tier-ctx 1.2.1__tar.gz

tier_ctx-1.2.1/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 000jd
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

tier_ctx-1.2.1/PKG-INFO

@@ -0,0 +1,273 @@
+Metadata-Version: 2.4
+Name: tier-ctx
+Version: 1.2.1
+Summary: TIER: layered context, repository graphs, code slicing, and agent memory
+Author: 000jd
+License: MIT
+Requires-Python: >=3.10
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Provides-Extra: ui
+Requires-Dist: rich>=13.0; extra == "ui"
+Provides-Extra: mcp
+Requires-Dist: mcp>=2.0; extra == "mcp"
+Provides-Extra: treesitter
+Requires-Dist: tree-sitter-language-pack>=0.7.0; extra == "treesitter"
+Provides-Extra: all
+Requires-Dist: rich>=13.0; extra == "all"
+Requires-Dist: mcp>=2.0; extra == "all"
+Requires-Dist: tree-sitter-language-pack>=0.7.0; extra == "all"
+Dynamic: license-file
+
+<div align="center">
+
+# TIER
+
+**Layered context for AI coding agents: a 4-layer document and repository format, a code graph
+indexer, and a zero-dependency MCP server.**
+
+[![CI](https://github.com/000jd/tier/actions/workflows/ci.yml/badge.svg)](.github/workflows/ci.yml)
+[![PyPI](https://img.shields.io/badge/pypi-tier--ctx-blue)](https://pypi.org/project/tier-ctx/)
+[![Python 3.10+](https://img.shields.io/badge/python-3.10%2B-3776AB?logo=python&logoColor=white)](pyproject.toml)
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue)](LICENSE)
+
+</div>
+
+---
+
+`tier` keeps an agent from loading whole repositories and documents into its context. Every
+document carries a small always-loadable index, one-line abstracts, structured bodies with symbol
+signatures and dependency edges, and exact source-line ranges it can fetch on demand. A hybrid
+retriever routes a question to the right section, an incremental repository map keeps fingerprints
+and line ranges current, and a reverse-dependency graph answers what an edit affects.
+
+The core is pure Python standard library: `dependencies = []`, no network calls, no telemetry.
+It runs entirely offline on Python 3.10 and later. Optional extras add terminal styling, the
+official MCP SDK adapter, and Tree-sitter grammars.
+
+## Installation
+
+Zero-clone installs, in order of least commitment:
+
+```sh
+uvx --from tier-ctx tier install     # run the installer once, leave nothing installed
+uv tool install tier-ctx             # put `tier`, `tier-mcp`, and `tier-mcp-sdk` on PATH
+pipx install tier-ctx                # the same, where pipx is preferred
+```
+
+Until the PyPI release is live, install from Git:
+
+```sh
+uv tool install git+https://github.com/000jd/tier
+uv tool install "tier-ctx[all] @ git+https://github.com/000jd/tier"   # + rich UI, MCP SDK
+```
+
+From a checkout, inside a virtual environment:
+
+```sh
+git clone https://github.com/000jd/tier && cd tier
+python3 -m venv .venv && . .venv/bin/activate
+pip install -e ".[all]"
+```
+
+> The distribution is named `tier-ctx`; the import package and console scripts are `tier`,
+> `tier-mcp`, and `tier-mcp-sdk`. The bare PyPI name `tier` belongs to an unrelated package.
+
+## Example
+
+An end-to-end session on a small project — install once, then one call per question:
+
+````console
+$ tier install --clients claude-code --yes
+Target: /home/me/myproject
+Clients: claude-code
+cache (.git/info/tier): /home/me/myproject/.git/info/tier/repo.tier (created)
+Tokens: raw source 197 -> L0 130 (34.0% saved)
+Instructions: /home/me/myproject/CLAUDE.md, /home/me/myproject/AGENTS.md
+
+$ tier context . "how does the retry backoff work" --budget-tokens 800
+## Context: how does the retry backoff work
+
+@src-http-client.py src/http/client.py
+? HTTP client with retry and backoff.; symbols: backoff_delay, request
+|symbols[3]{kind,name,signature,lines}
+function,backoff_delay,def backoff_delay(attempt: int) -> float,7-9
+function,request,"def request(url: str, retries: int=DEFAULT_RETRIES)",12-19
+function,_send,def _send(url: str),22-23
+|calls[1]{caller,callee,module}
+request,backoff_delay,.
