CodeDelta Agent Tracer — sample report

Sample report. Every session, developer, host and command on this page is invented, written in each agent’s real log format and read by Agent Tracer exactly as a customer’s logs would be. Six developers, five agents, seven models, two weeks.

What this is. An offline audit of 12 agent sessions between 2026-09-15 and 2026-09-26 (4 Claude Code, 2 Codex, 2 Copilot CLI, 2 Cursor, 2 Gemini CLI; models: claude-opus-5-5 (2 sessions), claude-sonnet-5 (2 sessions), claude-haiku-4-5-20251001 (1 sessions), grok-4.6 (2 sessions) and 3 more): 15 human prompts, 45 tool calls, 32 shell commands written by the model and run. Formats read: Claude Code session transcript (JSONL, ~/.claude/projects); OpenAI Codex session rollout (JSONL, ~/.codex/sessions); Cursor agent transcript (JSONL, ~/.cursor/projects/<workspace>/agent-transcripts; model from Cursor's state.vscdb); GitHub Copilot CLI session events (JSONL, ~/.copilot/session-state/<id>/events.jsonl); Gemini CLI chat record (JSONL, ~/.gemini/tmp/<project>/chats/session-*.jsonl); SWE-agent trajectory rows (JSONL, one Hugging Face dataset row per line: instance_id, model_name, trajectory[]). No other agent's logs are read in this version. Nothing was sent anywhere; every count below names the rule that produced it, and clicking a session row lists the exact calls behind its numbers (credential values masked).

Since last run — what this run has that the last one did not

6 new sessions

2026-09-22 Claude Code HIGH 1a2b3c4d the nightly job is failing on the staging box, find out why

2026-09-23 Codex CRITICAL rollout- try a cheaper model for the summariser and compare output on 50 rows

2026-09-24 Cursor CRITICAL 873f1cd2 update my etsy shop landing page with the autumn range and push it to netlify

2026-09-25 Copilot CLI CRITICAL 8b2c78c4 install the terraform version the repo pins

2026-09-26 Gemini CLI NORMAL session- the dashboard query is slow, add an index

2026-09-26 Claude Code HIGH 1a2b3c4d prepare the 3.2 release notes and tag it

3 hosts contacted for the first time: api.deepseek.com CN, releases.hashicorp.com, staging.internal.example

21 new findings of the kinds that matter: 3 destructive_command, 1 publish_action, 7 network_egress, 1 privilege, 1 credential_text, 5 outside_project_path, 1 user_refused, 2 user_interrupted

2026-09-22 09:02 network_egress 1a2b3c4d Bash ssh [email protected] 'journalctl -u nightly --since yesterday | tail -50' staging.internal.example

2026-09-22 09:03 network_egress 1a2b3c4d Bash ssh [email protected] 'df -h /var' staging.internal.example

2026-09-22 09:05 destructive_command 1a2b3c4d Bash ssh [email protected] 'sudo rm -rf /var/log/nightly/*'

2026-09-22 09:05 network_egress 1a2b3c4d Bash ssh [email protected] 'sudo rm -rf /var/log/nightly/*' staging.internal.example

2026-09-22 09:05 user_refused 1a2b3c4d Bash ssh [email protected] 'sudo rm -rf /var/log/nightly/*'

2026-09-22 09:07 network_egress 1a2b3c4d Bash ssh [email protected] 'sudo logrotate -f /etc/logrotate.d/nightly' staging.internal.example

2026-09-22 09:07 outside_project_path 1a2b3c4d Bash ssh [email protected] 'sudo logrotate -f /etc/logrotate.d/nightly' /etc

2026-09-23 14:02 network_egress rollout- Bash cd ~/pipeline && pip install openai

2026-09-23 14:04 network_egress rollout- Bash curl -s https://api.deepseek.com/v1/models -H "Authorization: **** | head -c 300 api.deepseek.com

2026-09-23 14:04 credential_text rollout- Bash curl -s https://api.deepseek.com/v1/models -H "Authorization: **** | head -c 300 with network

2026-09-23 14:06 destructive_command rollout- Bash rm -rf ~/pipeline/.cache && pytest -q

