@narumitw/pi-langfuse 0.18.0 → 0.26.0

package/README.md

@@ -6,13 +6,14 @@
 
 ## ✨ Features
 
-- Names each Langfuse trace `pi.trace` and creates a native `pi.agent` observation for the Pi agent run.
-- Adds a `pi.turn` span for every Pi turn, with generations and tools nested beneath it.
-- Records finalized assistant outputs without retaining intermediate provider request payloads.
-- Records provider, model, stop reason, token usage, and known non-zero reported cost.
-- Records normalized tool inputs, finalized outputs, duration, and failures as child spans.
-- Groups traces with Pi's session id.
-- Records provider HTTP status codes when Pi exposes them.
+- Names each Langfuse trace `pi.trace` and uses a native `pi.agent` as its root observation.
+- Keeps the root agent open until settlement, with indexed `pi.attempt` spans for retries and queued continuations.
+- Records bounded provider-request snapshots, finalized assistant outputs, requested/response identity, TTFT, usage, and known cost buckets.
+- Retains ordered HTTP response history and safe diagnostic headers without marking recovered requests as errors.
+- Records final tool inputs and outputs, progress timing, duration, and failures without exporting partial-result content.
+- Records active compactions structurally without exporting generated summary text.
+- Groups traces with Pi's session id and adds bounded session/context snapshots and aggregate counters.
+- Adds the run-start Git branch and commit as metadata plus a filterable branch tag.
 - Reads Langfuse credentials and options only from a private `pi-langfuse.json` file.
 - Batches routine exports without delaying normal Pi agent completion.
 - Keeps its OpenTelemetry provider isolated so it coexists with other tracing extensions.
@@ -25,7 +26,13 @@
 pi install npm:@narumitw/pi-langfuse
 ```
 
-Try the local workspace package:
+Try the published package without installing it:
+
+```bash
+pi -e npm:@narumitw/pi-langfuse
+```
+
+Try a local checkout:
 
 ```bash
 pi -e ./extensions/pi-langfuse
@@ -35,13 +42,15 @@ The Langfuse v4 SDK requires Node.js 20 or newer.
 
 ## ⚙️ Configuration
 
-Create or update the private config interactively from Pi:
+Run the interactive manager, then choose **Set up Langfuse for this Pi agent directory** or **Update Langfuse for this Pi agent directory**:
 
 ```text
-/langfuse init
+/langfuse
 ```
 
-The command prompts for the secret key, public key, and base URL in the same order Langfuse presents them. Leave either key blank to preserve its existing value when updating a valid config. Leave the base URL blank to use `https://us.cloud.langfuse.com`. The file is saved atomically with mode `0600`; restart Pi after saving. In print or JSON mode, edit the file manually because interactive input is unavailable.
+The setup flow prompts for the secret key, public key, and base URL in the same order Langfuse presents them. Leave either key blank to preserve its existing value when updating a valid config. Leave the base URL blank to use `https://us.cloud.langfuse.com`. The file is saved atomically with mode `0600`.
+
+Configuration belongs to the displayed Pi agent directory, not just the current conversation. Restart each running Pi process after saving; the new connection applies to subsequent sessions in that process. `/reload` is not sufficient because the isolated Langfuse runtime is initialized once per process. In print or JSON mode, edit the file manually because the interactive manager is unavailable.
 
 You can also create the file manually:
 
@@ -58,7 +67,7 @@ You can also create the file manually:
 
 `publicKey` and `secretKey` are required literal strings. Environment-variable and command interpolation are intentionally unsupported. `baseUrl` defaults to `https://us.cloud.langfuse.com`; regional and self-hosted HTTP or HTTPS endpoints are supported. Prefer HTTPS because HTTP sends Langfuse credentials and trace content without transport encryption.
 
-`environment` and `release` are optional Langfuse trace attributes. Set `captureContent` to `false` to trace timing, model, usage, cost, and status metadata without sending prompts, responses, or tool content.
+`environment` and `release` are optional Langfuse trace attributes. An environment must match Langfuse's contract: at most 40 lowercase letters, numbers, hyphens, or underscores, and it cannot start with `langfuse`. Set `captureContent` to `false` to trace timing, model, usage, cost, status, and bounded diagnostic metadata without sending prompts, provider-request snapshots, responses, or tool content.
 
