context-loader 1.3.2__tar.gz → 1.3.4__tar.gz

{context_loader-1.3.2 → context_loader-1.3.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: context-loader
-Version: 1.3.2
+Version: 1.3.4
 Summary: Render deterministic local Git context for any executor
 License:                                  Apache License
                                     Version 2.0, January 2004
@@ -204,7 +204,7 @@ License:                                  Apache License
             See the License for the specific language governing permissions and
             limitations under the License.
          
-Requires-Python: >=3.12, <3.13
+Requires-Python: >=3.12, <3.15
 Project-URL: Repository, https://github.com/xuanheng-tech/context-loader
 Project-URL: Issues, https://github.com/xuanheng-tech/context-loader/issues
 Description-Content-Type: text/markdown
@@ -219,14 +219,14 @@ Python standard library.
 
 ## Open-source quick start
 
-Current stable release: **1.3.2**. Local CLI for coding-agent workflows; renders deterministic Markdown or JSON.
+Current stable release: **1.3.4**. Local CLI for coding-agent workflows; renders deterministic Markdown or JSON.
 
 ```bash
-uv tool install 'context-loader==1.3.2'
+uv tool install 'context-loader==1.3.4'
 ```
 
 ```bash
-pip install 'context-loader==1.3.2'
+pip install 'context-loader==1.3.4'
 ```
 
 ```bash
@@ -239,21 +239,21 @@ project-context --repo /path/to/repo --format json
 
 ## Install
 
-Source version: `1.3.2`. Install its matching published release or an exact source commit.
+Source version: `1.3.4`. Install its matching published release or an exact source commit.
 
 Install via `uv`:
 
 ```bash
-uv tool install 'context-loader==1.3.2'
+uv tool install 'context-loader==1.3.4'
 ```
 
 Install via `pip`:
 
 ```bash
-pip install 'context-loader==1.3.2'
+pip install 'context-loader==1.3.4'
 ```
 
-For development or source-based installs tracking current repository (`1.3.2`):
+For development or source-based installs tracking current repository (`1.3.4`):
 
 ```bash
 uv tool install git+https://github.com/xuanheng-tech/context-loader.git
@@ -279,7 +279,7 @@ part of either distribution; run them from a checkout of the matching tag.
 
 ## Platform and Runtime Requirements
 
-- **Python**: Python 3.12 (`>=3.12,<3.13`). Runtime code uses only the Python standard library with zero runtime dependencies.
+- **Python**: Python 3.12–3.14 (`>=3.12,<3.15`). Runtime code uses only the Python standard library with zero runtime dependencies.
 - **Git**: Requires the standard `git` CLI at the fixed absolute path `/usr/bin/git`. The
   executable is never resolved through `PATH`, so an installation elsewhere is not used and
   every invocation fails with exit `1`. Git runs with a sanitized environment that ignores
@@ -350,7 +350,7 @@ set differs, as described under “Compact model consumption”. The declared co
   "schema_version": 3,
   "tool": {
     "name": "context-loader",
-    "version": "1.3.2"
+    "version": "1.3.4"
   },
   "repository": {
     "requested_path": "/canonical/requested/path",
@@ -397,7 +397,14 @@ context, in assembly order, after the existing newline normalization and truncat
 distinguishes `repository` from `global`; version 1.0.0's fixed root-file selection currently emits
 only `repository` sources and does not add any global-file discovery.
 
-`statuses` lists machine-readable collection and render conditions that previously appeared only inside `context` Markdown — plus the nested-context truncation flags, which exist only in machine form: skipped or absent sources, truncated sources, unreadable directory-tree entries, directory-tree entries whose directory holds more entries than one enumeration examines, sections omitted under the global output budget, truncated working-tree, declared-command or directory-tree listings, and truncated nested-`AGENTS.md` presence reports. A skipped or absent source reports the code of the condition the collector observed (`not_present`, `skipped_symlink`, `skipped_not_regular`, `skipped_encoding`, `skipped_unreadable`); the sentence in `context` is rendered from that same condition, so wording and code cannot drift apart. Each entry has stable `code`, `subject_kind`, and `subject` fields. The two `nested_agents_*` codes mirror the `nested_context` truncation booleans for consumers that branch only on `statuses`; the full `nested_context` object remains a first-class field, never folded into `statuses`. Callers can ignore `statuses` safely.
+`statuses` lists skipped or absent sources, truncated sources, unreadable tree entries,
+directories whose listing exceeds its retention cap, globally omitted sections, and truncated
+working-tree, command, tree or nested-`AGENTS.md` listings. A skipped or absent source reports
+the observed condition (`not_present`, `skipped_symlink`, `skipped_not_regular`,
+`skipped_encoding`, `skipped_unreadable`); its sentence in `context` comes from that same
+condition. Each entry has stable `code`, `subject_kind`, and `subject` fields. The two
+`nested_agents_*` codes mirror the `nested_context` truncation booleans; the complete object
+remains a separate field. Callers can ignore `statuses` safely.
 
