Dev insights: rca, changes, blame, friction

Prismor records every tool call an agent makes. Four views turn that record into answers to the questions developers ask when an agent run goes sideways:

All four read the local store and the agent's own transcript. They don't change policy or screening, and everything they print goes through the same redaction as prismor query.

Where outcomes come from

When a Claude Code call fails, Claude fires PostToolUseFailure, not PostToolUse. Before 1.61.0 Prismor didn't subscribe to it, so a failed call was stored as an allowed call with no result. For Claude these views read the session transcript (transcript_path, on every hook payload), which records each result with is_error and marks hook denials. A call that a policy denied is not counted as a failure.

Agents often pipe test runs (pytest -q 2>&1 | tail -40), so the shell exits 0 even when tests failed. A call is also counted as failed when its output carries an unambiguous runner summary: pytest, jest, go, cargo, tsc, npm, command not found, or a Python traceback.

For agents whose hooks don't report outcomes, the views fall back to the events table: the calls are known, the failures aren't, and rca says so.

prismor rca [session]

What went wrong in one session. Without an id it uses the latest session.

• First failure, with the three calls before it.
• Retry loops: the same call failing 3 or more times, with the time span and tokens spent.
• Reverted edits: an edit that restores exactly what an earlier edit replaced.

$ prismor rca d0c441fc…
4 calls · 4 failed · 0 denied

Loop: failed 4× (09:38:08–09:38:16)
  Bash: python3 -c 'import yaml_extras'
    → ModuleNotFoundError: No module named 'yaml_extras'

Flags: --workspace, --json.

prismor session <id> --changes

What a session actually touched: each file it wrote as a red/green diff with +N −M, the commands it ran, and a drafted PR description built from the prompts, files and test runs. For Claude the diff comes from Claude's own structuredPatch, so line numbers match the editor.

prismor blame <path>[:<line>]

Which agent session wrote a file, or one line of it, and what that session was asked. The dashboard's Files page lists every file an agent wrote as a folder tree.

Agents often write files from the shell (cat > f <<'EOF', >>, tee). Those never become file-write events, so blame finds them by parsing the shell commands and marks them shell. The text they replaced was never recorded.

The first blame on a large store builds an index (about 11 s on a 2 GB store); later calls take a fraction of a second.

Flags: --workspace, --json.

prismor friction

Failures that recur across sessions, grouped by command (cd x && FOO=1 python3 -m pytest -q | tail counts as python3 -m pytest) and printed as lines you can paste into AGENTS.md, so the next agent doesn't have to rediscover the fix. A failure has to appear in at least 2 sessions.

$ prismor friction --all
Recurring failures, last 14 days (4 sessions). Paste into AGENTS.md:

- `python3 -m pytest` has failed in 3 sessions: ModuleNotFoundError: No module named 'invoice'

Flags: --days (default 14), --all (every workspace), --workspace, --json.

Live retry-loop warning (Claude Code)

install-hooks registers PostToolUseFailure for Claude Code. Only failed calls reach it, so successful calls cost nothing. On the third identical failure in a session, the agent is told:

Prismor: this exact Bash call has now failed 3 times in this session (ModuleNotFoundError: No module named 'yaml_extras'). Retrying it unchanged will fail again: read the error, check the assumption behind the call, and change approach.

It's advisory: nothing is blocked. Existing installs get it after re-running prismor setup or prismor install-hooks --agent claude.

Limits

• Non-Claude agents: hook payloads carry no result status, so rca reports their calls without failures.
• Blame has no commit linkage: transcripts carry the branch, not the commit.
• Shell writes are found in the most recent 3,000 redirecting commands.

See also: Dashboard · Query your data · CLI reference.