agsync 0.1.0__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (34) hide show
  1. agsync-0.1.0/.github/ISSUE_TEMPLATE/rule-proposal.md +17 -0
  2. agsync-0.1.0/.github/workflows/ci.yml +38 -0
  3. agsync-0.1.0/.github/workflows/publish.yml +78 -0
  4. agsync-0.1.0/.gitignore +17 -0
  5. agsync-0.1.0/CODE_OF_CONDUCT.md +6 -0
  6. agsync-0.1.0/CONTRIBUTING.md +67 -0
  7. agsync-0.1.0/LICENSE +21 -0
  8. agsync-0.1.0/PKG-INFO +291 -0
  9. agsync-0.1.0/README.md +269 -0
  10. agsync-0.1.0/action.yml +42 -0
  11. agsync-0.1.0/docs/design.md +262 -0
  12. agsync-0.1.0/pyproject.toml +42 -0
  13. agsync-0.1.0/src/agsync/__init__.py +24 -0
  14. agsync-0.1.0/src/agsync/__main__.py +5 -0
  15. agsync-0.1.0/src/agsync/cli.py +211 -0
  16. agsync-0.1.0/src/agsync/engine.py +147 -0
  17. agsync-0.1.0/src/agsync/model.py +138 -0
  18. agsync-0.1.0/src/agsync/parser.py +310 -0
  19. agsync-0.1.0/src/agsync/replay.py +524 -0
  20. agsync-0.1.0/src/agsync/reporters.py +107 -0
  21. agsync-0.1.0/src/agsync/rules/__init__.py +57 -0
  22. agsync-0.1.0/src/agsync/rules/decisions.py +112 -0
  23. agsync-0.1.0/src/agsync/rules/links.py +113 -0
  24. agsync-0.1.0/src/agsync/rules/tasks.py +168 -0
  25. agsync-0.1.0/src/agsync/scaffold.py +180 -0
  26. agsync-0.1.0/tests/fixtures/decayed/AGENTS.md +9 -0
  27. agsync-0.1.0/tests/fixtures/decayed/memory/decisions.md +28 -0
  28. agsync-0.1.0/tests/fixtures/decayed/tasks/09-verify-staging-deploy.md +9 -0
  29. agsync-0.1.0/tests/fixtures/decayed/tasks/25-cache-invalidation.md +5 -0
  30. agsync-0.1.0/tests/fixtures/decayed/tasks/26-rate-limit-headers.md +6 -0
  31. agsync-0.1.0/tests/fixtures/decayed/tasks/README.md +14 -0
  32. agsync-0.1.0/tests/replay_repo.py +133 -0
  33. agsync-0.1.0/tests/test_agsync.py +420 -0
  34. agsync-0.1.0/tests/test_replay.py +254 -0
@@ -0,0 +1,17 @@
1
+ ---
2
+ name: Rule proposal
3
+ about: Propose a new integrity check
4
+ labels: rule
5
+ ---
6
+
7
+ **What silently breaks today?**
8
+ Describe the failure as it appeared in a real repo, not in the abstract.
9
+
10
+ **Why can't a human catch it by reading?**
11
+ The best rules catch things that look fine to a careful reader.
12
+
13
+ **Minimal failing example**
14
+ The smallest memory files that reproduce it.
15
+
16
+ **Proposed severity**
17
+ `error` (the memory is now wrong) or `warn` (the memory is fragile).
@@ -0,0 +1,38 @@
1
+ name: CI
2
+
3
+ on:
4
+ push:
5
+ branches: [main]
6
+ pull_request:
7
+
8
+ jobs:
9
+ test:
10
+ runs-on: ${{ matrix.os }}
11
+ strategy:
12
+ fail-fast: false
13
+ matrix:
14
+ os: [ubuntu-latest, macos-latest, windows-latest]
15
+ python-version: ["3.11", "3.12", "3.13"]
16
+ steps:
17
+ - uses: actions/checkout@v4
18
+ - uses: actions/setup-python@v5
19
+ with:
20
+ python-version: ${{ matrix.python-version }}
21
+ - run: pip install -e ".[dev]"
22
+ - run: ruff check src tests
23
+ - run: pytest -q
24
+
25
+ scaffold:
26
+ runs-on: ubuntu-latest
27
+ steps:
28
+ - uses: actions/checkout@v4
29
+ - uses: actions/setup-python@v5
30
+ with:
31
+ python-version: "3.12"
32
+ - run: pip install -e .
