@dailephd/my-dev-kit-lab 0.8.2 → 0.9.1

package/CHANGELOG.md

@@ -2,8 +2,27 @@
 
 All notable changes to my-dev-kit-lab are documented here.
 
+## [0.9.1] - 2026-10-12
+
+### Normalized provider telemetry and resumable campaigns
+
+- Normalized provider token and duration provenance while preserving legacy parser compatibility and separating unavailable telemetry from task outcomes.
+- Added deterministic agent-success-rate campaign queues, immutable checkpoint generations, version 1 compatibility and version 2 recovery epochs, explicit resume and verified completed-unit reuse, and bounded provider-limit continuation.
+- Added public campaign CLI options, Codex and Claude timeout presets, campaign progress summaries, and strong/partial/unavailable metric-specific evidence classifications.
+- Preserved partial evidence without presenting it as complete scientific analysis; uncertain provider invocations are not replayed automatically. Real-provider use remains explicitly opted in and depends on the existing provider adapters.
+
 ## [Unreleased]
 
+## [0.9.0] - 2026-10-10
+
+### Agent success rate
+
+- Added the experimental `agent-success-rate` plugin as the eighth registered experiment plugin, running the fixed `raw-full-file` and `context-pack` treatments over a bundled six-task implementation corpus (two Node benchmark projects; localized, cross-module, and broad-change tasks) in guarded disposable copies, with trusted task and regression checks, behavior facts, changed-file capture, and preserved patch artifacts.
+- Added a deterministic-fixture default mode (reference patches, no coding agent) and an explicit real-agent mode (`--agent <codex|claude> --include-real-agents`) in which the Lab validates and applies a provider-proposed unified diff itself, with optional bounded repair (`--repair-attempts 0|1|2`, at most three attempts per treatment, a fresh copy per attempt).
+- Added edit-quality, blast-radius, duration, and provider-token metrics with explicit availability, separate initial-attempt and final success, descriptive paired comparisons, and JSON, HTML, and text reports; no composite score or automatic winner is produced.
+- Extended packed-package acceptance to cover the installed CLI, bundled corpus and resources, simulated providers, repair campaigns, cleanup, and package hygiene; the deterministic reference patches ship as required resources and stay hidden from real-agent prompts.
+- The shared measured-command runner now handles errors emitted by child output streams, so an unsupported deep Windows working directory ends in a controlled infrastructure failure instead of an unhandled `ENOTCONN` error.
+
 ## [0.8.2] - 2026-10-07
 
 ### Context-pack generation experiments

package/README.md

@@ -6,9 +6,11 @@ my-dev-kit provides local repository indexing and graph-guided retrieval. my-dev
 
 Whole-ecosystem workflows are centralized in [my-dev-kit/docs/ECOSYSTEM_DEVELOPMENT_WORKFLOWS.md](https://github.com/dailephd/my-dev-kit/blob/main/docs/ECOSYSTEM_DEVELOPMENT_WORKFLOWS.md). Its [command-surface compatibility map](https://github.com/dailephd/my-dev-kit/blob/main/docs/ECOSYSTEM_DEVELOPMENT_WORKFLOWS.md#915-command-surface-compatibility-map) documents which Lab outputs can safely seed my-dev-kit investigation, Observer reference work, or Orchestrator evidence, and which look-alike artifacts are not direct inputs.
 
-The latest published release is v0.8.2 (context-pack generation experiments), and the repository package version is 0.8.2. The previous published release is v0.8.1 (retrieval query strategy comparison). The v0.4.5 context-integrity evaluation remains intentionally frozen against the published `@dailephd/my-dev-kit@1.10.4` and `@dailephd/my-dev-kit-orchestrator@1.2.3` contracts; those versions are historical validation baselines, not a statement that they are the ecosystem's current releases. See [docs/CURRENT_STATE.md](docs/CURRENT_STATE.md) for current state.
+The latest release is v0.9.1 (normalized provider telemetry and resumable agent-success-rate campaigns), and the repository package version is 0.9.1. The previous release is v0.9.0 (agent-success-rate implementation-patch evaluation). The v0.4.5 context-integrity evaluation remains intentionally frozen against the published `@dailephd/my-dev-kit@1.10.4` and `@dailephd/my-dev-kit-orchestrator@1.2.3` contracts; those versions are historical validation baselines, not a statement that they are the ecosystem's current releases. See [docs/CURRENT_STATE.md](docs/CURRENT_STATE.md) for current state.
 