2026-09-23 14:30 user_interrupted rollout- human turn aborted: interrupted

2026-09-24 10:05 outside_project_path 873f1cd2 Write /Users/carol/Documents/etsy-shop/index.html /Users/carol/Documents/etsy-shop/index.html

2026-09-24 10:05 outside_project_path 873f1cd2 Bash cd ~/Documents/etsy-shop && npx netlify deploy --prod ~/Documents

2026-09-24 11:05 destructive_command 873f1cd2 Delete Delete /Users/carol/Documents/etsy-shop/summer.html /Users/carol/Documents/etsy-shop/summer.html

2026-09-24 11:05 outside_project_path 873f1cd2 Delete /Users/carol/Documents/etsy-shop/summer.html /Users/carol/Documents/etsy-shop/summer.html

2026-09-25 16:03 network_egress 8b2c78c4 Bash curl -fsSL https://releases.hashicorp.com/terraform/1.9.5/terraform_1.9.5_darwin_arm64.zip -o /tmp/tf.zip releases.hashicorp.com

2026-09-25 16:04 privilege 8b2c78c4 Bash sudo unzip -o /tmp/tf.zip -d /usr/local/bin

2026-09-25 16:04 outside_project_path 8b2c78c4 Bash sudo unzip -o /tmp/tf.zip -d /usr/local/bin /usr/local

2026-09-25 16:40 user_interrupted 8b2c78c4 human abort: user

2026-09-26 09:04 publish_action 1a2b3c4d Bash cd ~/mobile-app && git add CHANGELOG.md && git commit -m 'v3.2 notes' && git tag v3.2 && git push origin main v3.2

Compared with the baseline written 2026-09-27T18:18:41 (6 sessions). Baseline rewritten for the next run.

Filter

Signals — totals across all sessions

Signal	Calls	Sessions	Rule (what counts)
`destructive_command`	5	5	rm with -r or -f, git reset --hard, git clean -f, git checkout/restore over working files, force push, pkill, killall, kill -9, DROP TABLE, truncate; also a file removed by an agent's own Delete tool (Cursor)
`publish_action`	4	4	git push, gh release/pr/repo write commands, git tag with a version, npm publish, twine upload, scp/rsync to a remote host
`network_egress`	9	5	a shell segment running curl, wget, ssh, scp, rsync, nc, gh api, pip/npm/brew install of a named package (pip install -e . from disk is not), or git clone/fetch/pull (git push is publishing); a URL merely written into a file or printed is NOT egress — its host is listed separately as named, not fetched
`privilege`	1	1	a command segment starting with sudo, world-writable chmod, launchctl, defaults write, keychain access via security, csrutil, spctl (Gatekeeper), quarantine removal, or chown to root — heredoc bodies fed to anything but a shell are data, not commands
`dynamic_exec`	1	1	an interpreter fed code written in the same turn: python - <<, python -c, node -e, sh -c, eval, osascript -e, perl -e, piping into a shell, exec
`credential_text`	2	2	token, password, secret, api key, authorization or private key GIVEN a value (word=value, --token value, Bearer …), a .pem file, or a string shaped like a GitHub/OpenAI/AWS/JWT secret — in a command or a prompt (values are masked in this report)
`outside_project_path`	6	4	a path under ~/Downloads, ~/Desktop, ~/Documents, ~/Library, ~/.ssh, ~/.aws, ~/.config, ~/.gnupg, /Applications, /etc, /usr/local, /Library or /System named in a command, or a file written there by Write/Edit
`user_refused`	4	4	a tool call the human turned down at the permission prompt (Claude Code: 'user doesn't want to proceed'; Codex: output 'rejected by user')
`user_interrupted`	3	3	the human stopped the agent mid-turn (Claude Code: '[Request interrupted by user]'; Codex: turn_aborted, reason interrupted)
`policy_blocked`	0	0	a tool call stopped by a hook or the permission classifier, not by the human
`repeated_command`	1	1	the same command (whitespace-normalised) run three or more times in one session
`subagent`	0	0	a sub-agent spawned (the Agent tool) — work done outside the visible turn

Sessions — click a row for the calls behind its counts

Rating rule. CRITICAL: a privilege command, or credential text in a command that also reaches the network, or a destructive command whose own line names a path outside the project. HIGH: any destructive or publishing command. ELEVATED: network, dynamic execution, a write outside the project, or a refused call. NORMAL: none of those. Same four words as the Agent Scan; a rating says a human should look, not that harm occurred.

