debrief-cli 0.1.0__py3-none-any.whl
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.
- debrief/__init__.py +0 -0
- debrief/build_artifact.py +83 -0
- debrief/classify.py +812 -0
- debrief/cli.py +291 -0
- debrief/compare_definitions.py +147 -0
- debrief/compare_llm.py +250 -0
- debrief/config.py +204 -0
- debrief/db.py +341 -0
- debrief/export.py +156 -0
- debrief/export_site.py +661 -0
- debrief/github.py +540 -0
- debrief/normalize.py +635 -0
- debrief/paths.py +17 -0
- debrief/prune.py +117 -0
- debrief/recommend.py +271 -0
- debrief/relay.py +83 -0
- debrief/runner.py +163 -0
- debrief/scrape.py +208 -0
- debrief/site/app.js +1613 -0
- debrief/site/index.html +32 -0
- debrief/site/method.js +117 -0
- debrief/site/style.css +520 -0
- debrief/skills.py +107 -0
- debrief/sources.py +341 -0
- debrief/synthesize.py +1052 -0
- debrief/verify_evidence.py +154 -0
- debrief_cli-0.1.0.dist-info/METADATA +205 -0
- debrief_cli-0.1.0.dist-info/RECORD +31 -0
- debrief_cli-0.1.0.dist-info/WHEEL +4 -0
- debrief_cli-0.1.0.dist-info/entry_points.txt +2 -0
- debrief_cli-0.1.0.dist-info/licenses/LICENSE +21 -0
|
@@ -0,0 +1,154 @@
|
|
|
1
|
+
"""Check each idea's cited evidence against the idea itself.
|
|
2
|
+
|
|
3
|
+
An idea (a row in recommendations) cites the evidence of every finding it covers, so it can
|
|
4
|
+
inherit items that are about something else: a finding that lists eight bugs, split into eight
|
|
5
|
+
backlog ideas, gives each idea all of its messages. One model call per idea reads the idea's
|
|
6
|
+
title and problem plus each cited item, resolved as export_site.py shows it, and says which
|
|
7
|
+
items directly support this idea. Results land in evidence_checks, one row per (idea, item);
|
|
8
|
+
export_site.py then counts reach from supported items only.
|
|
9
|
+
|
|
10
|
+
Ideas that already have check rows are skipped unless --force.
|
|
11
|
+
|
|
12
|
+
uv run python -m debrief.verify_evidence # latest recommendations run
|
|
13
|
+
uv run python -m debrief.verify_evidence --run 20260925T182434Z --force
|
|
14
|
+
uv run python -m debrief.verify_evidence --dry-run # print the first prompt; no model call
|
|
15
|
+
"""
|
|
16
|
+
|
|
17
|
+
import argparse
|
|
18
|
+
import json
|
|
19
|
+
import shutil
|
|
20
|
+
import sys
|
|
21
|
+
from concurrent.futures import ThreadPoolExecutor, as_completed
|
|
22
|
+
from datetime import UTC, datetime
|
|
23
|
+
|
|
24
|
+
from debrief.config import choose, load_env
|
|
25
|
+
from debrief.db import connect
|
|
26
|
+
from debrief.export_site import labels_by_turn, ref_id, resolve
|
|
27
|
+
from debrief.synthesize import ClaudeCodeError, ask_claude_code, parse_json
|
|
28
|
+
|
|
29
|
+
SYSTEM = """You check evidence for one idea from a report on how a developer or a team works with AI coding
|
|
30
|
+
agents. The idea names a specific problem. It cites items: human messages in agent sessions (with the agent
|
|
31
|
+
text just before them), failed CI runs (with a log tail) and PR review comments.
|
|
32
|
+
|
|
33
|
+
For each item, decide whether it directly supports this idea's specific claim: the item shows the problem
|
|
34
|
+
the idea describes happening, or names the exact bug. An item about a different bug, a different file or
|
|
35
|
+
area, or a different kind of problem does not support it, even if it is related or came from the same
|
|
36
|
+
session. Judge each item on its own content, not on the idea's wording.