 The extension automatically restricts an existing config file to mode `0600` and refuses to load credentials if that protection cannot be enforced. You can also set it explicitly:
 
@@ -66,47 +75,93 @@ The extension automatically restricts an existing config file to mode `0600` and
 chmod 600 ~/.pi/agent/pi-langfuse.json
 ```
 
-Restart Pi after changing credentials, endpoint, environment, or release. The isolated OpenTelemetry tracer provider is initialized once per Pi process and selected only for Langfuse; it does not replace Pi's process-global provider or send Langfuse observations to another extension's exporter.
+Restart Pi after changing credentials, endpoint, environment, release, or `captureContent`. The isolated OpenTelemetry tracer provider is initialized once per Pi process and selected only for Langfuse; it does not replace Pi's process-global provider or send Langfuse observations to another extension's exporter.
 
 ## 🔭 What is traced
 
-Each `pi.trace` contains one `pi.agent` native `agent` observation, which contains:
+Each trace has this observation hierarchy:
+
+```text
+pi.trace
+└── pi.agent (agent: submitted prompt until Pi fully settles)
+    ├── pi.attempt (span: one agent_start/agent_end pair)
+    │   └── pi.turn (span)
+    │       ├── pi.llm (generation)
+    │       └── pi.tool.<tool-name> (tool)
+    ├── pi.compaction (span, only while the trace is active)
+    └── pi.attempt ...
+```
+
+All observations and the trace use schema version `2`. Schema version 2 adds indexed `pi.attempt` observations and active `pi.compaction` spans beneath the root agent.
+
+### Trace and attempt fields
+
+The trace and root `pi.agent` retain the submitted prompt, final assistant output, Pi session id, working directory, mode, initial provider/model, and optional Git context. Root metadata includes:
+
+- `pi.trace.schema_version`, `pi.trace.outcome`, and `pi.trace.stop_reason`;
+- `pi.trace.attempt_count`, `pi.trace.turn_count`, `pi.trace.generation_count`, `pi.trace.tool_count`, `pi.trace.tool_error_count`, `pi.trace.compaction_count`, and `pi.trace.recovered_error_count`;
+- `pi.trace.start_leaf_id`, `pi.trace.end_leaf_id`, `pi.trace.start_context_tokens`, `pi.trace.end_context_tokens`, `pi.trace.start_context_window`, `pi.trace.end_context_window`, `pi.trace.start_context_percent`, and `pi.trace.end_context_percent` when Pi knows them;
+- `pi.git.branch`, `pi.git.commit`, and `pi.git.detached`, plus a `branch:<branch-name>` tag or `git:detached` tag.
+
+Outcomes are `success`, `recovered_success`, `error`, `aborted`, `length`, or `interrupted`. `pi.trace.recovered_error_count` includes recovered provider responses, tool failures handled by a later generation, and failed attempts followed by final success. Errors use Langfuse `ERROR`; aborts, output limits, shutdown, replacement, and other interruption closures use `WARNING`. High-cardinality correlation values stay in metadata rather than tags.
 
-- a `pi.turn` native `span` for every Pi turn, including its index, stop reason, tool-result count, duration, and failure status;
-- a `pi.llm` native `generation` under the active turn for every provider request;
-- a `pi.tool.<tool-name>` native `tool` observation under the active turn for every tool execution;
-- the Pi session id, working directory, mode, provider, and model;
-- generation token usage and positive total cost when Pi reports a known price;
-- the concrete response model when Pi reports one, with a differing requested alias retained in metadata;
-- error levels and status messages for failed provider responses, model calls, and tools.
+Each `pi.attempt` records `pi.attempt.index`, final `pi.attempt.outcome`, and `pi.attempt.stop_reason`. An attempt immediately following overflow compaction also sets `pi.attempt.reason` to `post_compaction`. Failed attempts remain errors even when a later attempt makes the root a recovered success.
 
-Pi does not expose a post-transform provider payload event, so generation request bodies are intentionally omitted rather than risking capture of a payload that a later extension rewrites or redacts. The agent observation still records the user prompt, and assistant output is reconciled after message transformers. Tool input is captured after argument preparation and `tool_call` mutations, while tool output is captured after `tool_result` transformers.
+### Generation fields
 
-Images and embedded base64 data URIs are represented without their payloads, including provider data URLs. Every captured input or output has one cumulative 64 KiB serialized UTF-8 budget, bounded object/array traversal, and deterministic truncation markers. Langfuse credentials are masked again in the span processor before network export.
+Each `pi.llm` generation records:
 
-Completed observations are exported in batches while Pi remains live. Normal `agent_end` handling never waits for Langfuse network I/O. Use `/langfuse flush` when you need to wait for completed exports; quit shutdown also drains the provider.