+
+@src-http-client.py#backoff_delay [L7-L9]
+```python
+def backoff_delay(attempt: int) -> float:
+    """Exponential backoff, capped at 30 seconds."""
+    return min(2 ** attempt, 30.0)
+```
+
+## Impact: @src-http-client.py#backoff_delay
+Direct dependents: @src-api-client.py
+Transitive dependents: @src-api-client.py
+Symbol callers: src/http/client.py#request [12-19]
+
+$ tier impact src/http/client.py#backoff_delay
+Impact target: @src-http-client.py
+Direct dependents: @src-api-client.py
+Transitive dependents: @src-api-client.py
+Dependencies: none
+Transitive dependencies: none
+Symbol: src/http/client.py#backoff_delay
+Symbol callers: src/http/client.py#request [12-19]
+Symbol calls: none
+
+$ tier slice src/http/client.py backoff_delay
+@src-http-client.py#backoff_delay [L7-L9]
+```python
+def backoff_delay(attempt: int) -> float:
+    """Exponential backoff, capped at 30 seconds."""
+    return min(2 ** attempt, 30.0)
+```
+````
+
+## How it works
+
+TIER splits every document into four layers, and an agent reads only the ones it needs.
+
+- **L0 — index.** Document metadata, global rules, and one line per section carrying that
+  section's token cost and rule count. Always loaded; a few hundred tokens.
+- **L1 — abstracts.** A one-line `?` summary per section saying *when* to read it, folded into
+  the L0 index.
+- **L2 — bodies.** Section contents; for codebases, a `|symbols` table with exact
+  `[start-end]` line ranges and `>` edges to local dependencies.
+- **L3 — slices.** Exact source lines for one symbol, fetched by name.
+
+`tier_context` collapses the whole loop into one call: resolve or build the repository index,
+route the query with the hybrid retriever, match symbol names, then pack L2 bodies, L3 slices,
+and a dependency summary until the token budget is spent. The granular commands remain available
+when you need a specific section id or symbol.
+
+## Architecture highlights
+
+**Header-once tables and indented key/value trees.** A record set is a header plus exactly N CSV
+rows, and nested data is a two-space-indented tree. Both keep the schema and drop the repeated
+punctuation that JSON and Markdown tables spend tokens on.
+
+**Deterministic line-start sigils.** Every line is classified by its first character: `%` header,
+`@` section, `?` abstract, `!` rule, `=` key/value, `|` counted table, `>` reference, `$` command.
+Fenced code, untrusted blocks, and table rows are opaque, so no sigil inside them is reinterpreted.
+
+**Untrusted content guards.** Text from the outside world is wrapped in `<<<source #guard ...
+>>>guard`. The guards are chosen so that the payload cannot contain its own terminator, which makes
+embedded instructions, fake headers, and rule sigils inert data.
+
+**Stateful session deduplication.** A section or source slice already delivered in the conversation
+is replaced by a one-line stub on the next request. `sync` evicts exactly the sections whose files
+changed, and their slices.
+
+**Symbol-level blast radius.** Repository maps record `>` edges between modules and a `|calls`
+table of caller/callee pairs. `tier impact` reports file-level importers and, given
+`module#symbol`, the exact functions that call it. `tier_context` filters `|calls` tables to the
+1-hop neighbourhood of the sections it returns; pass `include_calls="all"` to see every edge.
+
+**Invisible repository cache.** The map is a derived artifact, so it lives in
+`.git/info/tier/repo.tier` inside a worktree (which Git ignores) or in a user cache directory
+outside it. `.git` is followed when it is a pointer file rather than a directory, so linked
+worktrees and submodules resolve to their own git directory. A tracked `repo.tier` is still
+honoured when one exists. Nothing is written into your working tree, so `git status` stays clean.
