Direct answer: hiding content previews does not hide the run
The default JSON exporter replaces prompt, command and LLM-output previews with the string redacted. But its session summaries still include identifiers, hashes, models, timestamps, categorized file and network effects, and other structured evidence. The pprof exporter does not use that JSON preview switch at all: its samples carry semantic stacks plus a separate list of evidence labels. A folded stack or SVG presents fewer fields than the JSON session records, but its frames remain readable.
This is an inspection of the released v1.0.35 source at c1574366122a90c451bca6a84713e28813ab95e2, not a demonstration that a specific private token escaped or a claim that every session contains all these fields. For the general capture/export boundary, see AgentSight security and local data handling; for deciding what the flamegraph widths mean, see the Agent Flamegraph guide.
What each output format actually retains
| Export format or option | Evidence retained in v1.0.35 | Sharing decision |
|---|---|---|
| JSON, default (no --include-previews) | Prompt, command, and LLM-output preview fields become the literal text "redacted". The export still carries session IDs, file basename, timestamps, models, tags, hashes, process-chain information, path groups and domains when available. | Preview suppression is not anonymization. Correlation keys and metadata may identify a run or organization. |
| JSON with --include-previews | The prompt and LLM preview values and the tool command text are serialized instead of the redacted placeholders. | Do not use this flag for public artifacts unless the inputs have already been sanitized and the output reviewed. |
| Pprof protobuf (.pb or .pb.gz) | The stack functions plus sample labels: project, agent, session, source/evidence IDs, a normalized but unhashed call ID, prompt/response hashes, timestamp, skill and status when present. Pprof stores these strings in its own string table. | Keeping project, agent and session outside the default stack does not remove them: inspect sample labels separately. |
| Folded stacks (.folded) | One line per aggregated semantic frame sequence with a numeric width. Frame text survives as readable stack content. | No pprof-specific per-sample label set here, but semantic labels and object names can still reveal task or repository structure. |
| SVG flamegraph (.svg) | A visual projection of the same folded stack labels and their weights; labels and tooltip text are part of the generated SVG. | An image-looking artifact can still disclose readable text. Preview before embedding it on a public page. |
| Any format with customized --stack / --stack-rule / --op-map | Operator-selected fields and mapped labels become stack frames when their source values are present. | Custom mapping changes the disclosure surface. Do not assume the default stack and its sanitization properties still describe the output. |
These are source-derived format contracts, not measured occurrence frequencies. Actual field presence depends on the input agent transcript, parsed events, view, filters and tagging options. A profile containing zero samples will not disclose all of the fields listed above; a custom stack can disclose others.
The default JSON redaction is deliberately narrow
The implementation serializes session_to_json(session, include_previews)alongside each JSON profile. Without the flag, it substitutes the literal redacted for prompt previews, command previews, and LLM response previews. It does not replace the surrounding records with aggregate totals. For example, tool_events can retain process_chain, path_groups, domains, tool classifications and a command hash; llm_events retain model names and token counts.
A digest such as cwd_hash or command_hash is not encryption. It can support matching the same value across artifacts and, for a small candidate dictionary, sometimes guesses about the original value. Conversely, the presence of a hash does not establish that the original plaintext is exported. Judge each field by the concrete sharing threat model.
| Surviving field family | What the source guarantees | Question before publication |
|---|---|---|
session_id / agent_sight_session_id | Default JSON contains source-native and constructed session identity, not just a hashed anonymous label. | Remove or replace stable run identifiers before external sharing when they could be linked to other logs. |
session_file / cwd_hash / source | A source-file basename, working-directory hash, and agent family can survive preview redaction. | A basename may encode an identifier; a hash is a pseudonym and still allows equality matching. |
tool_events, path_groups, domains, process_chain | These structured metadata fields are exported independently of command_preview. | A host, tool, or path category may disclose environment details without showing the underlying prompt. |
model, usage counts, timestamps, tags | LLM model and token usage, operation times, and semantic task labels remain useful for profiling. | Their combination can reveal workflows and activity patterns even if content previews are absent. |
pprof evidence labels | The profile builder passes its evidence-label field set through safe_frame and serializes key/value strings into pprof samples. | Run pprof tag inspection; ordinary flamegraph screenshots do not show every label retained by the binary. |
safe_frame and private-marker handling | Source normalizes characters and some path/username/private-marker cases. It does not implement a universal sensitive-string redactor. | Test your actual labels and custom rules; never substitute normalization for data classification. |
Pprof sample labels are not the same as flamegraph frames
The default operation stack is task,skill,phase,action,object,repeat,result,outcome, with a further token frame in the token view. Project, agent and session are not default stack frames, but the pprof builder separately serializes a set of evidence labels. The source-defined list includes project, agent, session, source_kind, evidence_id, status, source_session, call_id (normalized, not hashed), prompt/response hashes, timestamp, and skill, where available. Thus a screenshot showing only broad task rectangles is not a complete inventory of what the corresponding .pb.gz file contains.