|
|
37
|
+
|
|
38
|
+
Reply with JSON only:
|
|
39
|
+
{"supports": ["<item id>", ...], "reasons": {"<item id>": "<under 15 words: why it does or doesn't>"}}
|
|
40
|
+
Give a reason for every item."""
|
|
41
|
+
|
|
42
|
+
|
|
43
|
+
def item_view(ref: dict) -> dict:
|
|
44
|
+
"""The parts of a resolved ref the model reads."""
|
|
45
|
+
if ref["type"] == "msg":
|
|
46
|
+
return {"type": "human message", "agent_before": ref["agent_before"], "text": ref["text"]}
|
|
47
|
+
if ref["type"] == "check":
|
|
48
|
+
return {"type": "failed CI run", "check": ref["name"], "failure_class": ref["failure_class"],
|
|
49
|
+
"log_tail": ref["log"]}
|
|
50
|
+
if ref["type"] == "review":
|
|
51
|
+
return {"type": "review comment", "path": ref["path"], "body": ref["body"], "held_up": ref["held_up"]}
|
|
52
|
+
return {"type": "missing"}
|
|
53
|
+
|
|
54
|
+
|
|
55
|
+
def idea_refs(conn, labels, rec, findings: dict[int, list[dict]]) -> dict[str, dict]:
|
|
56
|
+
"""ref id -> resolved ref, over the idea's findings, in the order export_site.py lists them."""
|
|
57
|
+
out: dict[str, dict] = {}
|
|
58
|
+
for fid in json.loads(rec["finding_ids"] or "[]"):
|
|
59
|
+
for ref in findings.get(fid, []):
|
|
60
|
+
if ref_id(ref) not in out:
|
|
61
|
+
out[ref_id(ref)] = resolve(conn, labels, ref)
|
|
62
|
+
return out
|
|
63
|
+
|
|
64
|
+
|
|
65
|
+
def prompt(rec, refs: dict[str, dict]) -> tuple[str, dict[str, str]]:
|
|
66
|
+
"""The user prompt, and short item ids (i1, i2, ...) -> ref ids."""
|
|
67
|
+
short = {f"i{n}": r for n, r in enumerate(refs, 1)}
|
|
68
|
+
body = {"idea": {"title": rec["title"], "problem": rec["problem"]},
|
|
69
|
+
"items": [{"id": s, **item_view(refs[r])} for s, r in short.items()]}
|
|
70
|
+
return json.dumps(body, ensure_ascii=False), short
|
|
71
|
+
|
|
72
|
+
|
|
73
|
+
def parse(text: str) -> list[dict]:
|
|
74
|
+
reply = parse_json(text)
|
|
75
|
+
return [reply] if isinstance(reply, dict) and isinstance(reply.get("supports"), list) else []
|
|
76
|
+
|
|
77
|
+
|
|
78
|
+
def main(argv: list[str] | None = None) -> None:
|
|
79
|
+
parser = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
|
|
80
|
+
parser.add_argument("--run", help="Recommendations run_id (default: latest)")
|
|
81
|
+
parser.add_argument("--force", action="store_true", help="Check ideas that already have check rows")
|
|
82
|
+
parser.add_argument("--parallel", type=int, default=3, help="Concurrent model calls (default: %(default)s)")
|
|
83
|
+
parser.add_argument("--dry-run", action="store_true", help="Print the first prompt; no model call")
|
|
84
|
+
args = parser.parse_args(argv)
|
|
85
|
+
load_env()
|
|
86
|
+
|
|
87
|
+
conn = connect()
|
|
88
|
+
run_id = args.run or conn.execute("SELECT MAX(run_id) FROM recommendations").fetchone()[0]
|
|
89
|
+
if not run_id:
|
|
90
|
+
sys.exit("no recommendations yet; run recommend.py")
|
|
91
|
+
findings = {r["finding_id"]: json.loads(r["evidence"] or "[]")
|
|
92
|
+
for r in conn.execute("SELECT finding_id, evidence FROM findings WHERE run_id = ?", (run_id,))}
|
|
93
|
+
recs = conn.execute("SELECT * FROM recommendations WHERE run_id = ? ORDER BY priority", (run_id,)).fetchall()
|
|
94
|
+