33
+ # A rule that fires on a structure agsync generated itself is a false
34
+ # positive by definition, so `init` output must always lint clean. This
35
+ # repository does not carry its own ledger to check instead: that is
36
+ # generated output, not source.
37
+ - run: agsync init "$RUNNER_TEMP/scaffold"
38
+ - run: agsync check "$RUNNER_TEMP/scaffold" --format github
@@ -0,0 +1,78 @@
1
+ name: Publish
2
+
3
+ on:
4
+ release:
5
+ types: [published]
6
+
7
+ jobs:
8
+ version:
9
+ runs-on: ubuntu-latest
10
+ steps:
11
+ - uses: actions/checkout@v4
12
+ - uses: actions/setup-python@v5
13
+ with:
14
+ python-version: "3.12"
15
+ # The tag names the release; pyproject names what actually gets uploaded.
16
+ # Nothing else ties them together, so a tag cut against the wrong commit
17
+ # publishes a version the tag does not describe — and PyPI never lets that
18
+ # version be replaced. Fails in seconds, in parallel with the suite.
19
+ - name: Tag must match the packaged version
20
+ env:
21
+ TAG: ${{ github.event.release.tag_name }}
22
+ run: |
23
+ packaged=$(python -c 'import tomllib; print(tomllib.load(open("pyproject.toml","rb"))["project"]["version"])')
24
+ echo "tag=$TAG packaged=$packaged"
25
+ if [ "$TAG" != "$packaged" ] && [ "$TAG" != "v$packaged" ]; then
26
+ echo "::error::release tag $TAG does not match pyproject version $packaged"
27
+ exit 1
28
+ fi
29
+
30
+ test:
31
+ runs-on: ubuntu-latest
32
+ strategy:
33
+ fail-fast: false
34
+ matrix:
35
+ # The ends of the supported range from pyproject's requires-python.
36
+ # A release that breaks on either end should never reach PyPI.
37
+ python-version: ["3.11", "3.13"]
38
+ steps:
39
+ - uses: actions/checkout@v4
40
+ - uses: actions/setup-python@v5
41
+ with:
42
+ python-version: ${{ matrix.python-version }}
43
+ - run: pip install -e ".[dev]"
44
+ - run: ruff check src tests
45
+ - run: pytest -q
46
+
47
+ build:
48
+ needs: [test, version]
49
+ runs-on: ubuntu-latest
50
+ steps:
51
+ - uses: actions/checkout@v4
52
+ - uses: actions/setup-python@v5
53
+ with:
54
+ python-version: "3.12"
55
+ - run: python -m pip install --upgrade build
56
+ - run: python -m build
57
+ - uses: actions/upload-artifact@v4
58
+ with:
59
+ name: dist
60
+ path: dist/
61
+
62
+ publish:
63
+ needs: build
64
+ runs-on: ubuntu-latest
65
+ # Trusted Publishing: PyPI mints a short-lived token from this job's OIDC
66
+ # identity, which is why no API token appears anywhere in this file. The
67
+ # identity PyPI checks is (repository, workflow filename, environment), so
68
+ # renaming this file or the environment breaks publishing until the
69
+ # trusted publisher on PyPI is updated to match.
70
+ environment: pypi
71
+ permissions:
72
+ id-token: write
73
+ steps:
74
+ - uses: actions/download-artifact@v4
75
+ with:
76
+ name: dist
77
+ path: dist/
78
+ - uses: pypa/gh-action-pypi-publish@release/v1
@@ -0,0 +1,17 @@
1
+ __pycache__/
2
+ *.py[cod]
3
+ .pytest_cache/
4
+ .ruff_cache/
5
+ dist/
6
+ build/
7
+ *.egg-info/
8
+ .venv/
9
+ venv/
10
+
11
+ # This repository's own agent memory. It is `agsync init` output kept locally:
12
+ # generated files do not belong in the repository that generates them. Paths are
13
+ # root-anchored so the deliberately-broken fixtures under tests/ stay tracked.
14
+ /AGENTS.md
15
+ /.agsync.toml
16
+ /memory/
17
+ /tasks/
@@ -0,0 +1,6 @@
1
+ # Code of Conduct
2
+
3
+ This project follows the [Contributor Covenant v2.1](https://www.contributor-covenant.org/version/2/1/code_of_conduct/).