 The optional `selection` object is present only on a rendered `AGENTS.md` source. Its section entries
 contain heading, heading level, and fixed selection reasons; it never contains the original focus or
@@ -417,7 +424,7 @@ separator, with the array's brackets charged to the first entry — so the emitt
 more than the budget it declares. Because quoting is counted, an array already near the cap can
 report one path fewer than an older release did, and then says so through `list_truncated`.
 `scan_truncated` means the depth (4) or directory-count (2,000) budget was reached, a directory
-held more entries than one enumeration examines (1,024), or a directory or entry could not be
+held more entries than one listing retains (1,024), or a directory or entry could not be
 read, so absence of a path is not proof of absence of the file. Nested paths
 are sanitized with the same escaping the Markdown uses for repository-derived text, so a nested name
 the filesystem could not decode never breaks the document;
@@ -426,6 +433,16 @@ see “Limits” for the root paths that are emitted verbatim. The corresponding
 Presence is not instruction: whether a nested file applies, and its text, remain the caller's
 judgment; Markdown output is unchanged.
 
+Escaped names remain reversible and distinct: a literal backslash becomes `\\`, a backtick
+becomes `\x60`, ASCII controls and undecodable filesystem bytes become `\xNN`, and other
+non-printable Unicode characters become `\uNNNN` or `\UNNNNNNNN`. Printable Unicode stays
+unchanged. After JSON decoding, recover POSIX path bytes by reading escapes from left to right:
+`\\` yields one backslash byte, `\xNN` yields that byte, and Unicode escapes and ordinary
+characters yield their UTF-8 bytes. Thus a backtick, a literal `\x60`, an undecodable byte
+`0x80`, and Unicode `U+0080` all have different representations. This applies to nested paths
+and displayed tree/change/status names; the verbatim root/source paths retain their existing
+semantics.
+
 ### Compact model consumption (`--format json-compact`)
 
 `--format json-compact` emits a `schema_version` 4 document: exactly the version-3 document with
@@ -453,7 +470,7 @@ although 1.1.0 added the top-level `statuses` key. From release 1.2.0 onward eac
 names exactly one key set, and a new number is issued whenever a key set would otherwise be reused.
 