-The next planned milestone after v0.8.2 is v0.9.0 (agent-success-rate experiments).
+v0.9.0 adds the `agent-success-rate` plugin, the eighth registered experiment plugin, which evaluates implementation patches under the `raw-full-file` and `context-pack` treatments.
+
+v0.9.1 adds normalized provider telemetry and resumable campaign execution. v0.9.2 remains planned.
 
 v0.4.6 adds a supported `my-dev-kit-lab` installed CLI router (`--help`, `--version`, `security validate`, `audit`, the `experiment` family, `report render`, `plots generate`, `gallery build`, `demo final`, and the historical direct final-demo invocation form), a writable lab workspace model kept separate from the installed package and the inspected target, and a permanent packed-tarball installation/execution acceptance gate (`npm run verify:packed-package`). See [docs/ROADMAP.md](docs/ROADMAP.md) and [docs/CURRENT_STATE.md](docs/CURRENT_STATE.md) for status detail and [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md) for the runtime path model. The "Installed CLI" section below documents the shipped command surface; the source-checkout `npm run` workflow in Quickstart remains available for contributors.
 
@@ -25,6 +27,8 @@ v0.4.6 adds a supported `my-dev-kit-lab` installed CLI router (`--help`, `--vers
 - **Measure retrieval precision and recall (v0.8.0, released):** the `retrieval-precision-recall` plugin runs the existing my-dev-kit search, lookup, slice, and source lifecycle for each case and compares the retrieved files and symbols with a deterministic answer key, without invoking any agent. It reports file precision and recall, symbol precision and recall, fact coverage (from explicit fact-to-context mappings), irrelevant context ratio, the estimated retrieved token count, and the missed and irrelevant identities, as `retrieval-precision-recall-execution.json` plus `report.json`, `report.html`, and `report.txt`. By default it runs a bundled 12-case corpus over two benchmark projects against the Lab itself. It can also run over an explicitly selected local Git repository with `--target <repository> --local-subject-config <path>`; in that mode the repository is treated as read-only and durable outputs redact the repository path and file, symbol, and fact identities while keeping the numeric evidence. These are set-based measures with no ranking, winner, or composite score. See [docs/COMMANDS.md](docs/COMMANDS.md), [docs/WORKFLOWS.md](docs/WORKFLOWS.md), and [docs/METRICS.md](docs/METRICS.md).
 - **Compare retrieval query strategies (v0.8.1, released):** the `retrieval-query-strategy-comparison` plugin compares seven fixed my-dev-kit retrieval strategies without running any coding agent: `keyword-search`, `symbol-lookup`, `graph-neighborhood`, `source-slice`, `data-model-graph`, `model-view-lineage`, and `combined-graph-guided`. In bundled mode it uses the same frozen 12-case warm-index benchmark corpus as `retrieval-precision-recall`, and every selected case receives all seven strategies (filters narrow cases or projects, never strategies). File and symbol quality keeps the inherited v0.8.0 set-based semantics and adds balanced file and symbol F1 (`2TP / (2TP + FP + FN)`); fact coverage and the estimated retrieved token count complete the four comparison objectives. Strategies are compared over matched complete cases only, as macro means for four scopes (`overall`, `localized`, `cross-module`, `broad-change`, reusing each case's existing task-locality label), using Pareto dominance on mean file F1, mean symbol F1, mean fact coverage (all maximized) and mean retrieved tokens (minimized). A best strategy is named only when the Pareto front has a single member; when several strategies remain nondominated the report states a tradeoff. There is no scalar or composite score and no ordinal ranking. The plugin can also run over an explicitly selected local Git repository through the existing `LocalRepositorySubject` boundary with `--target <repository> --local-subject-config <path>`; the repository stays read-only and durable output redacts private file, symbol, fact, semantic-node, warning, and case-title identities while preserving the numeric results. It writes `retrieval-query-strategy-comparison-execution.json`, `retrieval-query-strategy-comparison-analysis.json`, and `report.json`, `report.html`, and `report.txt`. No context-pack behavior is included. See [docs/COMMANDS.md](docs/COMMANDS.md), [docs/WORKFLOWS.md](docs/WORKFLOWS.md), and [docs/METRICS.md](docs/METRICS.md#retrieval-query-strategy-comparison-metrics-v081).