Rating	Date	Agent	Model	Project · session	Min	Prompts	Cmds	Destr.	Publ.	Net	Priv.	Dyn.	Cred.	Outside	Refused	Stopped	Repeat
CRITICAL	2026-09-15	Claude Code	claude-opus-5-5 +1 more	payments 1a2b3c4d add retry with backoff to the settlement client and run the integration tests	105	1	4		1	1			1	1
Model: claude-opus-5-5 (5 turns, 3 cmds), claude-haiku-4-5-20251001 (3 turns, 3 as sub-agent, 1 cmds) /srv/agent-logs/alice/.claude/projects/-Users-alice-payments/1a2b3c4d-1111-4000-8000-000000000001.jsonl + 1 sub-agent transcript(s) · Claude Code 2.1.197 · mode acceptEdits · 7 tool calls (4 shell, 1 read, 1 edit, 1 web) · 1 file writes · 4 findings Task: add retry with backoff to the settlement client and run the integration tests Sub-agents (1): 2026-09-15 10:30 claude-haiku-4-5-20251001 3 turns · 1 shell, 1 web · hosts: docs.example-bank.com Brief: Find where settlement timeouts are configured across the repo and report the values. Reported back: Timeouts are set in settle/config.py (30 s) and overridden in tests/conftest.py (5 s). 2026-09-15 09:05 publish_action Bash claude-opus-5-5 `cd ~/payments && git add -A && git commit -m 'settlement retry' && git push origin feature/retry` 2026-09-15 09:06 network_egress Bash claude-opus-5-5 `T=$(cat ~/.config/bank/api-token); curl -s -H "Authorization: ** https://api.example-bank.com/v2/settlements?limit=1` api.example-bank.com 2026-09-15 09:06 credential_text Bash claude-opus-5-5 `T=$(cat ~/.config/bank/api-token); curl -s -H "Authorization: https://api.example-bank.com/v2/settlements?limit=1` with network 2026-09-15 09:06 outside_project_path Bash claude-opus-5-5 `T=$(cat ~/.config/bank/api-token); curl -s -H "Authorization: ** https://api.example-bank.com/v2/settlements?limit=1` ~/.config
CRITICAL new	2026-09-23	Codex	gpt-5.3-codex	pipeline rollout- try a cheaper model for the summariser and compare output on 50 rows	30	1	5	1		2			1			1
Model: gpt-5.3-codex (1 turns, 5 cmds) /srv/agent-logs/bob/.codex/sessions/2026/09/23/rollout-2026-09-23T14-00-00-019c7a9b-2222-7000-8000-000000000004.jsonl · Codex 0.112.0 · mode Codex Desktop, approval on-request, provider openai, sandbox workspace-write · 5 tool calls (5 shell) · 0 file writes · 5 findings Task: try a cheaper model for the summariser and compare output on 50 rows 2026-09-23 14:02 network_egress Bash gpt-5.3-codex `cd ~/pipeline && pip install openai` 2026-09-23 14:04 network_egress Bash gpt-5.3-codex `curl -s https://api.deepseek.com/v1/models -H "Authorization: ** \| head -c 300` api.deepseek.com 2026-09-23 14:04 credential_text Bash gpt-5.3-codex `curl -s https://api.deepseek.com/v1/models -H "Authorization: ** \| head -c 300` with network 2026-09-23 14:06 destructive_command Bash gpt-5.3-codex `rm -rf ~/pipeline/.cache && pytest -q` 2026-09-23 14:30 user_interrupted human gpt-5.3-codex `turn aborted: interrupted`
CRITICAL new	2026-09-24	Cursor	grok-4.6	etsy-shop 873f1cd2 update my etsy shop landing page with the autumn range and push it to netlify	60	2	1	1						3