4
+
5
+ Report unacceptable behavior by opening a confidential issue or contacting the
6
+ maintainer directly. Reports are handled privately.
@@ -0,0 +1,67 @@
1
+ # Contributing
2
+
3
+ ```bash
4
+ git clone https://github.com/gabrielft-me/agsync
5
+ cd agsync
6
+ pip install -e ".[dev]"
7
+ pytest -q
8
+ ```
9
+
10
+ No dependencies in the runtime package. Ever. This runs inside git hooks on
11
+ other people's machines, and a virtualenv surprise there is a broken commit for
12
+ someone who never asked for our opinions about packaging.
13
+
14
+ ## Before you start
15
+
16
+ [docs/design.md](docs/design.md) explains why the architecture is the way it is
17
+ — why the parser normalizes instead of demanding, why severity is configuration,
18
+ why fingerprints hash structured fields, why replay clones instead of checking
19
+ out, and why intent is never inferred from a diff. Worth reading if a change
20
+ starts to feel like it is fighting the codebase.
21
+
22
+ ## Adding a rule
23
+
24
+ A rule is one function. Start here — it's the easiest contribution and the most
25
+ valuable one.
26
+
27
+ 1. Write it in `src/agsync/rules/` (decisions, tasks, or links):
28
+
29
+ ```python
30
+ @rule("no-undated-decisions", ERROR, "One line: what silently breaks.")
31
+ def no_undated_decisions(memory):
32
+ for decision in memory.decisions.values():
33
+ if "date" not in decision.fields:
34
+ yield Finding("no-undated-decisions", decision.path,
35
+ decision.line, f"{decision.id} has no date")
36
+ ```
37
+
38
+ 2. Add the failing case to `tests/fixtures/decayed/` and a test asserting the
39
+ rule fires with a useful message.
40
+ 3. Add a row to the rules table in `README.md`.
41
+
42
+ ### What makes a good rule
43
+
44
+ - **It catches something invisible to a careful reader.** A duplicate `D-021`
45
+ fifty lines apart is the archetype.
46
+ - **It has zero false positives on a healthy repo.** Run it against
47
+ `agsync init` output; that must stay clean.
48
+ - **The message says what broke and why it matters**, not just which check
49
+ failed. Compare: `"duplicate ID"` versus `"D-021 is redefined here (first
50
+ defined at line 7); 9 reference(s) are now ambiguous"`.
51
+ - **Default to `warn` if it encodes a style opinion**, `error` only if the
52
+ memory is now factually wrong.
53
+
54
+ ## Parser changes
55
+
56
+ The parser normalizes; it never demands. If a real repo uses a format we don't
57
+ read, that's a parser bug, not a user error. Add the shape to
58
+ `tests/fixtures/` and make the parser absorb it.
59
+
60
+ Never make a rule reach into raw markdown. If a rule needs something the model
61
+ doesn't carry, add the field to the model. The reasoning is in
62
+ [docs/design.md](docs/design.md).
63
+
64
+ ## Style
65
+
66
+ `ruff check src tests` must pass. Comments explain *why*, not *what* — the code
67
+ already says what.
agsync-0.1.0/LICENSE ADDED
@@ -0,0 +1,21 @@
1
+ MIT License
2
+
3
+ Copyright (c) 2026 Gabriel Fagundes
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
agsync-0.1.0/PKG-INFO ADDED
@@ -0,0 +1,291 @@
1
+ Metadata-Version: 2.5
2
+ Name: agsync
3
+ Version: 0.1.0
4
+ Summary: ESLint for your agents' memory — catch decision-log rot before it reaches a new session.