+
+**Hybrid search with precomputed IDF.** BM25-style term scoring over split identifiers, a compact
+synonym map, character 3-gram fallback for typos, and navigation priors. Document frequencies are
+computed once per query, which keeps a search over 800 sections near 200 ms with no index to build.
+
+## Performance
+
+Measured with the `o200k_base` tokenizer. These are token counts and retrieval hit rates, not LLM
+answer accuracy. Methodology and reproduction steps are in
+[`benchmarks/README.md`](benchmarks/README.md).
+
+| Workload | Measure | Result |
+|---|---|---|
+| Repository overview | raw source → full map → L0 → L0 + one module | 72,859 → 14,375 → 828 → 1,784 tokens (−98%) |
+| End-to-end 4-layer workflow | index, find, L2, L3 slices, impact + sync | 5,003 tokens (−93.1%), 3/3 exact slices |
+| 36-file polyglot repository | surgical L0 + target L2 + 7 slices | 3,008 vs 8,249 tokens (−64%), 116/116 symbols |
+| 197-section blind retrieval | 15 queries | 15/15 hit@1, mean top-1 fetch 134 tokens |
+
+[`benchmarks/BENCH.txt`](benchmarks/BENCH.txt) carries the full Track A–I breakdown, including the
+record-encoding and nested-config tracks, the per-query retrieval logs, and the raw output behind
+every row above.
+
+## Supported clients
+
+`tier install` configures 25 agent clients across JSON, TOML, YAML, and `SKILL.md` schemas. It
+backs up every file it touches, merges instead of overwriting, supports `--dry-run` and
+`--uninstall`, and validates the result with `tier doctor`.
+
+<details>
+<summary><b>Client list and config paths</b></summary>
+
+| Client | Scope | Config |
+|---|---|---|
+| Claude Code | project / user | `.mcp.json` + `.claude/settings.json`, `~/.claude.json` |
+| Claude Desktop | user | `claude_desktop_config.json` |
+| Command Code | project / user | `.commandcode/mcp.json`, `~/.commandcode/mcp.json` |
+| OpenCode | project / global | `opencode.json`, `~/.config/opencode/opencode.json` |
+| OpenAI Codex CLI | project / user | `.codex/config.toml`, `~/.codex/config.toml` |
+| Gemini CLI | project / user | `.gemini/settings.json` + `GEMINI.md` |
+| Cursor | project / global | `.cursor/mcp.json`, `~/.cursor/mcp.json` |
+| VS Code Copilot | project | `.vscode/mcp.json` |
+| Windsurf | user | `~/.codeium/windsurf/mcp_config.json` |
+| Roo Code / Cline | project | `.roo/mcp.json` |
+| Trae | project | `.trae/mcp.json` |
+| JetBrains Junie | project / user | `.junie/mcp/mcp.json`, `~/.junie/mcp/mcp.json` |
+| Kiro | project | `.kiro/settings/mcp.json` |
+| Warp | user | `~/.warp/.mcp.json` |
+| Zed | user | `~/.config/zed/settings.json` (`context_servers`) |
+| Continue.dev | project | `.continue/mcpServers/tier.yaml` |
+| Agent Skills | project / user | `.agents/skills/tier/SKILL.md` |
+
+Any other client can be configured with `--custom-config PATH --schema NAME`, and
+`tier config-snippet` prints ready-to-copy JSON, TOML, YAML, and `SKILL.md` snippets plus native
+CLI one-liners.
+
+</details>
+
+<details>
+<summary><b>Optional extras</b></summary>
+
+```sh
+uv tool install "tier-ctx[ui] @ git+https://github.com/000jd/tier"          # rich terminal UI
+uv tool install "tier-ctx[mcp] @ git+https://github.com/000jd/tier"         # official MCP SDK
+uv tool install "tier-ctx[treesitter] @ git+https://github.com/000jd/tier"  # multi-language AST
+uv tool install "tier-ctx[all] @ git+https://github.com/000jd/tier"
+```
+
+- `[ui]` adds `rich` rendering for `tier install` and `tier doctor`.