checked = {r[0] for r in conn.execute("SELECT DISTINCT rec_id FROM evidence_checks")}
|
|
95
|
+
todo = [r for r in recs if args.force or r["rec_id"] not in checked]
|
|
96
|
+
print(f"run {run_id}: {len(recs)} ideas, {len(recs) - len(todo)} already checked, {len(todo)} to check")
|
|
97
|
+
if not todo:
|
|
98
|
+
return
|
|
99
|
+
|
|
100
|
+
labels = labels_by_turn(conn)
|
|
101
|
+
jobs = {}
|
|
102
|
+
for rec in todo:
|
|
103
|
+
refs = idea_refs(conn, labels, rec, findings)
|
|
104
|
+
if refs:
|
|
105
|
+
jobs[rec["rec_id"]] = (rec, refs, *prompt(rec, refs))
|
|
106
|
+
if not jobs:
|
|
107
|
+
print("no idea cites any evidence; nothing to check")
|
|
108
|
+
return
|
|
109
|
+
if args.dry_run:
|
|
110
|
+
_, _, user, _ = next(iter(jobs.values()))
|
|
111
|
+
print(f"~{sum(len(j[2]) for j in jobs.values()) // 4} input tokens over {len(jobs)} calls\n")
|
|
112
|
+
print(user[:6000])
|
|
113
|
+
return
|
|
114
|
+
if not shutil.which("claude"):
|
|
115
|
+
sys.exit("claude (Claude Code) is not on PATH")
|
|
116
|
+
|
|
117
|
+
_, model = choose("verify")
|
|
118
|
+
now = datetime.now(UTC).isoformat()
|
|
119
|
+
cost, kept, total = 0.0, 0, 0
|
|
120
|
+
with ThreadPoolExecutor(max_workers=args.parallel) as pool:
|
|
121
|
+
futures = {pool.submit(ask_claude_code, user, model, system=SYSTEM, parse=parse): rec_id
|
|
122
|
+
for rec_id, (_, _, user, _) in jobs.items()}
|
|
123
|
+
for fut in as_completed(futures):
|
|
124
|
+
rec, refs, _, short = jobs[futures[fut]]
|
|
125
|
+
try:
|
|
126
|
+
replies, usage = fut.result()
|
|
127
|
+
except ClaudeCodeError as e:
|
|
128
|
+
print(f"#{rec['rec_id']} failed: {e}")
|
|
129
|
+
continue
|
|
130
|
+
cost += usage["cost"]
|
|
131
|
+
if not replies:
|
|
132
|
+
print(f"#{rec['rec_id']} unreadable reply; no rows saved")
|
|
133
|
+
continue
|
|
134
|
+
reply = replies[0]
|
|
135
|
+
unknown = [s for s in reply["supports"] if s not in short]
|
|
136
|
+
if unknown:
|
|
137
|
+
print(f"#{rec['rec_id']} ignored unknown item ids: {unknown}")
|
|
138
|
+
yes = {short[s] for s in reply["supports"] if s in short}
|
|
139
|
+
reasons = reply.get("reasons") if isinstance(reply.get("reasons"), dict) else {}
|
|
140
|
+
conn.execute("DELETE FROM evidence_checks WHERE rec_id = ?", (rec["rec_id"],))
|
|
141
|
+
conn.executemany(
|
|
142
|
+
"INSERT INTO evidence_checks (rec_id, ref, supports, reason, model, created_at) "
|
|
143
|
+
"VALUES (?, ?, ?, ?, ?, ?)",
|
|
144
|
+
[(rec["rec_id"], r, int(r in yes), str(reasons.get(s) or ""), model, now) for s, r in short.items()])
|
|
145
|
+
conn.commit()
|
|
146
|
+
kept += len(yes)
|
|
147
|
+
total += len(refs)
|
|
148
|
+
print(f"#{rec['rec_id']} {len(yes)}/{len(refs)} supported {rec['title']} (${usage['cost']:.4f})")
|
|
149
|
+
print(f"\n{kept} of {total} cited items support their idea; {total - kept} rejected. "
|
|
150
|
+
f"~${cost:.4f} API-equivalent ({model} via claude -p)")
|
|
151
|
+
|
|
152
|
+
|
|
153
|
+
if __name__ == "__main__":
|
|
154
|
+
main()
|
|
@@ -0,0 +1,205 @@
|
|
|
1
|
+
Metadata-Version: 2.5
|
|
2
|
+
Name: debrief-cli
|
|
3
|
+
Version: 0.1.0
|
|
4
|
+
Summary: Review your coding-agent sessions: what went wrong, and what to change.