5
+ Project-URL: Homepage, https://github.com/gabrielft-me/agsync
6
+ Project-URL: Issues, https://github.com/gabrielft-me/agsync/issues
7
+ Author: Gabriel Fagundes
8
+ License-Expression: MIT
9
+ License-File: LICENSE
10
+ Keywords: agents-md,ai-agents,decision-log,git-hooks,linter,llm
11
+ Classifier: Development Status :: 3 - Alpha
12
+ Classifier: Intended Audience :: Developers
13
+ Classifier: License :: OSI Approved :: MIT License
14
+ Classifier: Programming Language :: Python :: 3.11
15
+ Classifier: Programming Language :: Python :: 3.12
16
+ Classifier: Topic :: Software Development :: Quality Assurance
17
+ Requires-Python: >=3.11
18
+ Provides-Extra: dev
19
+ Requires-Dist: pytest>=8; extra == 'dev'
20
+ Requires-Dist: ruff>=0.6; extra == 'dev'
21
+ Description-Content-Type: text/markdown
22
+
23
+ # agsync
24
+
25
+ **ESLint for your agents' memory — because the decision log your AI agents write is already lying to them.**
26
+
27
+ [![CI](https://github.com/gabrielft-me/agsync/actions/workflows/ci.yml/badge.svg)](https://github.com/gabrielft-me/agsync/actions/workflows/ci.yml)
28
+ [![PyPI](https://img.shields.io/pypi/v/agsync.svg)](https://pypi.org/project/agsync/)
29
+ [![License: MIT](https://img.shields.io/badge/license-MIT-blue.svg)](LICENSE)
30
+
31
+ ```console
32
+ $ agsync check
33
+ AGENTS.md
34
+ 6 error protocol-files-exist boot protocol requires 'memory/goal.md', which does not
35
+ exist — this step is unsatisfiable and every session skips it
36
+ memory/decisions.md
37
+ 11 error unique-decision-ids D-021 is redefined here (first defined at line 7);
38
+ 7 reference(s) to D-021 are now ambiguous
39
+ 88 error relations-resolve D-027 supersedes D-019, which is not defined
40
+ tasks/25-search-indexing.md
41
+ 3 warn status-not-qualified status is qualified with free text ('(all tiers except the
42
+ edge cache)'); keep the status machine-readable
43
+ tasks/README.md
44
+ - error index-matches-files task 26 exists (tasks/26-webhook-retries.md) but is absent
45
+ from the index — invisible to anyone reading it
46
+ 14 error index-matches-files task 25: index says 'todo', file says 'done'
47
+
48
+ 27 decisions, 26 tasks, 25 index rows
49
+ 9 error(s), 4 warning(s)
50
+ ```
51
+
52
+ ---
53
+
54
+ ## The problem
55
+
56
+ You gave your agents a memory. `AGENTS.md` with a boot protocol, a numbered
57
+ decision log, a task queue. Every session starts by reading it. It is the only
58
+ thing standing between a new agent and a cold start.
59
+
60
+ **Nothing checks that it is true.**
61
+
62
+ Markdown never fails. A decision ID reused for a second, unrelated decision. A
63
+ boot step pointing at a file nobody ever wrote. A status of `done (stub)`. An
64
+ index row that disagrees with the task file it names. Every one of these parses,
65
+ renders, and reviews perfectly well, and every one is read as ground truth by the
66
+ next agent that boots.
67
+
68
+ Four things make the decay invisible, and they compound:
69
+
70
+ - **The reference graph is prose.** `D-021` is a bare token, not a link. Nothing
71
+ resolves it, so nothing notices when a second entry claims the same ID and
72
+ every existing reference to it silently becomes ambiguous.
73
+ - **Contradictions live in separate files.** The index says `todo`, the task file
74
+ says `done`. The protocol says "read this", and the file was never created. No
75
+ single diff contains both halves, so no review can catch it.
76
+ - **Every writer has a partial view.** Each session is a different agent, and
77
+ each edit is locally reasonable. Rot is what locally reasonable edits add up to
78
+ when nothing holds the whole.
79
+ - **A false memory looks exactly like a true one.** No crash, no red test. The
80
+ agent reads a confident, well-formatted, wrong statement and acts on it.
81
+
82
+ The failure is not hypothetical and it is not slow. A repo written by several
83
+ agents over a couple of days will already contain several of these.
84
+
85
+ > A memory system whose integrity depends on being maintained by hand degrades
86
+ > at exactly the speed you use it.
87
+
88
+ The design is not the problem — `AGENTS.md` plus a decision log is the right
89
+ shape. The problem is that nothing enforces it.
90
+
91
+ ## The fix is not a smarter agent
92
+
93
+ Agents self-report badly. They tell you what they *intended* to do, not what
94
+ they did, and every harness writes in a different dialect. "Be more careful" does
95
+ not survive a new agent joining.
96
+
97
+ `git push` is a choke point nobody can skip. Put the validation there.
98
+
99
+ ## Install
100
+
101
+ ```bash
102
+ uvx agsync check # try it, no install
103
+ pipx install agsync # keep it
104
+ pip install agsync
105
+ ```
106
+
107
+ Python 3.11+. **Zero dependencies** — the whole thing is stdlib, so it runs in a
108
+ git hook without a virtualenv surprise.