 The JSON schema version and package version are independent: this build emits `schema_version` `3`
-for `--format json` and `4` for `--format json-compact`, while `tool.version` is `1.3.2`. Version 1
+for `--format json` and `4` for `--format json-compact`, while `tool.version` is `1.3.4`. Version 1
 and 2 are the exact shapes already published in release 1.2.0 and are never reused: because
 `nested_context` changes the default document's key set, version 3 names the current full-document
 key set and version 4 names the compact projection.
@@ -527,10 +544,11 @@ bullet is a per-component, read or capture bound, as each one states:
   16 KiB, including a 4-KiB maximum small head
 - `README.md`: 16 KiB
 - Each entry file: 8 KiB
-- All entry-file bodies: 24 KiB
+- All entry-file bodies: 24 KiB, including truncation markers. Room is reserved for markers
+  on later text sources, so exhausting the budget never hides that their bodies were cut.
 - Declared commands: 8 KiB
 - Directory tree: 12 KiB of listing body, 300 entries, and depth 2. Each directory is
-  enumerated up to 512 entries, keeping the alphabetically first names so the listing never
+  retains and classifies up to 512 entries, keeping the alphabetically first names so the listing never
   depends on operating system order. A directory that holds more is reported by a single
   `Listing incomplete:` line stating the total, the limit and up to three names, and by at most
   eight `directory_listing_incomplete` status entries naming individual directories plus one
@@ -541,8 +559,10 @@ bullet is a per-component, read or capture bound, as each one states:
   alphabetical prefix and the entry budget kept it, so capped siblings displaced by the budget are
   disclosed as their parent's truncation rather than as separate numbers.
   Three limits act in sequence and a name survives only if all three leave room. The 512-entry
-  enumeration limit decides which names a directory offers at all; above it only an alphabetical
-  prefix is ever seen. The 300-entry item budget is then shared between a directory's
+  retention limit decides which names a directory offers at all; above it only an alphabetical
+  prefix is classified. All names must still be scanned to find that prefix, subject to the
+  separate hard name-scan bound below; the retention cap bounds memory and classification work.
+  The 300-entry item budget is then shared between a directory's
   subdirectories and its non-directories: each may take up to half of what remains before the other
   claims the surplus, which stops alphabetically-earlier directories from evicting every file, but
   it is a share rather than a floor -- with a single slot left the non-directories take it. The 12
@@ -561,8 +581,14 @@ bullet is a per-component, read or capture bound, as each one states:
   fence say so. The heading and the code-fence lines are not metered, so the whole Markdown
   section can exceed 12 KiB by that fixed few dozen bytes.
 - Nested `AGENTS.md` presence scan: depth 4, 2,000 directories, at most 32 reported paths, a
-  4 KiB budget metered on the emitted array, and at most 1,024 entries examined per directory;
-  file contents are never read
+  4 KiB budget metered on the emitted array, and at most 1,024 entries retained and classified
+  per directory. Names are scanned to select that prefix, subject to the separate hard
+  name-scan bound below; file contents are never read
+- Directory name scans stop on the first entry beyond 65,536 names per directory. An
+  over-limit directory's entire partial prefix is discarded to keep results independent of
+  operating system order. The tree emits `Skipped: directory name scan limit reached.` and
+  a `directory_scan_limit_exceeded` status; the nested presence scan sets `scan_truncated`.
+  This bounds name iteration, not elapsed time on a blocked filesystem call.
 - Working-tree changes: 100 paths and 4 KiB
 - Recent commits: 8
 - JSON-only listings have no byte budget of their own: `statuses` subjects are escaped names bounded
@@ -570,6 +596,7 @@ bullet is a per-component, read or capture bound, as each one states:
   repository root the filesystem could not decode makes the machine formats exit 1 where Markdown
   still renders
 - Git subprocess output: 16 MiB (bounded while reading)
+- Git status preflight: at most 128 repositories, including initialized nested submodules
 - Candidate file validation: 16 MiB hard read bound
 
 Truncation occurs only at complete UTF-8 and line boundaries and is marked explicitly.
@@ -593,6 +620,13 @@ the remote.
   this is normal after cloning an empty remote.
 - `Ahead / behind: not available` means the commit relationship cannot currently be computed.
   Counts are reported only when the configured upstream commit is available locally.
+- Executable Git clean/process filters can run during `status`, even with optional locks
+  disabled. Before reading status, the loader checks effective repository configuration,
+  including includes, worktree configuration and initialized nested submodules. A non-empty
+  clean/process command causes a safe error with no stdout, even if no current attribute uses
+  it. Disabling a driver silently would change cleanliness semantics. Filter-dependent repositories
+  (including locally configured Git LFS drivers) therefore need ordinary Git inspection outside
+  this loader. Git protocols are also disabled to prevent lazy object retrieval.
 
 ## Exit Codes
 
@@ -605,7 +639,7 @@ or candidate-file content.
 