|
|
5
|
+
Project-URL: Homepage, https://github.com/mvwi/debrief
|
|
6
|
+
Project-URL: Issues, https://github.com/mvwi/debrief/issues
|
|
7
|
+
License-Expression: MIT
|
|
8
|
+
License-File: LICENSE
|
|
9
|
+
Keywords: claude-code,codex,coding-agent,report,session-logs
|
|
10
|
+
Classifier: Development Status :: 3 - Alpha
|
|
11
|
+
Classifier: Environment :: Console
|
|
12
|
+
Classifier: Intended Audience :: Developers
|
|
13
|
+
Classifier: Programming Language :: Python :: 3.13
|
|
14
|
+
Classifier: Topic :: Software Development
|
|
15
|
+
Requires-Python: >=3.13
|
|
16
|
+
Requires-Dist: httpx>=0.28.1
|
|
17
|
+
Requires-Dist: python-dotenv>=1.2.3
|
|
18
|
+
Description-Content-Type: text/markdown
|
|
19
|
+
|
|
20
|
+
# debrief
|
|
21
|
+
|
|
22
|
+
Review your coding-agent sessions: where you had to correct the agent, what it missed, and what
|
|
23
|
+
to change so it happens less.
|
|
24
|
+
|
|
25
|
+
debrief reads the session logs Claude Code and Codex already keep on your machine, labels each
|
|
26
|
+
of your turns, groups the labels into findings with evidence, and builds a local report site.
|
|
27
|
+
|
|
28
|
+
## Install
|
|
29
|
+
|
|
30
|
+
With Homebrew on a Mac:
|
|
31
|
+
|
|
32
|
+
brew install mvwi/tap/debrief
|
|
33
|
+
|
|
34
|
+
With uv or pipx, on Python 3.13 or newer:
|
|
35
|
+
|
|
36
|
+
uv tool install debrief-cli
|
|
37
|
+
pipx install debrief-cli
|
|
38
|
+
|
|
39
|
+
The package on PyPI is `debrief-cli`; the command it installs is `debrief`. The PyPI package
|
|
40
|
+
named `debrief` is a different project, so `uvx debrief` and `pip install debrief` fetch the
|
|
41
|
+
wrong tool. To run debrief once without installing it:
|
|
42
|
+
|
|
43
|
+
uvx --from debrief-cli debrief open
|
|
44
|
+
|
|
45
|
+
From source:
|
|
46
|
+
|
|
47
|
+
git clone https://github.com/mvwi/debrief && cd debrief
|
|
48
|
+
uv tool install .
|
|
49
|
+
|
|
50
|
+
The default labeler runs through `claude -p` on your Claude subscription, so it needs a
|
|
51
|
+
logged-in `claude` CLI. Without one, pick OpenRouter and save a key on the Settings page of
|
|
52
|
+
`debrief open`. `gh` is optional: with it, `debrief sync` adds pull-request data.
|
|
53
|
+
|
|
54
|
+
## Use
|
|
55
|
+
|
|
56
|
+
debrief sync --dry-run # ingest logs, then say how many model calls labeling would take
|
|
57
|
+
debrief sync # ingest, label, find patterns, build the report
|
|
58
|
+
debrief open # serve the report at http://127.0.0.1:8765 and open it
|
|
59
|
+
|
|
60
|
+
Useful flags for `sync`:
|
|
61
|
+
|
|
62
|
+
--project DIR add this project to the ones sync analyses
|
|
63
|
+
--since YYYY-MM-DD only logs and evidence from this date on
|
|
64
|
+
--import DIR also load a folder of JSONL logs (Claude Code, Codex or Pi format)
|
|
65
|
+
--include-headless also load scripted runs (see below)
|
|
66
|
+
--llm openrouter label and synthesize through OpenRouter, whatever config.toml says
|
|
67
|
+
|
|
68
|
+
A re-run only reads logs that changed and only labels turns that are new.
