaped This Way

gotta is deliberately native-first:

• prefer provider-native retrieval and mutation surfaces over ad hoc shell fallbacks
• keep durable truth in session state, not in terminal scrollback
• make every material read reopenable through native locators
• preserve enough provenance that compression leads to rehydration, not loss of working context

The normal working loop expands and compresses repeatedly:

1. retrieve one or two strong anchors through provider-native search or read
2. expand outward into adjacent evidence and materialize it into the session
3. compress that evidence web through session manifest, timeline, graph, leads, and analyze
4. record friction and continuity gaps in oops
5. refine the next retrieval wave from durable state instead of memory

state/oops.jsonl is the canonical friction log, and gotta oops is the live readable surface. It captures operator-visible seams: misleading contracts, interrupted continuations, coverage gaps, and workflow friction worth fixing.

Installation

If you do not already have uv, install it first from Astral's official installation guide.

Install with uv:

uv tool install gotta

That installs the gotta entrypoint.

If you want to try the CLI without installing it permanently:

uvx --from gotta gotta --help

gotta supports Python >=3.10.

If you are developing on the repo, sync a local uv-managed environment:

uv sync --python 3.10 --extra dev
./scripts/install-hooks

Installed entrypoints:

gotta ...

Development And Quality Gate

The blocking local gate is:

uv run pytest -q
uv run ruff check src tests
uv run ruff format --check src tests
uv run python -m vulture src tests --min-confidence 80

The repo-managed pre-commit hook is the default formatter path. It formats staged *.py files with uv run ruff format, re-stages them, and then runs uv run ruff check on those same files. Install it once per checkout with ./scripts/install-hooks. The study and release wrappers still keep ruff format --check as the hard gate. Partially staged Python files are rejected instead of being auto-formatted so the hook never smashes unstaged worktree edits into the index.

Structural pressure tools are part of the maintenance discipline:

uv run python -m radon cc src tests -s
uv run lizard src tests
uv run pytest --durations=15 -q
uv build --python 3.10 --clear
uvx twine check dist/*

pytest, ruff, and vulture are the correctness and hygiene gate. radon and lizard are not bug finders; they expose responsibility concentration and complexity hotspots so cleanup removes residue instead of burying it. The duration profile keeps performance and pathological tests visible.

For one-command study and maintenance passes, use:

./scripts/study
./scripts/study --deep
./scripts/study --types

./scripts/study runs the blocking gate, pressure tools, and any local study binaries already installed. Pressure-map tools such as radon and lizard are surfaced as advisory signals rather than correctness blockers. --deep adds architecture and semantic probes through import-linter and semgrep. --types adds a source-only pyright pass as a pressure map rather than a release gate.

Advanced Study Battery

This repo benefits from being studied in layers instead of file-by-file.

• pytest, ruff, vulture verify correctness and hygiene.
• radon, lizard, and pytest --durations expose responsibility and runtime concentration.
• cloc, universal-ctags, and ast-grep surface size, symbols, carrier types, and CLI grammar shape.
• import-linter checks explicit architectural contracts.
• semgrep codifies custom invariants such as environment reads and durable writes.
• pyan3 is useful for focused call-graph descent through orchestration or synthesis hubs.
• tree-sitter is a precise syntax substrate for structural queries once a grammar is configured intentionally.
• CodeQL is the heavyweight semantic audit path.
• scip-python and repomix produce machine-readable indexes and packed handoff artifacts for agent work.

On macOS, the most useful optional local binaries are usually installed with Homebrew:

brew install ast-grep cloc tree-sitter tree-sitter-python universal-ctags

Everything else in the study battery can be invoked through uv, uvx, or dedicated one-off tooling.