+- a bounded input snapshot from this extension's `before_provider_request` handler and `pi.request.payload_stage` set to `before_provider_request`;
+- `pi.request.provider`, `pi.request.model`, `pi.request.api`, and `pi.request.thinking_level`, with thinking level also exported through Langfuse-native model parameters;
+- the Langfuse-native response model plus `pi.response.provider`, `pi.response.api`, `pi.response.model`, and `pi.response.id` when Pi reports them;
+- Langfuse-native `completionStartTime` from the first non-empty text, thinking, or tool-call delta;
+- ordered `http.response.status_codes`, final `http.response.status_code`, `http.response.attempt_count`, and `http.response.retry_count`;
+- allowlisted `http.response.headers`: request ids, `cf-ray`, `retry-after`, and the supported OpenAI/Anthropic rate-limit headers. Authorization, cookies, and unrecognized headers are never exported;
+- additive input, output, cache-read, cache-write, and total token usage plus known positive input, output, cache-read, cache-write, and total cost buckets;
+- non-additive `pi.usage.reasoning_tokens` and `pi.usage.cache_write_1h_tokens` in metadata so subsets are not double-counted.
 
-Automatic retries or continuations that begin without a new user prompt are recorded as a new trace labeled `[automatic continuation]`, so provider activity is not lost on Pi versions without a final `agent_settled` extension event.
+The request snapshot is the payload visible at this handler, not a guaranteed final wire payload: later extensions can still replace it. Final assistant content is reconciled from `turn_end` and `agent_end` after message transformation. A recovered sequence such as `429 -> 200` remains queryable in HTTP metadata but is not an error; the final assistant outcome decides generation severity.
+
+### Tool and compaction fields
+
+A `pi.tool.<tool-name>` observation starts with raw `tool_execution_start` arguments as a fallback for calls that never execute, including calls blocked during `tool_call`. For executed calls, it uses the `tool_result` input as authoritative after all argument mutations. It captures final transformed output from `tool_execution_end`, final error state, `pi.tool.progress_update_count`, and `pi.tool.time_to_first_progress_ms` when progress occurs. An unrecovered tool failure also makes the attempt and root errors. Existing `pi.tool.call_id` and `pi.tool.name` correlation fields remain. `tool_execution_update` partial-result bodies are never captured. Duplicate, parallel, failed, no-progress, and interrupted tools are closed independently.
+
+An active `pi.compaction` records `pi.compaction.reason`, `pi.compaction.will_retry`, `pi.compaction.from_extension`, `pi.compaction.tokens_before`, `pi.compaction.messages_to_summarize`, `pi.compaction.turn_prefix_messages`, `pi.compaction.branch_entries`, and `pi.compaction.is_split_turn`. It adds `pi.compaction.read_file_count`, `pi.compaction.modified_file_count`, and `pi.compaction.usage.*` / `pi.compaction.cost.*` when Pi reports them. It never records the summary, custom instructions, or message bodies. Manual compaction outside an active agent trace is ignored; incomplete compaction closes as a warning at settlement or shutdown.
+
+### Boundaries and export
+
+Images and embedded base64 data URIs are represented without their payloads, including provider data URLs. Opaque `thinkingSignature`, `textSignature`, and `thoughtSignature` continuity values are always removed. Every captured input or output has one cumulative 64 KiB serialized UTF-8 budget, bounded object/array traversal, and deterministic truncation markers. Langfuse credentials are masked again in the span processor before network export.
+
+The root agent begins before the first agent loop and remains open across retries, overflow-compaction recovery, and queued continuations. `agent_end` closes only the current attempt; `agent_settled` closes the root after no automatic work remains. Activity that unexpectedly arrives without a submitted prompt gets a fallback root input labeled `[automatic continuation]`. Session replacement, reload, quit, and a new unexpected prompt close all descendants defensively and idempotently.
+
+At run start, the extension performs bounded, non-shell Git lookups in `ctx.cwd`. A branch switch therefore applies to the next run. Detached HEADs retain only commit/detached metadata and the `git:detached` tag. Missing Git, non-repositories, timeouts, and lookup failures silently omit Git context without affecting tracing.
+
+Completed observations are exported in batches while Pi remains live. Neither `agent_end` nor `agent_settled` waits for Langfuse network I/O. To wait for completed exports, run `/langfuse` and choose **Flush completed traces for this session**; quit shutdown also drains the provider.
 
 ## 💬 Command
 