|
|
69
|
+
|
|
70
|
+
You pick which projects get analysed. A project is the folder a session ran in. The first
|
|
71
|
+
`debrief sync` reads your logs, lists the projects it found with their size, and stops:
|
|
72
|
+
|
|
73
|
+
debrief sync --project ~/Dev/myrepo # pick it (kept in config.toml), then analyse
|
|
74
|
+
|
|
75
|
+
Only picked projects are labeled, analysed and shown. Each gets its own findings and ideas,
|
|
76
|
+
and an idea names the project it lands in (`myrepo: CLAUDE.md`). The picker in the report's
|
|
77
|
+
side bar scopes every page to one project. Findings about how you prompt come from the picked
|
|
78
|
+
projects together. To drop a project:
|
|
79
|
+
|
|
80
|
+
debrief config --drop ~/Dev/myrepo # its findings stay in the database, hidden
|
|
81
|
+
|
|
82
|
+
Unticking it on the Projects page does the same.
|
|
83
|
+
|
|
84
|
+
`debrief <stage> --help` runs one stage on its own: `ingest`, `label`, `synthesize`,
|
|
85
|
+
`recommend`, `verify`, `export`, `prune`.
|
|
86
|
+
|
|
87
|
+
## What gets loaded
|
|
88
|
+
|
|
89
|
+
- Claude Code logs in `~/.claude/projects` and Codex logs in `~/.codex/sessions`.
|
|
90
|
+
- Scripted runs are skipped by default: `claude -p`, `codex exec`, and sessions started through
|
|
91
|
+
an SDK or a wrapper app that have fewer than two turns from a person. On the author's machine
|
|
92
|
+
that was 1,038 of 1,274 logs.
|
|
93
|
+
- Sub-agent logs are skipped, so tool calls made inside a sub-agent are not counted.
|
|
94
|
+
- A resumed or forked session copies its history into the new log. Each turn is loaded once.
|
|
95
|
+
|
|
96
|
+
## Where your data goes
|
|
97
|
+
|
|
98
|
+
- Everything debrief writes is under `~/.local/share/debrief` (set `DEBRIEF_HOME` to move it).
|
|
99
|
+
- Labeling and synthesis send transcript text to a model: by default Claude, through your own
|
|
100
|
+
`claude` CLI. With `--llm openrouter` it goes to OpenRouter and needs `OPENROUTER_API_KEY`.
|
|
101
|
+
- `debrief sync` asks GitHub about the pull requests your sessions produced, through your own
|
|
102
|
+
logged-in `gh` CLI and read-only. It sends repo names, PR numbers, branch names and commit
|
|
103
|
+
SHAs, and nothing from your transcripts. `--no-github`, or `[github] enabled = false` in
|
|
104
|
+
config.toml, turns it off; without `gh` it is skipped.
|
|
105
|
+
- The report holds your transcripts. `debrief open` serves it on 127.0.0.1 only. Token-shaped
|
|
106
|
+
strings (GitHub, Anthropic, OpenAI, AWS, Slack, Stripe, private keys and others) are replaced
|
|
107
|
+
with `[redacted]` in the report, but the redaction is pattern-based: read before you share.
|
|
108
|
+
|
|
109
|
+
## Which labeler
|
|
110
|
+
|
|
111
|
+
| | Claude Haiku (default) | Jev (`--llm openrouter`) |
|
|
112
|
+
|---|---|---|
|
|
113
|
+
| Needs | a Claude subscription | an OpenRouter key |
|
|
114
|
+
| Sends transcripts to | Anthropic, via your CLI | OpenRouter |
|
|
115
|
+
| Confidence | the model's own estimate | token probabilities |
|
|
116
|
+
| 2,774 items (236 sessions) | 33 min, no extra charge | about 3 min, about $0.37 |
|
|
117
|
+
|
|
118
|
+
The Jev time and cost are scaled up from a 74-item run (4 s, $0.0098).