Configuration And Durable State

Durable provider config lives in GOTTA_CONFIG_FILE or, by default, under gotta's OS-native config directory:

• macOS: ~/Library/Application Support/gotta/gotta.toml
• Linux: ~/.config/gotta/gotta.toml
• Windows: %AppData%\\gotta\\gotta.toml

Example:

[providers.atlassian.env]
GOTTA_ATLASSIAN_OAUTH_CLIENT_ID = "your-client-id"
GOTTA_ATLASSIAN_OAUTH_CLIENT_SECRET = "your-client-secret"
GOTTA_ATLASSIAN_OAUTH_REDIRECT_URI = "http://localhost:8080/callback"
GOTTA_ATLASSIAN_TOOLSETS = "all"
GOTTA_JIRA_BASE_URL = "https://example.atlassian.net"
GOTTA_CONFLUENCE_BASE_URL = "https://example.atlassian.net/wiki"

[providers.google.env]
GOTTA_GOOGLE_OAUTH_CLIENT_ID = "your-client-id"
GOTTA_GOOGLE_OAUTH_CLIENT_SECRET = "your-client-secret"
GOTTA_GOOGLE_OAUTH_REDIRECT_URI = "http://localhost:8091/callback"

[providers.slack.env]
GOTTA_SLACK_WORKSPACE = "your-workspace"

[providers.grafana.env]
GOTTA_GRAFANA_BASE_URL = "https://grafana.example.com"
GOTTA_GRAFANA_SERVICE_ACCOUNT_TOKEN = "glsa_your_service_account_token"
# Optional when the token must be pinned to one org explicitly.
GOTTA_GRAFANA_ORG_ID = "1"

For canonical durable setup, use gotta config instead of relying on incidental side effects from operational commands. For Slack, the intended entrypoint is gotta config slack <workspace-or-link>, which persists the default workspace, can derive it from a workspace-hosted Slack permalink, and can also use a generic Slack link when one local workspace is already unambiguous. Ordinary Slack --workspace flags stay per-command; the low-level exact auth surface remains gotta slack auth --workspace <workspace>.

Durable OAuth state lands under gotta's OS-native state directory:

• macOS: ~/Library/Application Support/gotta/auth/
• Linux: ~/.local/state/gotta/auth/
• Windows: %LocalAppData%\\gotta\\auth\\

Canonical Session Model

gotta has two first-class session layouts:

• shared-topology sessions under gotta's OS-native data directory
• exact-root sessions scaffolded intentionally at one concrete path

Shared-topology sessions are the ambient default for stable interactive fingerprints such as Codex threads and terminal sessions:

<gotta data dir>/sessions/<session-id>/
  session.json
  content/
  actors/<actor-id>/

Examples:

• macOS: ~/Library/Application Support/gotta/sessions/<session-id>/
• Linux: ~/.local/share/gotta/sessions/<session-id>/

Each shared-topology session owns:

• session.json for shared session membership and actor metadata
• content/ for the shared evidence web and append-only manifest
• actors/<actor-id>/ for actor-local writable surfaces and state

Each actor root then owns its local working surfaces:

actors/<actor-id>/
  WANT.md
  GOAL.md
  state/
  content -> ../../content

Top-level fingerprint bindings live separately:

<gotta data dir>/bindings/<fingerprint>/
  binding.json
  root -> <concrete-session-scope>

The default private session for an unbound fingerprint is:

sessions/<fingerprint>/actors/<fingerprint>/

Exact-root sessions are the manual path when you intentionally want one workspace-local root with session metadata, actor surfaces, and content co-located:

gotta session init --session "$WS"
gotta session bind "$WS"

<exact-root>/
  WANT.md
  GOAL.md
  state/
  content/
  actors/<actor-id>/
  session.json          # appears once sibling actors are bound

The topology choice is intentional:

• shared-topology sessions are the canonical ambient/default path
• exact-root sessions are the explicit "this directory is the session" path
• gotta session bind accepts either a shared session id, an exact session root, or an explicit <session>/<actor> reference

Stored artifacts have two native reopen handles:

• artifact:<preferred-name>@<digest12>
• content:<sha256>

Both resolve through gotta read ..., so emitted evidence does not require manual manifest spelunking just to continue.