 - **Generate auditable task context packs (v0.8.2):** `context-pack-generation` runs both fixed treatments, `raw-full-file` and `context-pack`, for each selected case. A pack contains a task summary, relevant files and symbols, bounded source slices, call relationships, tests, and evidence notes. Analysis reuses retrieval fact coverage and file/symbol precision, recall, and F1, reports estimated context size and paired raw-versus-pack tokens saved, percent saved, and objective deltas, and names no composite winner or ranking. Bundled runs persist pack artifacts and reports include a bounded preview. External-local runs calculate science on real identities, redact before persistence, omit durable pack bodies, and show redacted report previews. See [docs/COMMANDS.md](docs/COMMANDS.md), [docs/WORKFLOWS.md](docs/WORKFLOWS.md), and [docs/METRICS.md](docs/METRICS.md).
+- **Run resumable agent-success-rate campaigns (v0.9.1):** provider telemetry is normalized with provenance, and explicit campaign runs support deterministic case ordering, checkpoints, one-case execution, validated resume, safe completed-unit reuse, provider-limit pauses, authorized bounded continuation, and plugin-specific progress evidence. Campaigns use the existing Codex and Claude adapters only after explicit real-provider opt-in. Checkpoint/progress evidence is operational; complete scientific reports are available only after finalization.
+- **Evaluate implementation patches (v0.9.0):** the `agent-success-rate` plugin runs both `raw-full-file` and `context-pack` treatments over a bundled six-task implementation corpus in disposable benchmark copies and decides success from trusted checks on the changed code. The default deterministic mode applies reference patches and runs no coding agent; real Codex or Claude runs (with optional bounded repair) are explicit opt-in. It runs from the installed CLI or a built source checkout (see [docs/COMMANDS.md](docs/COMMANDS.md#agent-success-rate-v090)).
 - **Assess affected-neighborhood evidence (v0.6.1, released; retained in the current package):** using the baseline my-dev-kit graph of the same prepared warm index (loaded once per session), each task's confirmed modified or missing indexed files, and the baseline symbols those files contain, are mapped to graph nodes. The affected neighborhood is those seed nodes plus their direct one-hop neighbors (edges of any kind, either direction). The task's `expectedFiles` and file-scoped `expectedSymbols` are mapped to graph nodes and compared with that neighborhood, which yields six numeric warm-side metrics (`changedFileCount`, `changedSymbolCount`, `affectedNodeCount`, `affectedEdgeCount`, `taskOverlapCount`, `taskOverlapPercent`) plus categorical evidence: a `related`, `unrelated`, or `unknown` relationship and a `recommended`, `not-indicated`, or `unknown` `reindexRecommendation`. Incomplete evidence is reported as `unknown`, never as unrelated. The assessment is recorded additively in `warm-index-execution.json` and shown in the existing `report.json`/`report.txt`/`report.html` warm-index reports. It is observational only: retrieval, execution status, and provider status are unchanged, nothing is reindexed automatically, and no CLI flag, experiment plugin, plot, or gallery item was added. See [docs/METRICS.md](docs/METRICS.md#affected-neighborhood-metrics-v061) and [docs/WORKFLOWS.md](docs/WORKFLOWS.md#warm-index-reuse-experiment).
 - **Audit repository health:** run conservative code-rot detectors for TypeScript/JavaScript, Python, Java, and Kotlin, or adapt security findings into the common audit report.
 - **Validate CLI/package security:** inspect dependencies, package contents, path and subprocess boundaries, malformed inputs, optional static scanners, and bounded fuzz targets.
@@ -196,6 +200,18 @@ For experiment plugins that support explicit targets — `context-strategy-compa
 