 ```text
-/langfuse status
-/langfuse flush
-/langfuse help
-/langfuse init
+/langfuse
 ```
 
-- `status` reports whether tracing is enabled, the endpoint, configuration source, and content-capture mode. It never displays credentials.
-- `flush` waits for all completed observations to export.
-- `help` displays command guidance.
-- `init` interactively creates or updates the private config without displaying existing credentials. Blank keys preserve valid existing values; a blank base URL uses the US cloud endpoint.
+The command opens one context-aware menu. Its title shows the current session's tracing state, endpoint, content-capture mode, initialization failure when applicable, and private configuration path. It never displays credentials.
+
+Available actions depend on that state:
+
+- **Flush completed traces for this session** appears first when tracing is active and waits for completed observations to export.
+- **Set up Langfuse for this Pi agent directory** appears when no valid config was loaded.
+- **Update Langfuse for this Pi agent directory** appears when a valid config exists.
+- **Show setup and privacy help** explains the agent-directory scope, manual configuration path, and content-capture risk.
+
+Connection actions state their agent-directory scope and per-process restart requirement before selection. Command arguments are intentionally ignored so remembered subcommands cannot silently bypass the menu. In non-interactive modes, the command reports that the menu is unavailable and points to the manual config path.
 
 ## 🔐 Privacy
 
 With content capture enabled, traces can contain user prompts, model responses, tool arguments, and tool results. These may include source code, file contents, shell output, or other sensitive project data. Review your Langfuse retention and access controls before enabling this extension.
 
-The built-in mask specifically protects Langfuse credentials; it is not a general secret scanner. Set `"captureContent": false` in `pi-langfuse.json` when content must remain local.
+Git branch names, commit ids, working directory, session/leaf ids, model identity, usage/cost, aggregate counts, and allowlisted response-header values are metadata. They remain exported when `captureContent` is `false`; branch names and diagnostic header values can themselves contain operational details.
+
+The built-in mask specifically protects Langfuse credentials; it is not a general secret scanner. Set `"captureContent": false` in `pi-langfuse.json` when prompts, provider-request snapshots, responses, and tool content must remain local. Compaction summaries, tool partial results, opaque continuation signatures, authorization headers, cookies, and unapproved response headers are never exported in either mode.
 
 ## 🗂️ Package layout
 
@@ -114,7 +169,8 @@ The built-in mask specifically protects Langfuse credentials; it is not a genera
 extensions/pi-langfuse/
 ├── src/
 │   ├── langfuse.ts  # Pi lifecycle integration and slash command
-│   ├── tracing.ts   # Trace lifecycle and content bounding
+│   ├── tracing.ts   # Observation lifecycle, outcomes, and bounded metadata
+│   ├── sanitizer.ts # Content bounding and opaque-signature removal
 │   ├── runtime.ts   # Langfuse/OpenTelemetry runtime
 │   └── config.ts    # Private pi-langfuse.json loading and validation
 ├── test/

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@narumitw/pi-langfuse",
-	"version": "0.18.0",
+	"version": "0.26.0",
 	"description": "Pi extension that traces LLM generations and tool activity to Langfuse.",
 	"type": "module",
 	"license": "MIT",
@@ -29,18 +29,18 @@
 		"typecheck": "tsc --noEmit"
 	},
 	"dependencies": {
-		"@langfuse/otel": "^4.0.0",
+		"@langfuse/otel": "^5.9.1",
 		"@langfuse/tracing": "^4.0.0",
-		"@opentelemetry/api": "^1.9.0",
+		"@opentelemetry/api": "^1.9.1",
 		"@opentelemetry/exporter-trace-otlp-http": "^0.220.0",
 		"@opentelemetry/sdk-trace-base": "^2.9.0",
 		"@opentelemetry/sdk-trace-node": "^2.9.0"
 	},
 	"devDependencies": {
 		"@biomejs/biome": "2.5.3",
-		"@earendil-works/pi-coding-agent": "0.80.3",
+		"@earendil-works/pi-coding-agent": "0.80.10",
 		"@types/node": "26.1.1",
-		"typescript": "6.0.3"
+		"typescript": "7.0.2"
 	},
 	"repository": {
 		"type": "git",

package/src/config.ts

@@ -130,6 +130,16 @@ export function normalizeLangfuseConfig(
 	}
 	const environment = optionalString(input.environment, "environment");
 	if (!environment.ok) return environment;
+	if (
+		environment.value &&
+		(environment.value.length > 40 || !/^(?!langfuse)[a-z0-9_-]+$/u.test(environment.value))
+	) {
+		return {
+			ok: false,
+			reason:
+				"pi-langfuse.json environment must be at most 40 lowercase letters, numbers, hyphens, or underscores and must not start with langfuse.",
+		};
+	}
 	const release = optionalString(input.release, "release");
 	if (!release.ok) return release;