The content store is first-class:

• the filesystem under $WS/content is the durable evidence graph
• manifest.jsonl is the append-only invocation index
• identical bytes land in the same content object
• repeated fetches create additional timestamped evidence links

This is the continuity boundary in concrete form: the session keeps enough of the working set externalized that later synthesis, follow-up research, or downstream actions can be rebuilt from durable state instead of a lossy memory of prior interactions.

Retrieval And Materialization

Remote discovery and evidence acquisition are the canonical front door:

# Durable provider setup and readiness
gotta config slack https://demo.slack.com/archives/C12345678/p1773085070240949
gotta slack status --workspace demo
gotta slack workspaces
gotta slack auth --workspace demo

# Plain-text remote discovery
gotta search 'jira:retry budget'
gotta search 'github:service ownership'
gotta slack search "handoff failure"

# Provider-exact discovery and retrieval
gotta jira status
gotta github https://github.com/org/repo/commits/HEAD
gotta github https://github.com/org/repo/actions/runs/1234567890
gotta grafana status
gotta grafana datasources
gotta grafana search --type dash-db
gotta grafana search --type dash-db production-overview
gotta grafana search "production overview"
gotta grafana get demo-dashboard-uid
gotta grafana query --dashboard demo-dashboard-uid 'sum(app_up)'
gotta grafana query --dashboard 'https://grafana.example.com/d/demo-dashboard-uid/view?orgId=1&from=now-6h&to=now' 'sum(app_up)'
gotta grafana query --datasource prom-main 'rate(http_requests_total[5m])'
gotta confluence search "queue architecture"
gotta gdocs search "incident response"
gotta granola list --time-range last_30_days --limit 5
gotta granola transcript 11111111-1111-1111-1111-111111111111 --query latency
gotta granola search-transcript latency --time-range last_30_days

# Canonical acquisition into the evidence web
gotta read https://github.com/org/repo/blob/main/README.md
gotta read https://slack.com/docs/T12345678/D12345678

Stable interactive fingerprints like Codex threads and terminal sessions automatically adopt their deterministic default session on the first session-aware command that actually targets session state or acquires remote/provider evidence. Provider search, provider get, remote/provider read, and read-only session surfaces all auto-bootstrap there. Plain local reads stay sessionless, and explicit --actor targeting still requires an existing session because actor selection is session topology, not session creation. Fallback synthetic fingerprints remain conservative and can still use provider surfaces sessionlessly until a session is explicitly bound or created.

gotta read is the canonical acquisition entrypoint. It supports:

• local files and directories
• bounded local rereads through --head, --tail, and --section
• bounded remote/provider reads whose limits only trim what is shown while the full canonical payload still lands underneath in the evidence web
• provider URLs routed to the correct plugin
• canonical provider locators emitted by session surfaces
• stored content rereads by artifact or digest

This surface is broader than search. It is the acquisition layer for the session's evidence web. Slack threads and docs, GitHub pages and Actions runs, Jira issues, Google Docs and Drive files, shared-drive documents, Grafana dashboards, Confluence pages, and Granola notes all become reopenable session artifacts rather than transient terminal output.

Granola extends that same model to personal notes and transcripts through the local desktop session and Granola's APIs, so meeting notes and transcripts participate in the same durable evidence graph as repository, ticket, and chat artifacts.

Once that context is grounded, the same native surfaces support follow-on action. The goal is not only to recover information, but also to preserve enough working state that research, planning, and provider-native actions can happen against the same session context.

Remote Discovery

gotta search is the canonical plain-text remote discovery surface:

gotta search jira:Architecture
gotta search 'slack:ABC reboot'
gotta search 'github:SomeFunction ownership'
gotta search confluence:Architecture
gotta search 'jira:retry --literal-token'
gotta search 'github:search SomeFunction ownership'

The top-level verb already means search, so the routed target omits the subcommand by default. Explicit aliases like github:search foo are still tolerated, but github:foo is canonical. Top-level gotta search takes exactly one provider-qualified plain-text query string; it does not forward provider-specific search flags. Quote the full <provider>:<query> argument when the query contains spaces or flag-shaped text, and use provider-exact surfaces such as gotta slack search ..., gotta github search ..., or gotta confluence search ... when you need structured flags.