+- `[mcp]` adds the official SDK adapter (`tier-mcp-sdk`), Streamable HTTP, and MCP Inspector.
+- `[treesitter]` adds Tree-sitter grammars for TypeScript, JavaScript, Go, Rust, C, C++, Java,
+  and Ruby. Without it the stdlib extractors are used, and repository maps are byte-identical.
+
+</details>
+
+## Documentation
+
+- [`docs/spec.md`](docs/spec.md) — the format, parser rules, and retrieval protocol
+- [`ARCHITECTURE.md`](ARCHITECTURE.md) — module boundaries, layer flow, and invariants
+- [`SKILL.md`](SKILL.md) — the canonical Agent Skill: entry point, CLI and MCP tables, working rules
+- [`docs/cli.md`](docs/cli.md) — every command and flag
+- [`docs/mcp.md`](docs/mcp.md) — the 12 MCP tools, one-turn workflow, and cache behaviour
+- [`benchmarks/README.md`](benchmarks/README.md) — workload definitions and methodology
+- [`CONTRIBUTING.md`](CONTRIBUTING.md) — development setup and extension guides
+- [`CHANGELOG.md`](CHANGELOG.md) — release history
+
+## Limits
+
+Only token counts, retrieval identifiers, and exact-slice coverage are measured — not LLM answer
+accuracy. Claude's tokenizer is not public, so counts use OpenAI `o200k` via a persistent Node
+worker when available, and a characters/4 estimate otherwise. `tier hook` treats its token budget
+as a contract: it skips a payload only when it is provably within budget, otherwise it counts
+exactly and verifies the output. `tier sync` trusts recorded mtime and size as a cache; use
+`tier sync --rehash` after tooling that restores timestamps. JS/TS/Go/Rust symbol extraction is
+heuristic without the `[treesitter]` extra.
+
+## License
+
+MIT © 2026 000jd. See [LICENSE](LICENSE).

tier_ctx-1.2.1/README.md

@@ -0,0 +1,252 @@
+<div align="center">
+
+# TIER
+
+**Layered context for AI coding agents: a 4-layer document and repository format, a code graph
+indexer, and a zero-dependency MCP server.**
+
+[![CI](https://github.com/000jd/tier/actions/workflows/ci.yml/badge.svg)](.github/workflows/ci.yml)
+[![PyPI](https://img.shields.io/badge/pypi-tier--ctx-blue)](https://pypi.org/project/tier-ctx/)
+[![Python 3.10+](https://img.shields.io/badge/python-3.10%2B-3776AB?logo=python&logoColor=white)](pyproject.toml)
+[![License: MIT](https://img.shields.io/badge/license-MIT-blue)](LICENSE)
+
+</div>
+
+---
+
+`tier` keeps an agent from loading whole repositories and documents into its context. Every
+document carries a small always-loadable index, one-line abstracts, structured bodies with symbol
+signatures and dependency edges, and exact source-line ranges it can fetch on demand. A hybrid
+retriever routes a question to the right section, an incremental repository map keeps fingerprints
+and line ranges current, and a reverse-dependency graph answers what an edit affects.
+
+The core is pure Python standard library: `dependencies = []`, no network calls, no telemetry.
+It runs entirely offline on Python 3.10 and later. Optional extras add terminal styling, the
+official MCP SDK adapter, and Tree-sitter grammars.
+
+## Installation
+
+Zero-clone installs, in order of least commitment:
+
+```sh
+uvx --from tier-ctx tier install     # run the installer once, leave nothing installed
+uv tool install tier-ctx             # put `tier`, `tier-mcp`, and `tier-mcp-sdk` on PATH
+pipx install tier-ctx                # the same, where pipx is preferred
+```
+
+Until the PyPI release is live, install from Git:
+
+```sh
+uv tool install git+https://github.com/000jd/tier
+uv tool install "tier-ctx[all] @ git+https://github.com/000jd/tier"   # + rich UI, MCP SDK
+```
+
+From a checkout, inside a virtual environment:
+
+```sh
+git clone https://github.com/000jd/tier && cd tier
+python3 -m venv .venv && . .venv/bin/activate
+pip install -e ".[all]"
+```
+
+> The distribution is named `tier-ctx`; the import package and console scripts are `tier`,
+> `tier-mcp`, and `tier-mcp-sdk`. The bare PyPI name `tier` belongs to an unrelated package.