The builder passes label values through safe_frame and puts them into the pprof string table. That function lowercases and normalizes punctuation, and performs some path or private-marker substitutions. The source does not promise to detect arbitrary API keys, repository names or confidential task labels. The separate AgentSight JSON snapshot contract is a different exporter and must not be used as evidence for agentpprof field behavior.
How to audit your own profile before sending it
Use a synthetic or otherwise authorized session first. These commands are inspection steps, not a measured sample run; they assume the released agentpprof executable is available and the selected local project has readable supported sessions. Do not paste unredacted outputs into an issue or public chat.
agentpprof --project-root . --view operations -o review.json
# Examine which keys remain in default no-preview JSON:
jq '{root_keys: keys, session_keys: ((.sessions[0] // {}) | keys)}' review.json
agentpprof --project-root . --view operations -o review.pb.gz
# pprof labels can disclose fields outside the displayed stack:
go tool pprof -tags review.pb.gz
agentpprof --project-root . --view operations -o review.folded
# Examine the human-readable semantic stack labels:
head -n 10 review.foldedBefore considering --include-previews, first check the default output for unique identifiers, recognizable paths, external destinations, semantic tags, model identifiers, timestamps and retained hash values. If you need to publish an example, use a deliberately synthetic transcript or a separately reviewed, minimal aggregate. Merely renaming the file or converting a binary pprof into SVG does not define a disclosure policy.
A six-question disclosure review
- Does the recipient need raw per-session evidence, or would a category-level total answer the question?
- Is the output JSON, pprof, folded, or SVG, and have you inspected that exact artifact?
- Did
--include-previews,--stack,--stack-rule, or--op-mapalter default disclosure? - Can a model name, domain, project path, source file basename, timestamp or session ID identify the task?
- Do label normalization and hashes only pseudonymize values that can be linked elsewhere?
- Does the recipient have explicit authorization to receive the selected telemetry, including its retention and access conditions?
A negative answer to the first question often means producing a smaller artifact, not increasing the number of redaction rules. A useful shared finding can retain category labels and relative weights while removing session-level identity and sensitive effects. That transformed artifact should itself be reviewed.
Limits of this source audit
This analysis checks the release implementation and documented formats, not every optional tagger, future output format, plugin, local transcript parser, or deployment. It does not prove a particular secret appears in a generated trace. Output may vary when custom rules, source-agent versions, or content are changed. The time view's widths represent event-derived elapsed intervals, not CPU samples, while tokens can use estimates when logs lack native token counts; neither should be silently relabeled as billed cost or CPU utilization. Local profiling is also distinct from exporting model-call telemetry via OpenTelemetry.
To reproduce the source audit against this release, inspect the PPROF_EVIDENCE_LABEL_FIELDS, session_to_json, safe_frame, write_pprof_samples, and write_projection definitions in the pinned file. Compare them with the CLI flag in main.rs and the privacy warning in the agentpprof README.
Primary sources
All format claims are anchored to the AgentSight v1.0.35 product source released 23 October 2026. This article was researched on 7 November 2026; it contains no invented user trace, real-world breach example, benchmark, or profile measurement.
- v1.0.35 profile serialization, output formats, labels, and safe_frame
- v1.0.35 agentpprof CLI options, including --include-previews
- v1.0.35 documented output and privacy caveats
- v1.0.35 session record model and discovery
- AgentSight agentpprof semantic profiling reference
- AgentSight v1.0.35 release
Next: set the telemetry sharing boundary before circulating a flamegraph, and use the Agent Flamegraph guide to interpret its axes.