 ## Not Included
 
-Version 1.3.2 does not provide AI summaries, project-type detection, loading or transport of
+Version 1.3.4 does not provide AI summaries, project-type detection, loading or transport of
 nested `AGENTS.md` contents (the JSON `nested_context` field reports bounded existence only),
 Memory retrieval, semantic ranking, ignore-rule parsing, plugins, profiles, caches, databases,
 network services, MCP, daemons, GUIs, CI/CD, telemetry, or automatic updates.
@@ -618,6 +652,44 @@ Run the complete local check:
 just check
 ```
 
+The shared check also builds a wheel and installs it into a fresh virtual environment. Package
+acceptance runs the installed CLI outside the source checkout, checking deterministic agreement
+between all three formats, linked worktrees, and safe failures. Build and installation run offline
+using the backend cache populated by `uv run --frozen`; the tests use only disposable repositories.
+Git deadline regressions verify that silent, streaming, and stdout-closed child processes are
+terminated and reaped, and that a timeout leaves CLI stdout empty.
+
+The lockfile uses public PyPI artifact URLs; no organization mirror or local Renovate preset
+is needed. Renovate keeps scheduled, rate-limited updates and manual merges using a public
+recommended preset plus the repository's explicit settings. The GitHub quality gate runs on
+Python 3.12.13, 3.13.13 and 3.14.5; the Gitea root-runner gate and release builds use 3.12.13.
+Other interpreter/platform combinations require validation before widening the supported range.
+
+Git test fixtures discard inherited `GIT_*` routing/configuration before creating repositories.
+Permission-sensitive tests must run as a non-root user. Gitea uses `just ci-check`, which installs
+Python 3.12.13 without user-level executable links and keeps its interpreter/cache under
+`$RUNNER_TEMP/context-loader-ci` (or ignored `.ci-runtime/context-loader-ci` locally).
+Non-root runners execute directly; root containers transfer ownership of that runtime and the
+checkout, then use `runuser` with an existing account named by `CONTEXT_LOADER_CI_USER` (default
+`ubuntu`). That account must be able to execute the installed `uv` and `just`; it does not need
+access to `/root` or a root-owned Python/cache. No `sudo` dependency is required for this CI step.
+
+### Performance baseline
+
+On Linux with GNU `/usr/bin/time`, measure an explicit installed CLI using an existing empty
+scratch directory:
+
+```bash
+python scripts/benchmark.py --cli /absolute/path/to/project-context --scratch /absolute/empty/scratch
+```
+
+The opt-in benchmark creates only synthetic repositories under that directory and reports median
+elapsed time, peak RSS (KiB) and JSON-compact output bytes over five runs. It checks identical output
+between runs and covers 1,000 AGENTS headings, a 5,000-file directory, and 1,000 untracked files.
+Keep the CLI version, scratch path, interpreter and machine comparable when comparing runs;
+measurements are observations, not timing thresholds in CI. Retain or remove the scratch directory
+through your own temporary-resource workflow after inspection.
+
 ## Version maintenance
 
 The versions in `pyproject.toml` and `context_loader/__init__.py`, the `tool_version` in
@@ -638,6 +710,14 @@ only verifies the public release, mirrors its exact formal tag object if missing
 the Gitea Release record. It builds no package and has no PyPI publishing credentials.
 No private endpoint or credential is needed on GitHub.
 
+For environments that permit direct public reads but share an exhausted proxy quota,
+set the Gitea Actions repository variable `RELEASE_PUBLIC_API_DIRECT=true`. The record
+step appends only public GitHub API/PyPI hosts to the existing proxy bypass lists. Without
+this explicit option, the workflow uses its inherited proxy settings.
+Both existing bypass lists and an explicit `*` wildcard are preserved. Public HTTPS GET
+transport failures get at most three attempts within the existing total retry budget;
+API writes stop on transport failure so their outcome can be verified before retrying.
+
 Release checks use Python 3.12, Git, and the runner's existing OpenSSL CLI. They compare downloaded
 PyPI file bytes, metadata and SHA-256, plus the publisher/tag/commit claims in PyPI's HTTPS-served
 provenance. This is an identity check, not an independent cryptographic Sigstore verifier.

{context_loader-1.3.2 → context_loader-1.3.4}/README.md

@@ -8,14 +8,14 @@ Python standard library.
 
 ## Open-source quick start
 
-Current stable release: **1.3.2**. Local CLI for coding-agent workflows; renders deterministic Markdown or JSON.
+Current stable release: **1.3.4**. Local CLI for coding-agent workflows; renders deterministic Markdown or JSON.
 
 ```bash
-uv tool install 'context-loader==1.3.2'
+uv tool install 'context-loader==1.3.4'
 ```
 
 ```bash
-pip install 'context-loader==1.3.2'
+pip install 'context-loader==1.3.4'
 ```
 
 ```bash
@@ -28,21 +28,21 @@ project-context --repo /path/to/repo --format json
 
 ## Install
 
-Source version: `1.3.2`. Install its matching published release or an exact source commit.
+Source version: `1.3.4`. Install its matching published release or an exact source commit.
 