+
+## Example
+
+An end-to-end session on a small project — install once, then one call per question:
+
+````console
+$ tier install --clients claude-code --yes
+Target: /home/me/myproject
+Clients: claude-code
+cache (.git/info/tier): /home/me/myproject/.git/info/tier/repo.tier (created)
+Tokens: raw source 197 -> L0 130 (34.0% saved)
+Instructions: /home/me/myproject/CLAUDE.md, /home/me/myproject/AGENTS.md
+
+$ tier context . "how does the retry backoff work" --budget-tokens 800
+## Context: how does the retry backoff work
+
+@src-http-client.py src/http/client.py
+? HTTP client with retry and backoff.; symbols: backoff_delay, request
+|symbols[3]{kind,name,signature,lines}
+function,backoff_delay,def backoff_delay(attempt: int) -> float,7-9
+function,request,"def request(url: str, retries: int=DEFAULT_RETRIES)",12-19
+function,_send,def _send(url: str),22-23
+|calls[1]{caller,callee,module}
+request,backoff_delay,.
+
+@src-http-client.py#backoff_delay [L7-L9]
+```python
+def backoff_delay(attempt: int) -> float:
+    """Exponential backoff, capped at 30 seconds."""
+    return min(2 ** attempt, 30.0)
+```
+
+## Impact: @src-http-client.py#backoff_delay
+Direct dependents: @src-api-client.py
+Transitive dependents: @src-api-client.py
+Symbol callers: src/http/client.py#request [12-19]
+
+$ tier impact src/http/client.py#backoff_delay
+Impact target: @src-http-client.py
+Direct dependents: @src-api-client.py
+Transitive dependents: @src-api-client.py
+Dependencies: none
+Transitive dependencies: none
+Symbol: src/http/client.py#backoff_delay
+Symbol callers: src/http/client.py#request [12-19]
+Symbol calls: none
+
+$ tier slice src/http/client.py backoff_delay
+@src-http-client.py#backoff_delay [L7-L9]
+```python
+def backoff_delay(attempt: int) -> float:
+    """Exponential backoff, capped at 30 seconds."""
+    return min(2 ** attempt, 30.0)
+```
+````
+
+## How it works
+
+TIER splits every document into four layers, and an agent reads only the ones it needs.
+
+- **L0 — index.** Document metadata, global rules, and one line per section carrying that
+  section's token cost and rule count. Always loaded; a few hundred tokens.
+- **L1 — abstracts.** A one-line `?` summary per section saying *when* to read it, folded into
+  the L0 index.
+- **L2 — bodies.** Section contents; for codebases, a `|symbols` table with exact
+  `[start-end]` line ranges and `>` edges to local dependencies.
+- **L3 — slices.** Exact source lines for one symbol, fetched by name.
+
+`tier_context` collapses the whole loop into one call: resolve or build the repository index,
+route the query with the hybrid retriever, match symbol names, then pack L2 bodies, L3 slices,
+and a dependency summary until the token budget is spent. The granular commands remain available
+when you need a specific section id or symbol.
+
+## Architecture highlights
+
+**Header-once tables and indented key/value trees.** A record set is a header plus exactly N CSV
+rows, and nested data is a two-space-indented tree. Both keep the schema and drop the repeated
+punctuation that JSON and Markdown tables spend tokens on.
+
+**Deterministic line-start sigils.** Every line is classified by its first character: `%` header,
+`@` section, `?` abstract, `!` rule, `=` key/value, `|` counted table, `>` reference, `$` command.
+Fenced code, untrusted blocks, and table rows are opaque, so no sigil inside them is reinterpreted.