|
|
119
|
+
|
|
120
|
+
To pick the model for each stage, write `~/.local/share/debrief/config.toml`:
|
|
121
|
+
|
|
122
|
+
[label]
|
|
123
|
+
llm = "openrouter" # Jev: faster and closer to a careful reader, needs a key
|
|
124
|
+
[synthesize]
|
|
125
|
+
model = "claude-sonnet-5" # optional; each llm has a default
|
|
126
|
+
|
|
127
|
+
Stages are `label`, `synthesize`, `recommend` and `verify`. `debrief config` prints what each
|
|
128
|
+
will use. A flag beats an environment variable, which beats the file.
|
|
129
|
+
|
|
130
|
+
### Settings page
|
|
131
|
+
|
|
132
|
+
`debrief open` has a Settings page that does the same without editing files. It picks the
|
|
133
|
+
provider and model for each stage, turns the GitHub lookups on or off, and saves an OpenRouter
|
|
134
|
+
key.
|
|
135
|
+
|
|
136
|
+
Saving needs the link `debrief open` prints, which carries a token for that run: someone else
|
|
137
|
+
on the same machine can reach the port but cannot change your settings. They can still read the
|
|
138
|
+
report, as they can read the files.
|
|
139
|
+
|
|
140
|
+
- Saving rewrites `config.toml`; comments you wrote in it are lost.
|
|
141
|
+
- The key goes in `~/.local/share/debrief/.env`, readable only by you. The page shows its last
|
|
142
|
+
four characters and never the key. `OPENROUTER_API_KEY` set in your shell wins over the saved one.
|
|
143
|
+
- The server takes these writes only from a page it served itself, on `127.0.0.1`.
|
|
144
|
+
|
|
145
|
+
### Run from the page
|
|
146
|
+
|
|
147
|
+
`debrief open` can also run the analysis, so you need not go back to the terminal. With no
|
|
148
|
+
report yet it reads your local logs first (no model calls) and serves what it found.
|
|
149
|
+
|
|
150
|
+
The Projects page is where you do it: tick the projects to analyse (each tick is saved to
|
|
151
|
+
`config.toml` at once), ask for an estimate, then press "Check my data now". It shows the stage
|
|
152
|
+
and the count while the run goes, and has a Cancel button. When the run ends every page shows
|
|
153
|
+
the new report without a refresh. Underneath it uses:
|
|
154
|
+
|
|
155
|
+
GET /api/run state (idle, running, done, failed, cancelled), stage, done/total,
|
|
156
|
+
start and end times, exit code, last 30 lines of output
|
|
157
|
+
POST /api/run start `debrief sync`; 409 if a run or an estimate is going
|
|
158
|
+
DELETE /api/run stop the run and the `claude -p` calls it started
|
|
159
|
+
POST /api/estimate run `debrief sync --dry-run`, rebuild the pages, return the counts
|
|
160
|
+
|
|
161
|
+
- The writes need the same token as the Settings page. Nothing in a request reaches the command
|
|
162
|
+
line: what to analyse comes from `config.toml`.
|
|
163
|
+
- One run at a time. Its output is in `~/.local/share/debrief/run.log`, replaced on each run,
|
|
164
|
+
with the OpenRouter key blanked.
|
|
165
|
+
- Ctrl-C stops a run too. If the server is killed outright, the next `debrief open` or the
|
|
166
|
+
next start stops the run it left behind.
|
|
167
|
+
- A stopped run loses little: labels and bundles are kept, so the next run goes on from there.
|
|
168
|
+
|
|
169
|
+
How they compare, measured on the prototype's data with no human-labeled set behind it:
|
|
170
|
+
|
|
171
|
+
- On 66 turns where the two disagreed, a blind reader sided with Jev 32 times and Haiku 17 on
|
|
172
|
+
the turn-level questions. Three definitions were then tightened.