 Install via `uv`:
 
 ```bash
-uv tool install 'context-loader==1.3.2'
+uv tool install 'context-loader==1.3.4'
 ```
 
 Install via `pip`:
 
 ```bash
-pip install 'context-loader==1.3.2'
+pip install 'context-loader==1.3.4'
 ```
 
-For development or source-based installs tracking current repository (`1.3.2`):
+For development or source-based installs tracking current repository (`1.3.4`):
 
 ```bash
 uv tool install git+https://github.com/xuanheng-tech/context-loader.git
@@ -68,7 +68,7 @@ part of either distribution; run them from a checkout of the matching tag.
 
 ## Platform and Runtime Requirements
 
-- **Python**: Python 3.12 (`>=3.12,<3.13`). Runtime code uses only the Python standard library with zero runtime dependencies.
+- **Python**: Python 3.12–3.14 (`>=3.12,<3.15`). Runtime code uses only the Python standard library with zero runtime dependencies.
 - **Git**: Requires the standard `git` CLI at the fixed absolute path `/usr/bin/git`. The
   executable is never resolved through `PATH`, so an installation elsewhere is not used and
   every invocation fails with exit `1`. Git runs with a sanitized environment that ignores
@@ -139,7 +139,7 @@ set differs, as described under “Compact model consumption”. The declared co
   "schema_version": 3,
   "tool": {
     "name": "context-loader",
-    "version": "1.3.2"
+    "version": "1.3.4"
   },
   "repository": {
     "requested_path": "/canonical/requested/path",
@@ -186,7 +186,14 @@ context, in assembly order, after the existing newline normalization and truncat
 distinguishes `repository` from `global`; version 1.0.0's fixed root-file selection currently emits
 only `repository` sources and does not add any global-file discovery.
 
-`statuses` lists machine-readable collection and render conditions that previously appeared only inside `context` Markdown — plus the nested-context truncation flags, which exist only in machine form: skipped or absent sources, truncated sources, unreadable directory-tree entries, directory-tree entries whose directory holds more entries than one enumeration examines, sections omitted under the global output budget, truncated working-tree, declared-command or directory-tree listings, and truncated nested-`AGENTS.md` presence reports. A skipped or absent source reports the code of the condition the collector observed (`not_present`, `skipped_symlink`, `skipped_not_regular`, `skipped_encoding`, `skipped_unreadable`); the sentence in `context` is rendered from that same condition, so wording and code cannot drift apart. Each entry has stable `code`, `subject_kind`, and `subject` fields. The two `nested_agents_*` codes mirror the `nested_context` truncation booleans for consumers that branch only on `statuses`; the full `nested_context` object remains a first-class field, never folded into `statuses`. Callers can ignore `statuses` safely.
+`statuses` lists skipped or absent sources, truncated sources, unreadable tree entries,
+directories whose listing exceeds its retention cap, globally omitted sections, and truncated
+working-tree, command, tree or nested-`AGENTS.md` listings. A skipped or absent source reports
+the observed condition (`not_present`, `skipped_symlink`, `skipped_not_regular`,
+`skipped_encoding`, `skipped_unreadable`); its sentence in `context` comes from that same
+condition. Each entry has stable `code`, `subject_kind`, and `subject` fields. The two
+`nested_agents_*` codes mirror the `nested_context` truncation booleans; the complete object
+remains a separate field. Callers can ignore `statuses` safely.
 
 The optional `selection` object is present only on a rendered `AGENTS.md` source. Its section entries
 contain heading, heading level, and fixed selection reasons; it never contains the original focus or
@@ -206,7 +213,7 @@ separator, with the array's brackets charged to the first entry — so the emitt
 more than the budget it declares. Because quoting is counted, an array already near the cap can
 report one path fewer than an older release did, and then says so through `list_truncated`.
 `scan_truncated` means the depth (4) or directory-count (2,000) budget was reached, a directory
-held more entries than one enumeration examines (1,024), or a directory or entry could not be
+held more entries than one listing retains (1,024), or a directory or entry could not be
 read, so absence of a path is not proof of absence of the file. Nested paths
 are sanitized with the same escaping the Markdown uses for repository-derived text, so a nested name
 the filesystem could not decode never breaks the document;
@@ -215,6 +222,16 @@ see “Limits” for the root paths that are emitted verbatim. The corresponding
 Presence is not instruction: whether a nested file applies, and its text, remain the caller's
 judgment; Markdown output is unchanged.
 
+Escaped names remain reversible and distinct: a literal backslash becomes `\\`, a backtick
+becomes `\x60`, ASCII controls and undecodable filesystem bytes become `\xNN`, and other
+non-printable Unicode characters become `\uNNNN` or `\UNNNNNNNN`. Printable Unicode stays
+unchanged. After JSON decoding, recover POSIX path bytes by reading escapes from left to right:
+`\\` yields one backslash byte, `\xNN` yields that byte, and Unicode escapes and ordinary
+characters yield their UTF-8 bytes. Thus a backtick, a literal `\x60`, an undecodable byte
+`0x80`, and Unicode `U+0080` all have different representations. This applies to nested paths
+and displayed tree/change/status names; the verbatim root/source paths retain their existing
+semantics.
+
 ### Compact model consumption (`--format json-compact`)
 
 `--format json-compact` emits a `schema_version` 4 document: exactly the version-3 document with
@@ -242,7 +259,7 @@ although 1.1.0 added the top-level `statuses` key. From release 1.2.0 onward eac
 names exactly one key set, and a new number is issued whenever a key set would otherwise be reused.
 
 The JSON schema version and package version are independent: this build emits `schema_version` `3`
