Phoenix CLI

Open Coding

Free-form note-writing against sampled traces, spans, or sessions, before any taxonomy exists. After you pick a sample at the right unit (see Choosing the unit of analysis), read each one and write a short, specific observation of what went wrong. These raw notes feed axial coding, where they get grouped into named failure categories — and ultimately into eval targets or fix priorities.

Reach for this whenever the user wants to look at LLM traffic without a fixed taxonomy yet — e.g., "what's going wrong with this agent", "I just instrumented my app, where do I start", "review these traces", "the chatbot keeps losing context", "what kinds of mistakes is the model making", "help me make sense of these conversations", or any framing that needs grounded observations before categories.

Choosing the unit of analysis

The right unit — trace, span, or session — depends on the question and the system. Pick deliberately before recording; the choice determines whether you call px trace, px span, or px session throughout, and a wrong default is expensive to undo mid-run.

The unit is about where the failure modes you're investigating actually live:

• Trace — one input → one call graph → one output. Right for classifiers, single-shot summarizers, stateless tool-using agents, single-query RAG. Failure modes that live here: wrong answer, malformed output, missed retrieval, bad tool selection within one request.

• Span — one operation inside a trace. Right for in-isolation mechanical failures (an exception fired, a tool returned an error response, an output is malformed) or when you can attribute on sight to a specific component. Reach for span when the trace as a whole is fine but one piece inside it is the unit of interest.

• Session — a sequence of traces sharing a session.id. Right for multi-turn conversational agents, agents with episodic memory, anything where the failure mode is a trajectory: context loss across turns, drift from the user's stated goal, the agent forgetting a stated preference, repeated user clarifications. These failures don't exist on any single trace; they only exist across traces.

Diagnostic — three signals to read

1. User framing. Tilts session: "conversation", "agent forgot", "drift", "memory", "across turns", "user had to repeat themselves". Tilts trace: "this trace", "this call", "the response was wrong", "wrong output". Tilts span: "exception", "error response", "malformed", "the retrieval failed".

2. Data shape. Probe before the loop. The session id lives at rootSpan.attributes["session.id"] (it is not a top-level field on the trace JSON), and is "" for traces that aren't session-wired — filter both:

   Copy

   px trace list --limit 200 --format raw --no-progress \
     | jq '
       [ .[] | .rootSpan.attributes["session.id"] // empty | select(. != "") ]
       | { with_session: length,
           distinct_sessions: (group_by(.) | length),
           median_traces_per_session:
             (group_by(.) | map(length) | sort | .[length/2|floor] // 0) }
     '

   with_session: 0 → sessions not wired; trace is the grain. median_traces_per_session: 1 → single-trace sessions; still trace. median_traces_per_session: 5+ → sessions are meaningful; session is plausibly right.

3. System type. Open one recent trace and inspect the root span's input. A single user message → one turn or one shot. A message array ([{role: user}, {role: assistant}, ...]) → that's a turn within a longer dialogue; the dialogue lives at the session level.

   Copy

   px trace get <trace-id> --format raw \
     | jq '.rootSpan.attributes["input.value"] | (try fromjson catch .) | (type, length?)'

Commit out loud, then proceed

State the unit explicitly before recording any note:

"Question: 'the chatbot keeps losing context'. Data: median 7 traces per session, message-array inputs. Recording at the session level; will drop to trace for single-turn observations, span for mechanical failures."

The unit can shift if data demands it — a trace-level investigation that surfaces "the agent never remembers earlier turns" should pivot to session. Record the observation, then refocus the next batch. The unit is a starting hypothesis, not a contract.

Coding annotation identifier (pick this first)

Every artifact this workflow produces — open-coding notes, axial-coding labels, the local sidecar files, and the UI-filter annotation — is tagged with one coding annotation identifier so the run is queryable, revertible, and viewable as a unit. Pick a descriptive, unique identifier before recording any notes. Format suggestion:

coding-run:<short-topic>-<YYYY-MM-DD>

Examples: coding-run:chatbot-context-loss-2026-05-06, coding-run:agent-tool-misuse-q2. Descriptive ids carry meaning for whoever opens the data later — better than an opaque uuid. The coding-run: prefix is a visual convention; the value is the workflow's coding annotation identifier, not a px session id.