+
+**Untrusted content guards.** Text from the outside world is wrapped in `<<<source #guard ...
+>>>guard`. The guards are chosen so that the payload cannot contain its own terminator, which makes
+embedded instructions, fake headers, and rule sigils inert data.
+
+**Stateful session deduplication.** A section or source slice already delivered in the conversation
+is replaced by a one-line stub on the next request. `sync` evicts exactly the sections whose files
+changed, and their slices.
+
+**Symbol-level blast radius.** Repository maps record `>` edges between modules and a `|calls`
+table of caller/callee pairs. `tier impact` reports file-level importers and, given
+`module#symbol`, the exact functions that call it. `tier_context` filters `|calls` tables to the
+1-hop neighbourhood of the sections it returns; pass `include_calls="all"` to see every edge.
+
+**Invisible repository cache.** The map is a derived artifact, so it lives in
+`.git/info/tier/repo.tier` inside a worktree (which Git ignores) or in a user cache directory
+outside it. `.git` is followed when it is a pointer file rather than a directory, so linked
+worktrees and submodules resolve to their own git directory. A tracked `repo.tier` is still
+honoured when one exists. Nothing is written into your working tree, so `git status` stays clean.
+
+**Hybrid search with precomputed IDF.** BM25-style term scoring over split identifiers, a compact
+synonym map, character 3-gram fallback for typos, and navigation priors. Document frequencies are
+computed once per query, which keeps a search over 800 sections near 200 ms with no index to build.
+
+## Performance
+
+Measured with the `o200k_base` tokenizer. These are token counts and retrieval hit rates, not LLM
+answer accuracy. Methodology and reproduction steps are in
+[`benchmarks/README.md`](benchmarks/README.md).
+
+| Workload | Measure | Result |
+|---|---|---|
+| Repository overview | raw source → full map → L0 → L0 + one module | 72,859 → 14,375 → 828 → 1,784 tokens (−98%) |
+| End-to-end 4-layer workflow | index, find, L2, L3 slices, impact + sync | 5,003 tokens (−93.1%), 3/3 exact slices |
+| 36-file polyglot repository | surgical L0 + target L2 + 7 slices | 3,008 vs 8,249 tokens (−64%), 116/116 symbols |
+| 197-section blind retrieval | 15 queries | 15/15 hit@1, mean top-1 fetch 134 tokens |
+
+[`benchmarks/BENCH.txt`](benchmarks/BENCH.txt) carries the full Track A–I breakdown, including the
+record-encoding and nested-config tracks, the per-query retrieval logs, and the raw output behind
+every row above.
+
+## Supported clients
+
+`tier install` configures 25 agent clients across JSON, TOML, YAML, and `SKILL.md` schemas. It
+backs up every file it touches, merges instead of overwriting, supports `--dry-run` and
+`--uninstall`, and validates the result with `tier doctor`.
+
+<details>
+<summary><b>Client list and config paths</b></summary>
+
+| Client | Scope | Config |
+|---|---|---|
+| Claude Code | project / user | `.mcp.json` + `.claude/settings.json`, `~/.claude.json` |
+| Claude Desktop | user | `claude_desktop_config.json` |
+| Command Code | project / user | `.commandcode/mcp.json`, `~/.commandcode/mcp.json` |
+| OpenCode | project / global | `opencode.json`, `~/.config/opencode/opencode.json` |
+| OpenAI Codex CLI | project / user | `.codex/config.toml`, `~/.codex/config.toml` |
+| Gemini CLI | project / user | `.gemini/settings.json` + `GEMINI.md` |
+| Cursor | project / global | `.cursor/mcp.json`, `~/.cursor/mcp.json` |
+| VS Code Copilot | project | `.vscode/mcp.json` |
+| Windsurf | user | `~/.codeium/windsurf/mcp_config.json` |
+| Roo Code / Cline | project | `.roo/mcp.json` |
+| Trae | project | `.trae/mcp.json` |
+| JetBrains Junie | project / user | `.junie/mcp/mcp.json`, `~/.junie/mcp/mcp.json` |
+| Kiro | project | `.kiro/settings/mcp.json` |
+| Warp | user | `~/.warp/.mcp.json` |
+| Zed | user | `~/.config/zed/settings.json` (`context_servers`) |
+| Continue.dev | project | `.continue/mcpServers/tier.yaml` |
+| Agent Skills | project / user | `.agents/skills/tier/SKILL.md` |
+
+Any other client can be configured with `--custom-config PATH --schema NAME`, and
+`tier config-snippet` prints ready-to-copy JSON, TOML, YAML, and `SKILL.md` snippets plus native
+CLI one-liners.