|
|
173
|
+
- After that, on 155 ordinary turns, Haiku and Jev agree on 93% for "agent named a problem and
|
|
174
|
+
did not file it" (kappa 0.77) and 92% for "the agent could have found this itself" (kappa
|
|
175
|
+
0.66), where Haiku says yes about 1.7 times as often as Jev, mostly at low confidence.
|
|
176
|
+
|
|
177
|
+
## Limits in this version
|
|
178
|
+
|
|
179
|
+
- A local session is linked to a pull request only when its log shows one: the URL that
|
|
180
|
+
`gh pr create` printed, or a branch it pushed that is the head of a PR. On the author's
|
|
181
|
+
machine that linked 33 of the 53 sessions that pushed. Of the 20 left out, 12 pushed
|
|
182
|
+
straight to main or another trunk branch.
|
|
183
|
+
- A session that pushes straight to main gets no CI or review data, even when CI runs on
|
|
184
|
+
those commits. If you work alone and never open pull requests, the report will show no
|
|
185
|
+
CI for you. The same goes for a PR opened by hand from a branch pushed outside the session.
|
|
186
|
+
- A CI failure counts for a session when its commit was made while the session ran (give or
|
|
187
|
+
take ten minutes). Review comments count for every session linked to the PR.
|
|
188
|
+
- Sessions fetched from Entire still need `debrief github --repo owner/name`; `sync` does not
|
|
189
|
+
fetch their GitHub data.
|
|
190
|
+
- A session's project is the folder it ran in, or the project of the repo it opened a PR in.
|
|
191
|
+
A session run in a git worktree joins its repo's project when ingest saw the worktree
|
|
192
|
+
while it still existed, or when the log names the repo, as Codex logs do. A Claude Code
|
|
193
|
+
session in a worktree that was deleted before the first ingest, and that opened no PR,
|
|
194
|
+
shows up as a project of its own, named after the worktree folder.
|
|
195
|
+
- A picked project that has not been synced yet shows in the report with no labels.
|
|
196
|
+
- "How it works" describes the whole run; the project picker does not scope its numbers.
|
|
197
|
+
- Labels vary between runs: on a set of hard cases Haiku changed its answer on about 15% of
|
|
198
|
+
rows between two identical runs.
|
|
199
|
+
- Pi logs are parsed but only tested against a made-up log.
|
|
200
|
+
|
|
201
|
+
## Develop
|
|
202
|
+
|
|
203
|
+
uv sync
|
|
204
|
+
uv run pytest
|
|
205
|
+
uv run ruff check .
|
|
@@ -0,0 +1,31 @@
|
|
|
1
|
+
debrief/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
|
|
2
|
+
debrief/build_artifact.py,sha256=sgP-qdr1zNXhIwzbjcKjluapxkPzZUHo8H_PxQOjAy0,3410
|
|
3
|
+
debrief/classify.py,sha256=WdIb5fZI35ZUar1VhmhoGH2hKUCrOplBIfKycDdLc0Y,41694