Workflow term vs. server annotation name. The skill calls this value the coding annotation identifier. The server-side annotation NAME used for the UI filter is unchanged — coding_session_id — for data compatibility with rows already written. Don't try to rename it.

Pass the identifier explicitly on every px call. A shell variable for readability is fine, but do not rely on shell inheritance — many agent harnesses spawn each command in a fresh subshell, so CODING_ANNOTATION_IDENTIFIER may not propagate.

CODING_ANNOTATION_IDENTIFIER="coding-run:chatbot-context-loss-2026-05-06"

The local sidecar lives at .px/coding/<sanitized-identifier>.jsonl (CWD-relative, matching the .px/docs precedent). Sanitization rule: replace any character not matching [a-zA-Z0-9_-] with - before using the value in the filename — colons, slashes, and other shell-fragile characters get normalized. For CODING_ANNOTATION_IDENTIFIER="coding-run:chatbot-context-loss-2026-05-06" the sidecar path is .px/coding/coding-run-chatbot-context-loss-2026-05-06.jsonl.

Verify this run hasn't already started — uniqueness is a local file check, not a server query:

SLUG=$(echo -n "$CODING_ANNOTATION_IDENTIFIER" | sed 's/[^a-zA-Z0-9_-]/-/g')
SIDECAR=".px/coding/${SLUG}.jsonl"
test ! -f "$SIDECAR" || { echo "Sidecar already exists at $SIDECAR — pick a new identifier or delete the file"; exit 1; }
mkdir -p .px/coding

If $SIDECAR already exists, append a disambiguator (-v2, -dustin, etc.) to CODING_ANNOTATION_IDENTIFIER, re-derive SLUG, and re-check. The agent harness can run open coding and axial coding in independent invocations: each step re-derives SLUG from CODING_ANNOTATION_IDENTIFIER and reads/writes the same file.

Process

1. Pick a coding annotation identifier — choose a descriptive value and verify the sidecar file does not yet exist (see Coding annotation identifier)

2. Pick the unit — work through Choosing the unit of analysis and commit to trace, span, or session

3. Inspect — fetch one entity at the chosen unit (trace / span / session)

4. Read — input, output, exceptions, tool calls, retrieved context, and (at session level) the trajectory across child traces

5. Note — write one specific sentence describing what went wrong (or skip if correct)

6. Record — px {trace,span,session} add-note <id> --text "..." --identifier "$CODING_ANNOTATION_IDENTIFIER" --format raw --no-progress, add/update one JSONL sidecar row for the note, then write the matching UI-filter annotation

7. Iterate — move to the next entity; repeat until the sample is exhausted or saturation hits

8. Hand off — axial coding reads the sidecar directly (no shared shell required); see Wrapping up for the UI link

Inspection

Use px to read context at the unit committed in Choosing the unit:

• Trace unit — read one trace's input → tool calls → retrieved context → output as one story.

• Span unit — read one operation's input/output and surrounding spans for context.

• Session unit — read the sequence of traces in order; the trajectory (turns, retrievals, tool-call patterns across traces) is the data, not any single trace's inputs and outputs.

Don't filter the sample by --status-code ERROR. OTel's status_code only flips to ERROR when an instrumentor catches a raised Python exception (network failure, 5xx, parse error). Hallucinations, wrong tone, retrieval misses, and bad tool selection all complete cleanly and arrive as OK or UNSET. Sampling for open coding by --status-code ERROR excludes the population this workflow exists to surface.

# Sample recent traces — the unit of inspection in open coding
px trace list --limit 100 --format raw --no-progress | jq '
  .[] | {trace_id: .traceId, root: .rootSpan.name, status,
         input: .rootSpan.attributes["input.value"],
         output: .rootSpan.attributes["output.value"]}
'

# Trace-level context — all spans in one trace, ordered by start_time
px trace get <trace-id> --format raw | jq '
  .spans | sort_by(.start_time) | map({span_id: .context.span_id, name, status_code,
    input: .attributes["input.value"],
    output: .attributes["output.value"]})
'

# Drill to one span (px span get does not exist; filter via span list)
px span list --trace-id <trace-id> --format raw --no-progress \
  | jq '.[] | select(.context.span_id == "<span-id>")'

# Check existing notes on traces (default) or spans you are about to review
# Notes are stored as annotations with name="note"; use --include-notes (not --include-annotations)
px trace list --include-notes --limit 10 --format raw --no-progress | jq '
  .[] | select((.notes // []) | length > 0)
  | {trace_id: .traceId, notes: [.notes[] | .result.explanation]}
'
# Same shape on spans — swap px trace for px span and use .context.span_id

Always pipe through jq with --format raw --no-progress when scripting.