109
+
110
+ ## Use
111
+
112
+ ```bash
113
+ agsync check # lint the repo
114
+ agsync check --warn-only # first run: see everything, fail nothing
115
+ agsync check --baseline # accept today's mess, fail on new mess
116
+ agsync check --no-baseline # what an agent runs at boot: everything, unsuppressed
117
+ agsync check --format github # inline annotations on the PR diff
118
+ agsync check --format json # for your own tooling
119
+ agsync replay <repo> # how many past pushes the gate would have stopped
120
+ agsync rules # what it checks and why
121
+ agsync init # scaffold a memory structure from scratch
122
+ agsync install-hooks # gate commits locally
123
+ ```
124
+
125
+ ### Adopting an existing repo
126
+
127
+ Do not turn everything on at once — a linter that rejects most of your pushes on
128
+ day one gets uninstalled on day one.
129
+
130
+ ```bash
131
+ agsync check --warn-only # 1. look at the damage
132
+ agsync check --baseline # 2. freeze it; new violations now fail
133
+ ```
134
+
135
+ The baseline records a fingerprint per violation that survives line shifts, so
136
+ unrelated edits above a known problem don't resurrect it as "new". Delete
137
+ entries from `.agsync-baseline.json` as you fix them.
138
+
139
+ **The baseline governs the gate, and never the reader.** It records what should
140
+ stop a push — a decision about your rollout, not a statement about what is true.
141
+ An agent about to act on this memory needs the whole picture, including the parts
142
+ you chose not to block, so the boot protocol runs:
143
+
144
+ ```bash
145
+ agsync check --no-baseline
146
+ ```
147
+
148
+ Suppressing a finding is how you keep working. Believing it was suppressed
149
+ because it was fixed is how a new session acts on something false.
150
+
151
+ ### Replay: how bad is it already?
152
+
153
+ `agsync replay` runs the same check engine at every commit in a repository's
154
+ history and counts the pushes the gate would have stopped. Point it at a repo
155
+ that has been running on hand-maintained memory for a while, and it will tell you
156
+ when each violation started and how long it has been sitting there.
157
+
158
+ ```console
159
+ $ agsync replay . --ref main --first-seen
160
+ sha date errors rules
161
+ 3ad9e51 2026-03-02 0 —
162
+ ...
163
+ b7e2d13 2026-03-14 4 index-matches-files, links-resolve, unique-decision-ids
164
+
165
+ 18 of 42 pushes would have been rejected
166
+
167
+ rule first failed survived
168
+ protocol-files-exist 9f1c0a4 2026-03-04 31 commits, 9 days (still failing at HEAD)
169
+ unique-decision-ids c40b8f2 2026-03-09 14 commits, 4 days (still failing at HEAD)
170
+ index-matches-files c40b8f2 2026-03-09 14 commits, 4 days (still failing at HEAD)
171
+ ```
172
+
173
+ The second table is the uncomfortable one. A broken boot step usually dates from
174
+ the commit that introduced the protocol and is usually still broken at HEAD,
175
+ because nothing has ever checked it.
176
+
177
+ Replay clones the target to a temp directory and moves HEAD only inside that
178
+ clone, so it cannot leave your repository detached at an old commit. It takes a
179
+ path or a clone URL, exits 0 unless the run itself fails — it reports on history,
180
+ it does not gate it — and ignores any baseline, because the question is what the
181
+ rules would have caught, not what someone had already told the gate to overlook.
182
+
183
+ One caveat worth stating out loud: replay judges old commits by today's rules.
184
+ The claim is "these would be rejected now", not "these were rejected then".