-for `--format json` and `4` for `--format json-compact`, while `tool.version` is `1.3.2`. Version 1
+for `--format json` and `4` for `--format json-compact`, while `tool.version` is `1.3.4`. Version 1
 and 2 are the exact shapes already published in release 1.2.0 and are never reused: because
 `nested_context` changes the default document's key set, version 3 names the current full-document
 key set and version 4 names the compact projection.
@@ -316,10 +333,11 @@ bullet is a per-component, read or capture bound, as each one states:
   16 KiB, including a 4-KiB maximum small head
 - `README.md`: 16 KiB
 - Each entry file: 8 KiB
-- All entry-file bodies: 24 KiB
+- All entry-file bodies: 24 KiB, including truncation markers. Room is reserved for markers
+  on later text sources, so exhausting the budget never hides that their bodies were cut.
 - Declared commands: 8 KiB
 - Directory tree: 12 KiB of listing body, 300 entries, and depth 2. Each directory is
-  enumerated up to 512 entries, keeping the alphabetically first names so the listing never
+  retains and classifies up to 512 entries, keeping the alphabetically first names so the listing never
   depends on operating system order. A directory that holds more is reported by a single
   `Listing incomplete:` line stating the total, the limit and up to three names, and by at most
   eight `directory_listing_incomplete` status entries naming individual directories plus one
@@ -330,8 +348,10 @@ bullet is a per-component, read or capture bound, as each one states:
   alphabetical prefix and the entry budget kept it, so capped siblings displaced by the budget are
   disclosed as their parent's truncation rather than as separate numbers.
   Three limits act in sequence and a name survives only if all three leave room. The 512-entry
-  enumeration limit decides which names a directory offers at all; above it only an alphabetical
-  prefix is ever seen. The 300-entry item budget is then shared between a directory's
+  retention limit decides which names a directory offers at all; above it only an alphabetical
+  prefix is classified. All names must still be scanned to find that prefix, subject to the
+  separate hard name-scan bound below; the retention cap bounds memory and classification work.
+  The 300-entry item budget is then shared between a directory's
   subdirectories and its non-directories: each may take up to half of what remains before the other
   claims the surplus, which stops alphabetically-earlier directories from evicting every file, but
   it is a share rather than a floor -- with a single slot left the non-directories take it. The 12
@@ -350,8 +370,14 @@ bullet is a per-component, read or capture bound, as each one states:
   fence say so. The heading and the code-fence lines are not metered, so the whole Markdown
   section can exceed 12 KiB by that fixed few dozen bytes.
 - Nested `AGENTS.md` presence scan: depth 4, 2,000 directories, at most 32 reported paths, a
-  4 KiB budget metered on the emitted array, and at most 1,024 entries examined per directory;
-  file contents are never read
+  4 KiB budget metered on the emitted array, and at most 1,024 entries retained and classified
+  per directory. Names are scanned to select that prefix, subject to the separate hard
+  name-scan bound below; file contents are never read
+- Directory name scans stop on the first entry beyond 65,536 names per directory. An
+  over-limit directory's entire partial prefix is discarded to keep results independent of
+  operating system order. The tree emits `Skipped: directory name scan limit reached.` and
+  a `directory_scan_limit_exceeded` status; the nested presence scan sets `scan_truncated`.
+  This bounds name iteration, not elapsed time on a blocked filesystem call.
 - Working-tree changes: 100 paths and 4 KiB
 - Recent commits: 8
 - JSON-only listings have no byte budget of their own: `statuses` subjects are escaped names bounded
@@ -359,6 +385,7 @@ bullet is a per-component, read or capture bound, as each one states:
   repository root the filesystem could not decode makes the machine formats exit 1 where Markdown
   still renders
 - Git subprocess output: 16 MiB (bounded while reading)
+- Git status preflight: at most 128 repositories, including initialized nested submodules
 - Candidate file validation: 16 MiB hard read bound
 
 Truncation occurs only at complete UTF-8 and line boundaries and is marked explicitly.
@@ -382,6 +409,13 @@ the remote.
   this is normal after cloning an empty remote.
 - `Ahead / behind: not available` means the commit relationship cannot currently be computed.
   Counts are reported only when the configured upstream commit is available locally.
+- Executable Git clean/process filters can run during `status`, even with optional locks
+  disabled. Before reading status, the loader checks effective repository configuration,
+  including includes, worktree configuration and initialized nested submodules. A non-empty
+  clean/process command causes a safe error with no stdout, even if no current attribute uses
+  it. Disabling a driver silently would change cleanliness semantics. Filter-dependent repositories
+  (including locally configured Git LFS drivers) therefore need ordinary Git inspection outside
+  this loader. Git protocols are also disabled to prevent lazy object retrieval.
 
 ## Exit Codes
 
@@ -394,7 +428,7 @@ or candidate-file content.
 
 ## Not Included
 
-Version 1.3.2 does not provide AI summaries, project-type detection, loading or transport of
+Version 1.3.4 does not provide AI summaries, project-type detection, loading or transport of
 nested `AGENTS.md` contents (the JSON `nested_context` field reports bounded existence only),
 Memory retrieval, semantic ranking, ignore-rule parsing, plugins, profiles, caches, databases,
 network services, MCP, daemons, GUIs, CI/CD, telemetry, or automatic updates.
@@ -407,6 +441,44 @@ Run the complete local check:
 just check
 ```
 