 ---
 
+### Evaluate implementation patches (v0.9.0)
+
+The `agent-success-rate` plugin ships in the package and runs from the installed CLI or a built source checkout. The default run is deterministic: it applies each task's reference patch in a disposable copy, runs the trusted checks, and launches no coding agent and no provider.
+
+```bash
+npm run build
+node dist/scripts/cli.js experiment describe --experiment agent-success-rate
+node dist/scripts/cli.js experiment run --experiment agent-success-rate --case asr-board-title-normalization --out lab-output/agent-success-rate
+```
+
+This validates the evaluation pipeline and the benchmark corpus; it does not measure agent ability. Real Codex or Claude evaluation requires both `--agent <codex|claude>` and `--include-real-agents`, and optional `--repair-attempts <0|1|2>`; see [docs/COMMANDS.md](docs/COMMANDS.md#agent-success-rate-v090) and [docs/WORKFLOWS.md](docs/WORKFLOWS.md#agent-success-rate-evaluation-v090). A live paid-provider campaign has not yet been run.
+
 ## Where to find outputs
 
 | Artifact | Location |
@@ -244,7 +260,7 @@ See [docs/METRICS.md](docs/METRICS.md) for full metric definitions.
 - Claude and Codex token totals are used only when the CLI/run exposes them; missing provider token evidence stays explicitly unavailable (never zero, and never substituted with an estimated context-token count)
 - Codex and Claude campaign runs can produce timeouts, invalid-output, agent-unavailable, or agent-limit-reached outcomes; these are structured provider states, not warm-index infrastructure failures
 - Small projects may make raw-full-file cheaper than my-dev-kit-guided; larger localized tasks are where my-dev-kit is expected to become more useful
-- The generic experiment-plugin framework has seven registered plugins: `context-strategy-comparison`, `warm-index-reuse`, `incremental-change-staleness`, `context-window-scaling`, `retrieval-precision-recall`, `retrieval-query-strategy-comparison`, and `context-pack-generation`. The v0.7.0 context-window-scaling plugin is released; v0.7.1 added caller-supplied synthetic repositories; v0.7.2 added an explicitly selected local Git repository. The context-pack-generation plugin shipped in v0.8.2; agent-success plugins remain planned.
+- The generic experiment-plugin framework has eight registered plugins: `context-strategy-comparison`, `warm-index-reuse`, `incremental-change-staleness`, `context-window-scaling`, `retrieval-precision-recall`, `retrieval-query-strategy-comparison`, `context-pack-generation`, and `agent-success-rate` (v0.9.0). The v0.7.0 context-window-scaling plugin is released; v0.7.1 added caller-supplied synthetic repositories; v0.7.2 added an explicitly selected local Git repository. The context-pack-generation plugin shipped in v0.8.2; agent-success-rate shipped in v0.9.0 and gained resumable campaigns in v0.9.1.
 - `retrieval-precision-recall` (v0.8.0) measures set-based retrieval quality, not ranked retrieval: it has no ranking metrics, no retrieval-strategy comparison of its own (the separate `retrieval-query-strategy-comparison` plugin added in v0.8.1 compares strategies), no winner or composite score, and no real agents. Its lifecycle expands only the top search candidate. An external-local run fails as a whole if any case fails a safety or execution check, and then writes no normal artifact or report family. Durable external-local identities are redacted, although the console summary of a successful run can still show the physical output directory. Platform path-length limits can make a run with very long paths fail cleanly instead of executing.
 - `retrieval-query-strategy-comparison` (v0.8.1) runs no coding agent and does not evaluate answer quality. Its retrieved token count is a context-size estimate, not provider billing telemetry. Semantic strategies (`data-model-graph`, `model-view-lineage`) can truthfully be partial or unavailable for a case, and a case is then excluded from the matched cross-strategy comparison for every strategy while its per-strategy evidence is kept. The Pareto front is a set of nondominated strategies in canonical strategy order, not an ordinal ranking. The plugin reports no significance test or confidence interval and selects no strategy automatically. External-local mode rejects `--case` and `--benchmark-project`, builds one private base index per configured case, and fails as a whole, writing no normal artifact or report family, if any safety or privacy check fails.
 - `warm-index-reuse` without `--campaign` still uses the deterministic fake agent only. With `--campaign`, exactly one real provider (Codex or Claude) runs per campaign preset — there is no multi-agent matrix, retry, provider switching, or scheduler. The v0.5.1 expanded corpus is deterministic fake-agent benchmark evidence when run without `--campaign`. No warm-index mode calculates a token-savings percentage, break-even task, winner, or ranking, and component-duration sums are not wall-clock latency

package/benchmarks/contracts/agent-success-rate-project-profiles.json

@@ -0,0 +1,33 @@
+{
+  "schemaVersion": "my-dev-kit-lab-agent-success-project-profiles-v1",
+  "projects": [
+    {
+      "projectId": "agent-success-task-board-node",
+      "displayName": "Task Board (Node)",
+      "description": "Dependency-free in-memory task board used as an implementation benchmark: title validation, task storage, external-task import, project ownership and per-project summaries.",
+      "rootPath": "benchmarks/projects/agent-success-task-board-node",
+      "sourceRoots": [
+        "src"
+      ],
+      "testRoots": [
+        "tests"
+      ],
+      "primaryLanguage": "javascript",
+      "runtime": "node>=24"
+    },
+    {
+      "projectId": "agent-success-inventory-node",
+      "displayName": "Inventory and Fulfillment (Node)",
+      "description": "Dependency-free in-memory inventory used as an implementation benchmark: quantity validation, stock and reservation storage, reservation lifecycle and a fulfillment report.",
+      "rootPath": "benchmarks/projects/agent-success-inventory-node",
+      "sourceRoots": [
+        "src"
+      ],
+      "testRoots": [
+        "tests"
+      ],
+      "primaryLanguage": "javascript",
+      "runtime": "node>=24"
+    }
+  ]
+}