Session Synthesis Surfaces

gotta session is how the raw evidence web becomes navigable:

• session manifest
• session timeline
• session graph
• session leads
• session analyze
• session scan

Examples:

gotta session show
gotta session manifest --plugin jira
gotta session timeline --filter retry --limit 20
gotta session graph --filter jira --output mermaid
gotta session leads artifact:ticket.md@0123deadbeef
gotta session analyze --mode lineage --output mermaid
gotta session analyze --output markdown
gotta session scan "retry ownership"

These surfaces intentionally compress the evidence web without severing it:

• manifest summarizes what has been materialized
• timeline reconstructs chronology
• graph renders lineage and continuity
• leads extracts explicit next reads from existing artifacts
• analyze renders focused lineage and semantic synthesis directly from durable state
• scan searches projected artifact text across the materialized corpus

Where helpful, manifest, timeline, graph, and leads also surface top-N hotspots in both text and JSON so the densest actors, providers, plugins, and artifact kinds are visible without reading the entire payload.

session analyze always renders the requested --output format to stdout and no longer writes graph bundles or summary artifacts as a side effect. Redirect stdout when you want durable editor-visible output. Textual stdout is budgeted by default; pass --full-output to disable terminal budgeting. Successful operator surfaces emit a compact JSON receipt on stderr only when gotta has non-obvious side effects to report, such as truncation or artifact locators; pass --quiet to suppress informational stderr output. Raw Mermaid output requires --mode lineage or --mode semantic; use --output markdown for the combined two-graph bundle. Text and Markdown analysis outputs now start with a compact anchor shortlist plus lineage and lead previews before the deeper sections.

Empty manifest, graph, leads, and analyze output means the session has not materialized enough evidence yet.

After compaction pressure, these are the surfaces that make rehydration cheap. They preserve the shape of the work so the next pass starts from durable structure rather than a thin summary.

Session Coordination

When you want one explicit shared session, bind it and then rewrite the operator-owned charter surfaces inside the active actor root:

gotta session bind retry-review
gotta actor bind Claude
gotta want --stdin <<'EOF'
Queue retry context review.
EOF
gotta goal --stdin <<'EOF'
Build the execution charter for the current session from live context.
EOF

That yields one shared session plus one actor-local working root:

sessions/retry-review/
  session.json
  content/
  actors/<actor-id>/
    WANT.md
    GOAL.md
    state/
    content -> ../../content

Inside this topology:

• WANT.md and GOAL.md are actor-local intentional rewrites
• state/todo.jsonl is actor-local checklist truth and gotta todo is the live readable/mutable surface
• state/logs.jsonl is actor-local procedural/system trace and gotta logs is the live readable trace surface
• state/oops.jsonl is actor-local friction truth and gotta oops is the live readable friction surface
• state/notes.jsonl is actor-authored narration truth and gotta notes is the live readable narration surface
• sessions/<session-id>/content/ is shared across all actors bound into that session

Read-only session-rooted surfaces such as gotta oops, gotta logs, gotta todo, gotta want, gotta goal, gotta actor status, and gotta session show auto-bootstrap the deterministic default session for stable interactive fingerprints like Codex threads and terminal sessions. Fallback synthetic fingerprints still require either an existing bound session or an explicit --session <session-id>.

This split is deliberate:

• narrative framing stays editable
• operational truth stays append-only and durable
• the CLI renders readable output from canonical state on demand

In stable interactive contexts, the first session-aware gotta command implicitly scaffolds and adopts the deterministic session for that context. Future commands in that same context resolve there ambiently. gotta session init remains the manual exact-root path when you want to scaffold one specific root intentionally. For shared-topology sessions, other contexts should reuse the shared session id with --session <shared-session-id> or gotta session bind <shared-session-id>. Exact-root reuse with --session <session-root> or gotta session bind '<session-root>' is the low-level path for non-shared roots or intentionally reusing one concrete exact-root session scope.