+The shared check also builds a wheel and installs it into a fresh virtual environment. Package
+acceptance runs the installed CLI outside the source checkout, checking deterministic agreement
+between all three formats, linked worktrees, and safe failures. Build and installation run offline
+using the backend cache populated by `uv run --frozen`; the tests use only disposable repositories.
+Git deadline regressions verify that silent, streaming, and stdout-closed child processes are
+terminated and reaped, and that a timeout leaves CLI stdout empty.
+
+The lockfile uses public PyPI artifact URLs; no organization mirror or local Renovate preset
+is needed. Renovate keeps scheduled, rate-limited updates and manual merges using a public
+recommended preset plus the repository's explicit settings. The GitHub quality gate runs on
+Python 3.12.13, 3.13.13 and 3.14.5; the Gitea root-runner gate and release builds use 3.12.13.
+Other interpreter/platform combinations require validation before widening the supported range.
+
+Git test fixtures discard inherited `GIT_*` routing/configuration before creating repositories.
+Permission-sensitive tests must run as a non-root user. Gitea uses `just ci-check`, which installs
+Python 3.12.13 without user-level executable links and keeps its interpreter/cache under
+`$RUNNER_TEMP/context-loader-ci` (or ignored `.ci-runtime/context-loader-ci` locally).
+Non-root runners execute directly; root containers transfer ownership of that runtime and the
+checkout, then use `runuser` with an existing account named by `CONTEXT_LOADER_CI_USER` (default
+`ubuntu`). That account must be able to execute the installed `uv` and `just`; it does not need
+access to `/root` or a root-owned Python/cache. No `sudo` dependency is required for this CI step.
+
+### Performance baseline
+
+On Linux with GNU `/usr/bin/time`, measure an explicit installed CLI using an existing empty
+scratch directory:
+
+```bash
+python scripts/benchmark.py --cli /absolute/path/to/project-context --scratch /absolute/empty/scratch
+```
+
+The opt-in benchmark creates only synthetic repositories under that directory and reports median
+elapsed time, peak RSS (KiB) and JSON-compact output bytes over five runs. It checks identical output
+between runs and covers 1,000 AGENTS headings, a 5,000-file directory, and 1,000 untracked files.
+Keep the CLI version, scratch path, interpreter and machine comparable when comparing runs;
+measurements are observations, not timing thresholds in CI. Retain or remove the scratch directory
+through your own temporary-resource workflow after inspection.
+
 ## Version maintenance
 