|
|
4
|
+
debrief/cli.py,sha256=l3IaBXd6ruPXQ2zgVKSJlraolvfZbuHfUmqNOxirvqI,13864
|
|
5
|
+
debrief/compare_definitions.py,sha256=MtfXjpGse2Y8ofMfiAhNdi_g0wv4daIB1SstQkVKciA,7591
|
|
6
|
+
debrief/compare_llm.py,sha256=bC940jTLiM8B2YbtDDHkwtaoK4VAwc1eBzfec3f354s,13654
|
|
7
|
+
debrief/config.py,sha256=_EKjd9-UhGMWetnjiQLJSxiEbUVQ2_XFJJVf1dt7D48,8886
|
|
8
|
+
debrief/db.py,sha256=P3WGrLN7k-JjgN6PRa59aRhvk0kJw0SR1k7mVkLmCFM,15112
|
|
9
|
+
debrief/export.py,sha256=yg0RefQv-us0MqPiwxfWyA6qjUAjtmhAZmlXgGz-t2s,6894
|
|
10
|
+
debrief/export_site.py,sha256=MeDZfOehFTGnLgzzpHXJ2iznv3RHhOXZxTIoKh4oj2Y,35854
|
|
11
|
+
debrief/github.py,sha256=4iTigTeki2crbOmUzI65C_nEFobGauGtn1clcoWI5rs,27791
|
|
12
|
+
debrief/normalize.py,sha256=nAlPTjETkjQt0vq44jT5ALXX6HM-ARsZYRt88bdZNPI,32045
|
|
13
|
+
debrief/paths.py,sha256=Ih_yPDRRTOXnjdKq8_nkfHXETIadE4UAtn9slg6QmpE,852
|
|
14
|
+
debrief/prune.py,sha256=yOande7YGWWS5Oxu2FPccz-ocCWNLV2kCAgrli-LA88,6622
|
|
15
|
+
debrief/recommend.py,sha256=nn_4cyH_7b3UB25CzJqqo3jBVUzZmvUUH_4AfMbUSfc,15449
|
|
16
|
+
debrief/relay.py,sha256=HQxgh-770-7Bh3zyiXn-yj5ipAGHnO8VIAqekOxLK94,4826
|
|
17
|
+
debrief/runner.py,sha256=4CCMxHoqe6xzv4oc5qd1NlOwOycwcObFGqJvfHfJTdU,6968
|
|
18
|
+
debrief/scrape.py,sha256=cvJ2HGKAdfM8XWjxem1E4y-nTDoVsMwoamidtd4Pjig,9096
|
|
19
|
+
debrief/skills.py,sha256=ML5UzD2S2LnguoZIxtGpUPWn_2K5xqX0u5NGeksksrQ,4931
|
|
20
|
+
debrief/sources.py,sha256=i0KjBgE-7Jw_5tygNOet1iBvI8N3EizsYX3fbSdomg0,20469
|
|
21
|
+
debrief/synthesize.py,sha256=Jwifsyoy-XDgVsnkvRYmIndhRAk_ka2YtxXFZYS_FzA,58050
|
|
22
|
+
debrief/verify_evidence.py,sha256=xT_gY9d41sejDS2sHp0V1ZkKkr9hZO-TPXfkwcu-kJc,7648
|
|
23
|
+
debrief/site/app.js,sha256=EZtp_WyHzJwkJgQr13LvZIwFZXNknBMiIsPht2YQMbU,121129
|
|
24
|
+
debrief/site/index.html,sha256=8VVs4ZUAivzRolea3r-gqUyR_JDGTETP5oQB-J8EOqU,1150
|
|
25
|
+
debrief/site/method.js,sha256=vA_J2apgriH9xDvLZbjFCYMUiT_s45VQmdsNONMKtmU,9814
|
|
26
|
+
debrief/site/style.css,sha256=tjJwve_do_VUgjvGjuXcW4AfmHBq65pfsVcMk2nKI6Q,36657
|
|
27
|
+
debrief_cli-0.1.0.dist-info/METADATA,sha256=1XbDEHLNTxXftKPnXllVkwKGqZD6F0aNFMdPVCTqUzU,10296
|
|
28
|
+
debrief_cli-0.1.0.dist-info/WHEEL,sha256=2iZrJkDIkeDh0k9SikKMqFuRP-ua6mb4Ir1AsER4BEw,87
|
|
29
|
+
debrief_cli-0.1.0.dist-info/entry_points.txt,sha256=Y7nRq52bggFLkBSmLUQ_ZKpsMtKmdtbmP8LCoriT_Ek,45
|
|
30
|
+
debrief_cli-0.1.0.dist-info/licenses/LICENSE,sha256=wz5XvWjejmAu9Am30iUrerq0X5Py2d1Tc-l0wlFonUA,1073
|
|
31
|
+
debrief_cli-0.1.0.dist-info/RECORD,,
|
|
@@ -0,0 +1,21 @@
|
|
|
1
|
+
MIT License
|
|
2
|
+
|
|
3
|
+
Copyright (c) 2026 Michael Williams
|
|
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.
|