Examples:

gotta want --stdin <<'EOF'
Queue retry context review.
EOF
gotta goal --stdin <<'EOF'
Trace retry handling from the first strong source anchor and keep the charter
current as evidence lands.
EOF
gotta todo append <<'EOF'
Inspect duplicate materializations in `gotta session analyze`.
EOF
gotta logs append <<'EOF'
Captured the first GitHub, Jira, and Confluence anchor set.
EOF
gotta oops append <<'EOF'
Direct fetch coverage should preserve the same continuity guarantees as search.
EOF

Shared Sessions

Each shared session owns its evidence web directly and carries nested actor-local session areas beneath it.

• shared session roots live at sessions/<session-id>/
• actor-local session areas live at sessions/<session-id>/actors/<actor-id>/
• the active fingerprint points at one actor-local root through bindings/<fingerprint>
• shared evidence lives under sessions/<session-id>/content/
• actor-local planning and lifecycle state stay local to each actor root
• actor targeting is explicit actor selection inside the current session, not path traversal under actors/

Examples:

gotta session bind retry-review
gotta actor bind Claude
ACTOR=<bound-actor-id>
gotta want --actor "$ACTOR" --stdin <<'EOF'
Trace retry ownership from the first strong source anchor.
EOF
gotta goal --actor "$ACTOR" --stdin <<'EOF'
Materialize the actor-local evidence contract before launch.
EOF
gotta actor launch "$ACTOR"
gotta notes show --actor "$ACTOR"
gotta todo extend --actor "$ACTOR" <<'EOF'
- Compare retry behavior across the relevant design docs.
EOF

Important invariants:

• gotta session bind ... switches the active fingerprint binding to one concrete session scope
• binding a shared session id joins that shared session through the caller's own actor root; binding an exact session root reuses that exact root; binding an explicit <session>/<actor> reference reuses that exact actor root
• gotta actor bind ... binds sibling actor sessions inside that shared session but does not launch them
• actor-local WANT.md and GOAL.md are seeded placeholders that must be rewritten before launch with gotta want --actor <actor> ... and gotta goal --actor <actor> ...
• actor-local checklist state is seeded automatically and surfaced through gotta todo
• actor notes are the canonical actor-authored narration surface
• short one-line notes are valid; polished synthesis notes are optional
• actor evidence often lands in the shared manifest, timeline, and graph before notes fully catch up

The important property is continuity under delegation. Actors can branch, gather evidence, and rejoin the shared working set without flattening everything into a single chat transcript.

oops As Canonical Alignment

gotta oops is a first-class operator surface.

Bare gotta oops shows the friction ledger. gotta oops show is the explicit read form. gotta oops append ... and gotta oops extend ... write new entries. Bare prose, real piped stdin, --stdin, and --from-file still imply append when no read action is named.

Use it to record:

• incomplete or misleading contracts
• native surfaces that should have been followable but were not
• provider coverage limits that materially shaped the working path
• workflow friction that forced avoidable detours

The point is not merely bug tracking. The point is preserving operator-visible misalignment in canonical shared state so the tool can be refined from observed behavior rather than taste or memory.

Key Concepts

These terms appear throughout gotta and its documentation:

• session — a durable working root that survives context loss. Sessions own a content store, actor roots, and synthesis surfaces. They can be reopened, extended, handed off, or compacted and rehydrated.

• actor — a named participant in a session. Each actor has its own charter files (WANT.md, GOAL.md), canonical state, and lifecycle. gotta notes is actor-authored narration, gotta logs is procedural/system trace, gotta oops is friction, and gotta todo is the checklist surface. Actors can be human operators, AI agents, or automated workflows.

• fingerprint — the context identity that binds the current terminal, thread, or environment to one concrete session scope: an exact root, a shared session through the caller's actor root, or an explicit actor root. Fingerprints are derived from terminal session IDs, Codex thread IDs, or explicit bindings.