 The versions in `pyproject.toml` and `context_loader/__init__.py`, the `tool_version` in
@@ -427,6 +499,14 @@ only verifies the public release, mirrors its exact formal tag object if missing
 the Gitea Release record. It builds no package and has no PyPI publishing credentials.
 No private endpoint or credential is needed on GitHub.
 
+For environments that permit direct public reads but share an exhausted proxy quota,
+set the Gitea Actions repository variable `RELEASE_PUBLIC_API_DIRECT=true`. The record
+step appends only public GitHub API/PyPI hosts to the existing proxy bypass lists. Without
+this explicit option, the workflow uses its inherited proxy settings.
+Both existing bypass lists and an explicit `*` wildcard are preserved. Public HTTPS GET
+transport failures get at most three attempts within the existing total retry budget;
+API writes stop on transport failure so their outcome can be verified before retrying.
+
 Release checks use Python 3.12, Git, and the runner's existing OpenSSL CLI. They compare downloaded
 PyPI file bytes, metadata and SHA-256, plus the publisher/tag/commit claims in PyPI's HTTPS-served
 provenance. This is an identity check, not an independent cryptographic Sigstore verifier.

{context_loader-1.3.2 → context_loader-1.3.4}/context_loader/__init__.py

@@ -1,3 +1,3 @@
 """Project Context Loader."""
 
-__version__ = "1.3.2"
+__version__ = "1.3.4"

{context_loader-1.3.2 → context_loader-1.3.4}/context_loader/application.py

@@ -26,7 +26,6 @@ from .model import (
 from .render import (
     RenderedSourceBody,
     render_markdown_with_details,
-    rendered_source_contents,
 )
 
 # Each value names one exact document shape and is never reused: 1 and 2 are the
@@ -70,9 +69,7 @@ class ContextStatus:
     code: str
     subject_kind: str
     subject: str
-    # The raw fact behind the displayed subject. Display escaping is not injective, so
-    # deduplication must key on this and never on `subject`, or two different directories
-    # could collapse into one claim. It is not serialized.
+    # Deduplicate raw facts independently of display formatting. Not serialized.
     identity: str = field(default="", compare=False)
 
 
@@ -191,7 +188,19 @@ def _build_statuses(
     for entry in project.directory_tree.entries:
         if entry.kind == "unreadable_directory":
             subject = entry.path if entry.path else "."
-            statuses.append(_status("unreadable", "tree_entry", display_text(subject)))
+            statuses.append(
+                _status("unreadable", "tree_entry", display_text(subject), identity=subject)
+            )
+        elif entry.kind == "scan_limit_directory":
+            subject = entry.path if entry.path else "."
+            statuses.append(
+                _status(
+                    "directory_scan_limit_exceeded",
+                    "tree_entry",
+                    display_text(subject),
+                    identity=subject,
+                )
+            )
     tree = project.directory_tree
     named = list(dict.fromkeys(tree.incomplete_directories))
     for path in named:
@@ -212,7 +221,7 @@ def _build_statuses(
                 "directory_listing_incomplete",
                 "tree",
                 f"{tree.incomplete_count} directories exceeded the "
-                f"{tree.enumeration_limit}-entry per-directory enumeration limit; "
+                f"{tree.enumeration_limit}-entry per-directory retention limit; "
                 f"{unnamed} further {stem} not named individually",
                 identity=f"aggregate:{tree.incomplete_count}:{unnamed}",
             )
@@ -254,7 +263,7 @@ def load_project_context(
         raise ContextLoaderError(str(exc), exit_code=2) from None
     rendered = render_markdown_with_details(state, project)
     context = rendered.output.decode("utf-8")
-    rendered_bodies = rendered_source_contents(project, rendered.included_sections)
+    rendered_bodies = rendered.source_bodies
     sources = _sources(state.repository, rendered_bodies)
     statuses = _build_statuses(
         project,