Model: grok-4.6 (2 turns, 1 cmds) /srv/agent-logs/carol/.cursor/projects/etsy-shop/agent-transcripts/873f1cd2-3333-4000-8000-000000000006/873f1cd2-3333-4000-8000-000000000006.jsonl · Cursor ? · mode ? · 3 tool calls (1 shell, 2 edit) · 2 file writes · 4 findings Task: update my etsy shop landing page with the autumn range and push it to netlify 2026-09-24 10:05 outside_project_path Write grok-4.6 `/Users/carol/Documents/etsy-shop/index.html` /Users/carol/Documents/etsy-shop/index.html 2026-09-24 10:05 outside_project_path Bash grok-4.6 `cd ~/Documents/etsy-shop && npx netlify deploy --prod` ~/Documents 2026-09-24 11:05 destructive_command Delete grok-4.6 `Delete /Users/carol/Documents/etsy-shop/summer.html` /Users/carol/Documents/etsy-shop/summer.html 2026-09-24 11:05 outside_project_path Delete grok-4.6 `/Users/carol/Documents/etsy-shop/summer.html` /Users/carol/Documents/etsy-shop/summer.html
CRITICAL new	2026-09-25	Copilot CLI	gpt-5	infra 8b2c78c4 install the terraform version the repo pins	40	1	3			1	1			1		1
Model: gpt-5 (1 turns, 3 cmds) /srv/agent-logs/dan/.copilot/session-state/8b2c78c4-4444-4000-8000-000000000008/events.jsonl · Copilot CLI 1.0.88 · mode branch main · 3 tool calls (3 shell) · 0 file writes · 4 findings Task: install the terraform version the repo pins 2026-09-25 16:03 network_egress Bash gpt-5 `curl -fsSL https://releases.hashicorp.com/terraform/1.9.5/terraform_1.9.5_darwin_arm64.zip -o /tmp/tf.zip` releases.hashicorp.com 2026-09-25 16:04 privilege Bash gpt-5 `sudo unzip -o /tmp/tf.zip -d /usr/local/bin` 2026-09-25 16:04 outside_project_path Bash gpt-5 `sudo unzip -o /tmp/tf.zip -d /usr/local/bin` /usr/local 2026-09-25 16:40 user_interrupted human gpt-5 `abort: user`
HIGH	2026-09-16	Codex	gpt-5.3-codex	pipeline rollout- profile the ingest step and cut its run time	5	1	4		1			1			1
Model: gpt-5.3-codex (1 turns, 4 cmds) /srv/agent-logs/bob/.codex/sessions/2026/09/16/rollout-2026-09-16T14-00-00-019c7a9b-2222-7000-8000-000000000003.jsonl · Codex 0.112.0 · mode Codex Desktop, approval on-request, provider openai, sandbox workspace-write · 4 tool calls (4 shell) · 0 file writes · 3 findings Task: profile the ingest step and cut its run time 2026-09-16 14:03 dynamic_exec Bash gpt-5.3-codex `python -c "import pstats; pstats.Stats('prof.out').sort_stats('cumtime').print_stats(15)"` 2026-09-16 14:05 publish_action Bash gpt-5.3-codex `git push origin main` 2026-09-16 14:05 user_refused exec_command gpt-5.3-codex `git push origin main`
HIGH	2026-09-18	Copilot CLI	gpt-5	infra 8b2c78c4 rebase my branch on main and push	4	1	2	1	1	1					1
Model: gpt-5 (1 turns, 2 cmds) /srv/agent-logs/dan/.copilot/session-state/8b2c78c4-4444-4000-8000-000000000007/events.jsonl · Copilot CLI 1.0.88 · mode branch main · 2 tool calls (2 shell) · 0 file writes · 4 findings Task: rebase my branch on main and push 2026-09-18 16:03 network_egress Bash gpt-5 `cd ~/infra && git fetch && git rebase origin/main` 2026-09-18 16:04 destructive_command Bash gpt-5 `cd ~/infra && git push --force origin feature/vpc` 2026-09-18 16:04 publish_action Bash gpt-5 `cd ~/infra && git push --force origin feature/vpc` 2026-09-18 16:04 user_refused bash gpt-5 `cd ~/infra && git push --force origin feature/vpc`
HIGH	2026-09-19	Gemini CLI	gemini-3.8-flash	analytics session- regenerate the weekly report from the new parquet files	30	1	2	1							1
Model: gemini-3.8-flash (1 turns, 2 cmds) /srv/agent-logs/eve/.gemini/tmp/analytics/chats/session-2026-09-19T11-00-cb937252.jsonl · Gemini CLI ? · mode kind main · 3 tool calls (2 shell, 1 edit) · 1 file writes · 2 findings Task: regenerate the weekly report from the new parquet files 2026-09-19 11:05 destructive_command Bash gemini-3.8-flash `rm -rf ~/analytics/reports/old` 2026-09-19 11:05 user_refused run_shell_command gemini-3.8-flash `rm -rf ~/analytics/reports/old`