Recording Notes

Use the add-note command matching the unit committed in Choosing the unit: px trace add-note, px span add-note, or px session add-note. Every call carries an explicit --identifier "$CODING_ANNOTATION_IDENTIFIER" and --format raw --no-progress.

Passing --identifier "$CODING_ANNOTATION_IDENTIFIER" does two things:

• Tags the note row with the coding annotation identifier on the server, so the cleanup px <entity>-annotations delete --identifier "$CODING_ANNOTATION_IDENTIFIER" --all sweep removes every artifact this run produced.

• Makes the call upsert on (entity_id, name='note', identifier) — re-running open coding on the same entity within the same coding annotation identifier overwrites the prior note instead of appending a second row. (Without --identifier, the server stamps a unique px-{kind}-note:<uuid> and each call appends.)

After every successful add-note, record one JSONL line in $SIDECAR. The sidecar is what axial coding reads — no server round-trip. It is a content handoff, not code: keep it readable, inspect it directly, and use whatever simple tooling is convenient.

Sidecar JSONL line shape (one per add-note):

{"entity_kind":"trace","entity_id":"<trace-id>","note":"<text>","identifier":"<original identifier value, unsanitized>","ts":"<ISO-8601 UTC>"}

Fields:

• entity_kind — "trace", "span", or "session" (matches the add-note subcommand used)

• entity_id — the entity argument passed to add-note (trace id, span id, or session id)

• note — the --text value, verbatim

• identifier — the original $CODING_ANNOTATION_IDENTIFIER value, unsanitized; the sanitized form lives only in the filename

• ts — ISO-8601 UTC timestamp (e.g. 2026-05-08T17:14:09Z) of the local append

If you revise a note for the same entity under the same coding annotation identifier, either replace that row or append a newer row. When duplicate (entity_kind, entity_id) rows exist, the newest ts is the current note. This matches the server upsert behavior of add-note --identifier.

Minimal trace example:

px trace add-note <trace-id> \
  --text "Asked about returns; final answer covered shipping policy instead" \
  --identifier "$CODING_ANNOTATION_IDENTIFIER" \
  --format raw --no-progress

Then add a matching JSONL row to $SIDECAR using the line shape above. For span or session notes, change entity_kind, entity_id, and the px subcommand accordingly.

Bulk auto-tagging by status code (e.g. px span list --status-code ERROR | xargs ... add-note "error") is not open coding — open coding is manual, observation-grounded, and ranges over all failure modes, not just spans where Python raised. Skip the bulk-by-status-code shortcut; it produces fewer, less informative notes than walking traces.

UI-filter annotation

Every entity that receives an open-coding note (or an axial-coding label later) also needs a UI-filter annotation so the Phoenix UI can filter by coding annotation identifier. Phoenix's UI filter language is name-based, not identifier-based — there is no UI primitive for filtering by identifier, so an annotation whose name is the constant coding_session_id and whose label is the coding annotation identifier value is what the wrap-up UI link actually filters on.

The annotation NAME coding_session_id is the load-bearing data key on the server and is unchanged in this rewrite. The skill's workflow term is "coding annotation identifier"; the server key stays coding_session_id for compatibility with rows already written.

Run this once per touched entity, alongside the add-note (and again later when axial coding labels a different entity):

px trace annotate <trace-id> \
  --name coding_session_id \
  --label "$CODING_ANNOTATION_IDENTIFIER" \
  --identifier "$CODING_ANNOTATION_IDENTIFIER"
# or px span annotate / px session annotate at matching levels

The annotation's --identifier matches $CODING_ANNOTATION_IDENTIFIER, so the wrap-up DELETE cleans it up in the same call as the notes and the axial-coding labels.