• materialization — the act of capturing command output as a durable, content-addressed artifact in the session's content store. Materialized evidence has a SHA-256 digest and can be reopened by locator.

• canonical locator — a provider-normalized reference for a materialized artifact. Examples: github:org/repo/blob/main/README.md, jira:PROJ-123, slack:C01234/p1234567890. Locators are emitted when evidence lands and can be followed with gotta read.

• artifact locator — a session-relative reference to stored content. Format: artifact:<preferred-name>@<digest12>. Resolves through gotta read without requiring manifest spelunking.

• content locator — a digest-based reference to stored bytes. Format: content:<sha256>. Two identical fetches share the same content object.

• evidence web — the accumulated set of materialized artifacts, their metadata, and the relationships between them. The web grows through retrieval waves and can be navigated through synthesis surfaces.

• synthesis surface — a compressed, navigable view over the evidence web. manifest summarizes what has been materialized. timeline reconstructs chronology. graph renders lineage. leads extracts followable references. analyze renders focused synthesis directly from durable state. scan searches projected artifact text across the corpus.

• friction — operator-visible misalignment captured in oops. Not bug tracking. Friction records seams: misleading contracts, continuity gaps, workflow detours. The canonical log is state/oops.jsonl; gotta oops is the readable surface.

• projection — an on-demand terminal render from canonical state or a provider/content transform. Projections are no longer seeded as live root files; the CLI is the live readable surface.

• rehydration — recovering prior working state from durable session artifacts after context has been compressed. The synthesis surfaces make rehydration cheap: they preserve the shape of the work so the next pass starts from structure rather than a thin summary.

Plugin Architecture

gotta uses two entry-point groups:

• gotta.plugins for top-level plugins
• gotta.ask for optional ask-family extensions

Core is a PEP 420 namespace package. Plugins can live in separate distributions and still contribute modules under the shared gotta namespace.

Core ships these top-level plugins:

• ask
• actor
• confluence
• gdocs
• gdrive
• goal
• grafana
• granola
• github
• gsheets
• jira
• logs
• notes
• oops
• read
• search
• session
• slack
• todo
• want

read is the URL-shaped dispatcher. It routes recognized targets through installed provider plugins by asking those plugins whether they own the target.

Ask Extensions

gotta ask is a generic host for separately installed ask-family extensions. Ask surfaces are provided by extensions that register the gotta.ask entry-point group.

An ask extension registers the gotta.ask entry-point group and then becomes available as:

gotta ask <surface> ...

Use gotta ask --help-all to inspect installed ask surfaces recursively.

Contributing And Release Discipline

The project is maintained with a few deliberate rules:

• remove residue instead of layering over it
• prefer tool-observable truth to conversational assumption
• keep friction canonical in oops
• treat complexity as measurable pressure, not just aesthetic discomfort
• preserve session continuity and reopenable evidence paths

If you contribute:

• run the full quality gate before proposing changes
• keep examples and fixtures generic and public
• preserve the native evidence-first workflow
• prefer behavior-level cleanup over compatibility ballast

See CONTRIBUTING.md for the repository baseline.

Release

Build and validate artifacts with uv:

./scripts/release prepare patch
./scripts/release prepare minor
./scripts/release publish
./scripts/release patch
./scripts/release minor

The script is the canonical release path. It runs ./scripts/study on the unbumped tree, then bumps the version with uv, builds the wheel and sdist, validates them with twine check, smoke-installs the wheel on Python 3.10, commits the release metadata, and then either stops for review or pushes and publishes. The blocking correctness gate still comes from pytest, ruff, vulture, build, and packaging validation; pressure-map output from ./scripts/study is surfaced for review without blocking the release candidate. prepare creates the reviewable release commit without pushing. publish validates the current prepared version and then pushes main, publishes to PyPI, and waits for public propagation. One-shot patch and minor still do the full flow in one pass.

It reads the PyPI token from ~/.pypirc under [pypi].password.