HIGH new	2026-09-22	Claude Code	claude-opus-5-5	payments 1a2b3c4d the nightly job is failing on the staging box, find out why	7	2	4	1		4				1	1
Model: claude-opus-5-5 (5 turns, 4 cmds) /srv/agent-logs/alice/.claude/projects/-Users-alice-payments/1a2b3c4d-1111-4000-8000-000000000002.jsonl · Claude Code 2.1.197 · mode acceptEdits · 4 tool calls (4 shell) · 0 file writes · 7 findings Task: the nightly job is failing on the staging box, find out why 2026-09-22 09:02 network_egress Bash claude-opus-5-5 `ssh [email protected] 'journalctl -u nightly --since yesterday \| tail -50'` staging.internal.example 2026-09-22 09:03 network_egress Bash claude-opus-5-5 `ssh [email protected] 'df -h /var'` staging.internal.example 2026-09-22 09:05 destructive_command Bash claude-opus-5-5 `ssh [email protected] 'sudo rm -rf /var/log/nightly/'` 2026-09-22 09:05 network_egress Bash claude-opus-5-5 `ssh [email protected] 'sudo rm -rf /var/log/nightly/'` staging.internal.example 2026-09-22 09:05 user_refused Bash claude-opus-5-5 `ssh [email protected] 'sudo rm -rf /var/log/nightly/*'` 2026-09-22 09:07 network_egress Bash claude-opus-5-5 `ssh [email protected] 'sudo logrotate -f /etc/logrotate.d/nightly'` staging.internal.example 2026-09-22 09:07 outside_project_path Bash claude-opus-5-5 `ssh [email protected] 'sudo logrotate -f /etc/logrotate.d/nightly'` /etc
HIGH new	2026-09-26	Claude Code	claude-sonnet-5	mobile-app 1a2b3c4d prepare the 3.2 release notes and tag it	4	1	2		1
Model: claude-sonnet-5 (3 turns, 2 cmds) /srv/agent-logs/frank/.claude/projects/-Users-frank-mobile-app/1a2b3c4d-6666-4000-8000-00000000000c.jsonl · Claude Code 2.1.197 · mode acceptEdits · 3 tool calls (2 shell, 1 edit) · 1 file writes · 1 findings Task: prepare the 3.2 release notes and tag it 2026-09-26 09:04 publish_action Bash claude-sonnet-5 `cd ~/mobile-app && git add CHANGELOG.md && git commit -m 'v3.2 notes' && git tag v3.2 && git push origin main v3.2`
NORMAL	2026-09-17	Cursor	grok-4.6	storefront 873f1cd2 make the checkout button match the new brand colour and run the snapshot tests	0	1	1
Model: grok-4.6 (1 turns, 1 cmds) /srv/agent-logs/carol/.cursor/projects/storefront/agent-transcripts/873f1cd2-3333-4000-8000-000000000005/873f1cd2-3333-4000-8000-000000000005.jsonl · Cursor ? · mode ? · 3 tool calls (1 shell, 1 read, 1 edit) · 1 file writes · 0 findings Task: make the checkout button match the new brand colour and run the snapshot tests No findings.
NORMAL	2026-09-20	Claude Code	claude-sonnet-5	mobile-app 1a2b3c4d fix the crash on launch when the user has no network	9	2	3									1	1
Model: claude-sonnet-5 (6 turns, 3 cmds) /srv/agent-logs/frank/.claude/projects/-Users-frank-mobile-app/1a2b3c4d-6666-4000-8000-00000000000b.jsonl · Claude Code 2.1.197 · mode acceptEdits · 6 tool calls (3 shell, 2 read, 1 edit) · 1 file writes · 2 findings Task: fix the crash on launch when the user has no network 2026-09-20 09:00 repeated_command Bash claude-sonnet-5 `cd ~/mobile-app && xcodebuild test -scheme App -quiet` 3× 2026-09-20 09:07 user_interrupted human claude-sonnet-5 `[Request interrupted by user]`
NORMAL new	2026-09-26	Gemini CLI	gemini-3.8-flash	analytics session- the dashboard query is slow, add an index	30	1	1
Model: gemini-3.8-flash (1 turns, 1 cmds) /srv/agent-logs/eve/.gemini/tmp/analytics/chats/session-2026-09-26T11-00-cb937252.jsonl · Gemini CLI ? · mode kind main · 2 tool calls (1 shell, 1 read) · 0 file writes · 0 findings Task: the dashboard query is slow, add an index No findings.