+
+</details>
+
+<details>
+<summary><b>Optional extras</b></summary>
+
+```sh
+uv tool install "tier-ctx[ui] @ git+https://github.com/000jd/tier"          # rich terminal UI
+uv tool install "tier-ctx[mcp] @ git+https://github.com/000jd/tier"         # official MCP SDK
+uv tool install "tier-ctx[treesitter] @ git+https://github.com/000jd/tier"  # multi-language AST
+uv tool install "tier-ctx[all] @ git+https://github.com/000jd/tier"
+```
+
+- `[ui]` adds `rich` rendering for `tier install` and `tier doctor`.
+- `[mcp]` adds the official SDK adapter (`tier-mcp-sdk`), Streamable HTTP, and MCP Inspector.
+- `[treesitter]` adds Tree-sitter grammars for TypeScript, JavaScript, Go, Rust, C, C++, Java,
+  and Ruby. Without it the stdlib extractors are used, and repository maps are byte-identical.
+
+</details>
+
+## Documentation
+
+- [`docs/spec.md`](docs/spec.md) — the format, parser rules, and retrieval protocol
+- [`ARCHITECTURE.md`](ARCHITECTURE.md) — module boundaries, layer flow, and invariants
+- [`SKILL.md`](SKILL.md) — the canonical Agent Skill: entry point, CLI and MCP tables, working rules
+- [`docs/cli.md`](docs/cli.md) — every command and flag
+- [`docs/mcp.md`](docs/mcp.md) — the 12 MCP tools, one-turn workflow, and cache behaviour
+- [`benchmarks/README.md`](benchmarks/README.md) — workload definitions and methodology
+- [`CONTRIBUTING.md`](CONTRIBUTING.md) — development setup and extension guides
+- [`CHANGELOG.md`](CHANGELOG.md) — release history
+
+## Limits
+
+Only token counts, retrieval identifiers, and exact-slice coverage are measured — not LLM answer
+accuracy. Claude's tokenizer is not public, so counts use OpenAI `o200k` via a persistent Node
+worker when available, and a characters/4 estimate otherwise. `tier hook` treats its token budget
+as a contract: it skips a payload only when it is provably within budget, otherwise it counts
+exactly and verifies the output. `tier sync` trusts recorded mtime and size as a cache; use
+`tier sync --rehash` after tooling that restores timestamps. JS/TS/Go/Rust symbol extraction is
+heuristic without the `[treesitter]` extra.
+
+## License
+
+MIT © 2026 000jd. See [LICENSE](LICENSE).

tier_ctx-1.2.1/pyproject.toml

@@ -0,0 +1,31 @@
+[build-system]
+requires = ["setuptools>=61"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "tier-ctx"
+version = "1.2.1"
+description = "TIER: layered context, repository graphs, code slicing, and agent memory"
+readme = "README.md"
+requires-python = ">=3.10"
+authors = [{ name = "000jd" }]
+license = { text = "MIT" }
+dependencies = []
+
+[project.optional-dependencies]
+ui = ["rich>=13.0"]
+mcp = ["mcp>=2.0"]
+treesitter = ["tree-sitter-language-pack>=0.7.0"]
+all = ["rich>=13.0", "mcp>=2.0", "tree-sitter-language-pack>=0.7.0"]
+
+[project.scripts]
+tier = "tier.cli:main"
+tier-mcp = "tier.mcp.server:main"
+tier-mcp-sdk = "tier.mcp.sdk_server:main"
+
+[tool.setuptools]
+package-dir = {"" = "src"}
+
+[tool.setuptools.packages.find]
+where = ["src"]
+include = ["tier*"]

tier_ctx-1.2.1/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

tier_ctx-1.2.1/src/tier/__init__.py

@@ -0,0 +1,53 @@
+"""TIER: layered context, repository graphs, code slicing, and agent memory."""