Fallback write paths (one-line asides):

• POST /v1/trace_notes and POST /v1/span_notes and POST /v1/session_notes — accept one {data: {trace_id|span_id|session_id, note, identifier}} per request; the optional identifier field upserts on (entity_id, name='note', identifier) when non-empty.

• @arizeai/phoenix-client addTraceNote, addSpanNote, and addSessionNote wrap the same endpoints and accept an optional identifier field on the note object.

• The GraphQL endpoint rejects mutations with "Only queries are permitted." — write through px {trace,span,session} add-note or the REST endpoints above.

What Makes a Good Note

Write what you saw, not the category you think it belongs to — categorization happens in axial coding. Short prefixes like TONE: or FACTUAL: are a personal shorthand, not a repo convention.

Saturation

Stop writing notes when observations stop being new. Signals:

• Repeats — the last 10–15 traces produced notes that describe failures you've already seen.

• Paraphrase convergence — you catch yourself writing minor variations of earlier notes.

• Skips outnumber notes — most recent traces are correct and need no note.

At saturation, move on to axial coding to group what you have. Continuing past saturation adds traces but not insight. You do not need to annotate every trace — annotating correct ones dilutes signal.

Listing what this run produced

The local sidecar is the handoff record for notes written this run. Inspect it directly. Each line is one note record; if the same entity appears more than once, use the newest ts as the current note. Missing-file behavior: an absent sidecar means open coding has not yet started for this coding annotation identifier; treat that as zero notes, not an error. Malformed lines are line-local: fix or drop the bad line without editing neighbors.

Wrapping up

When the run is done, share the Phoenix UI link with the user. The link filters the project's traces page by the coding_session_id annotation written alongside each note. The UI route /projects/:projectId expects an encoded GraphQL node ID, not a project name — resolve it via px project get:

project_id=$(px project get "$PHOENIX_PROJECT" --format raw --no-progress | jq -r '.id')
encoded=$(python3 -c 'import urllib.parse, sys; print(urllib.parse.quote(sys.argv[1]))' \
  "annotations['coding_session_id'].label == '$CODING_ANNOTATION_IDENTIFIER'")
echo "Phoenix UI: $PHOENIX_HOST/projects/$project_id/traces?filterCondition=$encoded"

If the user wants to discard everything this run produced, three identifier-bound deletes handle the server side and one rm handles the local sidecars. Confirm with the user before running — this is destructive. Each call requires --all (or both --start-time and --end-time) to authorize the sweep; --identifier filters further but never authorizes on its own. Set PHOENIX_CLI_DANGEROUSLY_ENABLE_DELETES=true first if not already exported:

for kind in trace span session; do
  px "$kind-annotations" delete \
    --identifier "$CODING_ANNOTATION_IDENTIFIER" \
    --all -y \
    --format raw --no-progress
done
rm -f "$SIDECAR" ".px/coding/${SLUG}-axial.jsonl"

Each px <entity>-annotations delete call covers notes, structured annotations, and the coding_session_id annotation in one shot because they share the underlying annotation table.

Principles

• One coding annotation identifier per run — every server artifact and every sidecar line carries the same $CODING_ANNOTATION_IDENTIFIER; never mint a per-stage id.

• Pass --identifier explicitly — every px call gets --identifier "$CODING_ANNOTATION_IDENTIFIER"; do not rely on inherited env vars across harness-spawned subshells.

• Sidecar is the handoff record for notes — axial coding reads from the local sidecar, not from the server; if an entity appears more than once, the newest ts wins.

• Free-form over structured — do not pre-commit to a taxonomy during open coding; categories emerge in axial coding.

• Specific over general — quote or paraphrase the observed failure; vague labels ("bad response") carry no signal.

• Context before labeling — inspect input, output, and retrieved context before writing any note.

• Iterate before categorizing — work through the full sample first; resist grouping while still collecting.

• Skip is valid — a correct span needs no note; annotating everything dilutes signal.

• Revert is opt-in — the wrap-up DELETE only runs after explicit user confirmation; the default path prints the UI link and stops.