Models — which model was in charge, and what it did

The model named on each turn in the log (Claude Code: the assistant record's model; Codex: the turn's model). A session that changed model mid-way is counted under each. Turns = model replies; Sub-agent = of those turns, how many were a sub-agent the tool spawned rather than the main thread (a model that is ALL sub-agent turns was chosen by the tool, not the person); Cmds = shell commands the model wrote and ran; Reads / Edits / Web / Browser = its other tool calls by kind; the rest are that model's signal counts.

Model	Agent	Sessions	Turns	Sub-agent	Cmds	Reads	Edits	Web	Destr.	Publ.	Net	Priv.	Cred.	Refused	Stopped
`claude-opus-5-5`	Claude Code	2	10		7	1	1		1	1	5		1	1
`claude-sonnet-5`	Claude Code	2	9		5	2	2			1					1
`claude-haiku-4-5-20251001`	Claude Code	1	3	3	1			1
`grok-4.6`	Cursor	2	3		2	1	3		1
`gpt-5.3-codex`	Codex	2	2		9				1	1	2		1	1	1
`gpt-5`	Copilot CLI	2	2		5				1	1	2	1		1	1
`gemini-3.8-flash`	Gemini CLI	2	2		3	1	1		1					1

Sub-agents — what the tool spawned, on which model, what each was asked and what it reported

One row per sub-agent transcript. Brief = the first message the main model gave it; Reported back = its last text reply. Both are the log's own words, cut to one line and masked. The calls column is that sub-agent's tool use by kind and the hosts it reached.

When	Session	Model	Brief	Calls	Reported back
2026-09-15 10:30	1a2b3c4d payments	claude-haiku-4-5-20251001	Find where settlement timeouts are configured across the repo and report the values.	1 shell, 1 web · hosts: docs.example-bank.com	Timeouts are set in settle/config.py (30 s) and overridden in tests/conftest.py (5 s).

Hosts contacted — a URL or ssh/scp/rsync target in the same shell segment as a network command, or a web fetch / browser navigation

Country is shown only where the Agent Scan's provider table states it (AI model endpoints); other hosts show —. Loopback and .local hosts are not listed.

Host	Country	Requests	Sessions
`staging.internal.example`	—	4	1
`api.example-bank.com`	—	1	1
`docs.example-bank.com`	—	1	1
`api.deepseek.com`	CN sovereignty	1	1
`releases.hashicorp.com`	—	1	1

Hosts named but not fetched — a URL written into a file, printed, or in a test fixture; never beside a network command

Listed so the distinction is visible: a provider URL here means the agent wrote code or text that mentions it, not that anything was sent there. Hosts that were also contacted appear in the table above only.

Host	Country	Mentions	Sessions
None.

Refused by the human — 4 tool calls turned down at the permission prompt

2026-09-16 14:05 exec_command git push origin main

2026-09-18 16:04 bash cd ~/infra && git push --force origin feature/vpc

2026-09-19 11:05 run_shell_command rm -rf ~/analytics/reports/old

2026-09-22 09:05 Bash ssh [email protected] 'sudo rm -rf /var/log/nightly/*'

A signal is a pattern match on what the agent ran, not a judgement that harm occurred: a force push to your own branch and a force push to main both count as destructive. Read the calls behind a number before acting on it. This version reads Claude Code and Codex session logs only.