+from __future__ import annotations
+
+from .models import (
+    VERSION, Doc, Section, SIGIL_CHARS, SEC_RE, HDR_RE, TABLE_RE, DATA_OPEN_RE,
+    RULE_START, STOP, SHELL_LANGS, TABLE_MODES, _W, _atomic_write, _clip, _inside,
+    _read_text, _table_count,
+)
+from .tokens import count_tokens
+from .parser import (
+    _csv_cells, _data_closes, _data_open, _kv_split, _live_lines, _load, _parse_kv,
+    _render_table, _repo_file_meta, _table_rows, dumps, is_repo_doc, nodes, parse,
+    render_lines, render_section, to_obj, wrap_untrusted,
+)
+from .lint import lint
+from .converter import (
+    _auto_abstract, _clip as _converter_clip, _csv_row, _esc, _json_body, _json_cell,
+    _json_kv_lines, _json_scalar, _make_id, _md_cells, _one_line_text, _records_table,
+    _ref_topics, _table_name, from_json, from_md, from_payload, to_md,
+)
+from .retrieval import (
+    QSTOP, SYNONYMS, Session, _find_legacy, _index_tokens, _is_navigation_section,
+    _is_test_module, _norm, _query_terms, _rule_count, _synonym_map, _trigram_sim,
+    _trigrams, find, get, index, resolve, split_identifier,
+)
+from .repo import (
+    JS_CONTROL, LANG_BY_EXT, REPO_ENTRY_NAMES, REPO_SKIP_DIRS, REPO_SOURCE_EXTS,
+    _brace_end, _build_graph, _changed_symbols, _extract_symbols, _git_changed_paths,
+    _go_module, _go_symbols, _js_arrow_name, _js_expr_end, _js_rhs_arrow, _js_symbols,
+    _leading_comment, _py_module_name, _py_symbols, _reachable, _repo_add_deps, _repo_id,
+    _repo_info, _repo_section, _repo_summary, _resolve_go_imports, _resolve_impact_target,
+    _resolve_js_imports, _resolve_module, _resolve_py_imports, _resolve_rust_mods,
+    _rust_symbols, _sha256, _sig_line, _slice_range, _strip_code_strings, _symbol_rows,
+    _symbol_table, _top_symbol_names, from_repo, impact, render_impact, repo_source_files,
+    slice, slice_symbols, sync_repo,
+)
+from .memory import _validate_section_body, remember
+from .pipe import (
+    LOG_CMD_RE, LOG_ERR_RE, LOG_WARN_RE, _fit_text, _log_shape, _text_fields, _tokens,
+    detect_stream_format, from_stream, hook_process, log_doc, pipe_compress,
+)
+from .plugin import (
+    TIER_BEGIN, TIER_END, _hook_commands, _looks_like_tier_hook, _merge_hook_settings,
+    _plugin_instructions, _write_marked_instructions, init_plugin,
+)
+
+__version__ = VERSION
+__all__ = [
+    "VERSION", "Doc", "Section", "parse", "dumps", "index", "find", "get", "slice",
+    "slice_symbols", "impact", "sync_repo", "from_repo", "remember", "Session",
+    "wrap_untrusted", "from_md", "from_json", "from_payload", "to_md", "lint", "nodes",
+    "count_tokens", "pipe_compress", "hook_process", "init_plugin",
+]

tier_ctx-1.2.1/src/tier/__main__.py

@@ -0,0 +1,5 @@
+"""Enable `python -m tier`."""
+from .cli import main
+
+if __name__ == "__main__":
+    raise SystemExit(main())