185
+
186
+ ## Rules
187
+
188
+ | Rule | Default | Catches |
189
+ |---|---|---|
190
+ | `unique-decision-ids` | error | Two ledger entries sharing an ID, making every reference ambiguous |
191
+ | `decision-refs-resolve` | error | A `D-xxx` cited anywhere that the ledger never defines |
192
+ | `relations-resolve` | error | `(supersedes D-019)` where `D-019` doesn't exist |
193
+ | `no-self-reference` | error | A decision superseding itself |
194
+ | `decision-has-date` | error | An entry with no date, so the ledger can't be ordered |
195
+ | `protocol-files-exist` | error | A boot instruction pointing at a file that isn't there |
196
+ | `links-resolve` | error | Any relative markdown link with no target |
197
+ | `status-in-enum` | error | A status outside `todo/in-progress/done/superseded` |
198
+ | `status-not-qualified` | warn | `done (stub)` — not machine-readable |
199
+ | `index-matches-files` | error | Index and task files disagreeing, or a task missing entirely |
200
+ | `task-refs-resolve` | error | A dependency on a task that doesn't exist |
201
+ | `no-dependency-cycles` | error | Tasks that can never be executed in any order |
202
+ | `blocked-task-not-done` | warn | A task marked done while its dependency is still open |
203
+ | `no-orphan-memory-files` | warn | A memory file nothing references, so no agent reads it |
204
+
205
+ ## Configuration
206
+
207
+ `.agsync.toml`. Severity is configuration; rule logic is not.
208
+
209
+ ```toml
210
+ exclude = ["archive/"]
211
+
212
+ [rules]
213
+ status-not-qualified = "warn"
214
+ decision-has-date = "off"
215
+ ```
216
+
217
+ Also readable from `[tool.agsync]` in `pyproject.toml`.
218
+
219
+ ## CI
220
+
221
+ ```yaml
222
+ - uses: gabrielft-me/agsync@v1
223
+ ```
224
+
225
+ Or directly:
226
+
227
+ ```yaml
228
+ - run: pipx install agsync && agsync check --format github
229
+ ```
230
+
231
+ Findings render inline on the pull request diff. Pair it with branch protection
232
+ and the memory can't rot through a merge.
233
+
234
+ ## How it works
235
+
236
+ **Normalize, don't demand.** Real memory repos are written by several agents over
237
+ weeks and the format drifts: field sets vary between entries, values wrap across
238
+ lines, index tables use different column orders in the same file. The parser
239
+ absorbs all of it into a graph. Rules only ever see the graph.
240
+
241
+ **Rules are tiny and isolated.** Each is a function taking the parsed memory and
242
+ yielding findings. No shared state, no ordering, no knowledge of each other.
243
+
244
+ ```python
245
+ from agsync.rules import rule
246
+ from agsync.model import ERROR, Finding
247
+
248
+ @rule("no-undated-tasks", ERROR, "A task file carries no date.")
249
+ def no_undated_tasks(memory):
250
+ for task in memory.tasks.values():
251
+ if "date" not in task.status_raw:
252
+ yield Finding("no-undated-tasks", task.path, 1, "no date")
253
+ ```
254
+
255
+ **One report contract.** Every finding is `{rule, path, line, message, severity}`.
256
+ Text output, JSON, and GitHub annotations all derive from it, so a new output
257
+ format never touches a rule.
258
+
259
+ ## Roadmap
260
+
261
+ - [x] Linter, config, baseline, JSON/GitHub output
262
+ - [x] `init` scaffold, local hooks with chaining
263
+ - [x] `replay`: measure the decay across a repository's whole history
264
+ - [ ] `--fix`: regenerate the task index; interactive ID renumbering
265
+ - [ ] YAML front matter schema — make the D-ID ↔ task graph structured instead of prose
266
+ - [ ] Harness adapters (Claude Code, Cursor) writing commit trailers at commit time
267
+ - [ ] `agsync draft`: turn captured trailers into a ledger entry, BYO key, opened as a PR
268
+ - [ ] Single static binary
269
+
270
+ ### What this will never be
271
+
272
+ **An intent-inference engine.** A `post-receive` hook fires *after* the push —
273
+ the agent's turn is over and its context is gone. All the server sees is a diff
274
+ and a message, strictly less than the agent had. The reason a project pivoted
275
+ from one approach to another is not in the diff and no model recovers it from
276
+ there.
277
+
278
+ So: **capture at commit time, enforce at push time.** When the ledger entry
279
+ lands, the model formats a *why* that a human or agent already stated. It never
280
+ invents one. If nobody supplied a reason, the right outcome is a rejected push —
281
+ not a plausible-looking fabrication. A ledger that looks trustworthy and isn't is
282
+ worse than no ledger.
283
+
284
+ ## Contributing
285
+
286
+ See [CONTRIBUTING.md](CONTRIBUTING.md). New rules are welcome and are the easiest
287
+ place to start — one function, one test, one row in the table above.
288
+
289
+ ## License
290
+
291
+ MIT.