pi-autoresearch-vkf 0.5.0 → 0.5.2

package/CHANGELOG.md

@@ -1,6 +1,26 @@
 # Changelog
 
-## 0.5.0 — unreleased
+## 0.5.2
+
+Prefixed all skill names with `autoresearch-vkf-` to avoid namespace conflicts
+with other tooling. Renamed `knowledge-gather`, `claim-extract`, `claim-verify`,
+`contradiction-miner`, `cross-domain-transfer`, `idea-tournament`,
+`hypothesis-loop`, and `research-report`; all cross-references in the skills,
+the README, and the extension were updated accordingly. No behavior change.
+
+
+## 0.5.1
+
+Fix tool/skill name collisions with pi-autoresearch (both can now load together).
+
+- Tools `run_experiment` → **`vkf_run_experiment`** and `log_experiment` →
+  **`vkf_log_experiment`** (pi-autoresearch registers the bare names; pi requires
+  globally-unique tool names across loaded extensions).
+- Skill `autoresearch-create` → **`autoresearch-vkf`** (pi-autoresearch ships a
+  skill of the same name). Invoke the loop via the `autoresearch-vkf` skill now.
+- Docs/skills/benchmark updated accordingly. No behavior change.
+
+## 0.5.0 — first published release
 
 Self-contained workspace (breaking path change).
 

package/README.md

@@ -51,7 +51,7 @@ pi install file:/path/to/pi-autoresearch-vkf
 ### Knowledge sources (how ingestion works)
 
 The extension stores and reasons over knowledge; it does **not** fetch papers
-itself. Gathering is done by the host agent through the `knowledge-gather` skill,
+itself. Gathering is done by the host agent through the `autoresearch-vkf-knowledge-gather` skill,
 using the agent's built-in **`WebSearch` + `WebFetch`** against free, openly
 accessible databases — no API keys, no paid services, no MCP setup:
 
@@ -74,7 +74,7 @@ In a project you want to optimize:
 optimize the test suite runtime, using the research literature and remembering what works
 ```
 
-The **autoresearch-create** skill drives it: confirm goal/metric/command → init →
+The **autoresearch-vkf** skill drives it: confirm goal/metric/command → init →
 gather literature → extract & verify claims → loop (recall → experiment →
 write-back) → report. All state lives in one self-contained `.autoresearch-vkf/`
 folder at the project root, so work **survives restarts and context resets**.
@@ -86,12 +86,12 @@ goal ─► recall_memory ─► gather literature ─► remember_claim (candid
    │                                              │
    │                                         verify_claim ──► trusted claims
    ▼                                              │
- hypothesis-loop:  recall ─► pick idea ─► run_experiment ─► log_experiment
+ autoresearch-vkf-hypothesis-loop:  recall ─► pick idea ─► vkf_run_experiment ─► vkf_log_experiment
    │                                                            │
    │                                  writes experiment card back to memory,
    │                                  updates the claim's belief & lifecycle
    ▼
- research-report   (paper → claim → hypothesis → patch → metric Δ → memory update)
+ autoresearch-vkf-research-report   (paper → claim → hypothesis → patch → metric Δ → memory update)
 ```
 
 ### One self-contained workspace
@@ -135,8 +135,8 @@ verifier — is the defense against **memory poisoning**.
 | `score_ideas` | Rank untested ideas by `EV × feasibility × evidence × novelty × info_gain ÷ cost`. |
 | `find_contradictions` | Mine memory for tensions between claims — each a seed for a novel hypothesis. |
 | `find_transfers` | Cross-domain mechanism search: same *how*, different *where*. |
-| `run_experiment` | Run the measurement command; capture `METRIC name=value`. |
-| `log_experiment` | Record a result, write it back to memory, update belief & lifecycle. |
+| `vkf_run_experiment` | Run the measurement command; capture `METRIC name=value`. |
+| `vkf_log_experiment` | Record a result, write it back to memory, update belief & lifecycle. |
 | `promote_to_global` | Copy a trusted card into the cross-project global memory. |
 | `export_dashboard` | Write browser dashboards: a live progress page + the `vkf html` idea-lineage graph. |
 | `research_status` | Show session experiments + memory lifecycle. |
@@ -145,15 +145,15 @@ verifier — is the defense against **memory poisoning**.
 
 | Skill | Role |
 |-------|------|
-| `autoresearch-create` | Orchestrator / spine — the entry point. |
-| `knowledge-gather` | Find candidate techniques via WebSearch/WebFetch (arXiv / Semantic Scholar / OpenAlex / GitHub). |
-| `claim-extract` | Distill sources into reusable claim cards. |
-| `claim-verify` | Check citations & codebase fit — the trust layer. |
-| `contradiction-miner` | Turn tensions in memory into novel hypotheses. |
-| `cross-domain-transfer` | Import a mechanism from another field. |
-| `idea-tournament` | Multi-perspective debate to pick the 2–3 ideas worth testing. |
-| `hypothesis-loop` | Pick the next idea and run the smallest falsifying experiment. |
-| `research-report` | The auditable lineage report. |
+| `autoresearch-vkf` | Orchestrator / spine — the entry point. |
+| `autoresearch-vkf-knowledge-gather` | Find candidate techniques via WebSearch/WebFetch (arXiv / Semantic Scholar / OpenAlex / GitHub). |
+| `autoresearch-vkf-claim-extract` | Distill sources into reusable claim cards. |
+| `autoresearch-vkf-claim-verify` | Check citations & codebase fit — the trust layer. |
+| `autoresearch-vkf-contradiction-miner` | Turn tensions in memory into novel hypotheses. |
+| `autoresearch-vkf-cross-domain-transfer` | Import a mechanism from another field. |
+| `autoresearch-vkf-idea-tournament` | Multi-perspective debate to pick the 2–3 ideas worth testing. |
+| `autoresearch-vkf-hypothesis-loop` | Pick the next idea and run the smallest falsifying experiment. |
+| `autoresearch-vkf-research-report` | The auditable lineage report. |
 
 ### The `.autoresearch-vkf/` workspace
 
@@ -282,7 +282,7 @@ Verify what will ship first with `npm pack --dry-run`.
 
 All four planned phases are in: the lean MVP (Phase 1), the **novelty scorer**
 (Phase 2), the **hypothesis-synthesis layer** (Phase 3 — `find_contradictions`,
-`find_transfers`, `idea-tournament`), and **global cross-project memory + the
+`find_transfers`, `autoresearch-vkf-idea-tournament`), and **global cross-project memory + the
 benchmark** (Phase 4).
 
 Possible next steps:

package/extensions/pi-autoresearch-vkf/index.ts

@@ -112,7 +112,7 @@ export default function autoresearchExtension(pi: ExtensionAPI): void {
       if (existing) {
         refreshWidget(ctx, root);
         return textResult(
-          `A research session already exists: "${existing.name}".\nSession: ${sp.dir}\nMemory:  ${memoryPaths(root).dir}\nContinue the loop with recall_memory → run_experiment → log_experiment.`,
+          `A research session already exists: "${existing.name}".\nSession: ${sp.dir}\nMemory:  ${memoryPaths(root).dir}\nContinue the loop with recall_memory → vkf_run_experiment → vkf_log_experiment.`,
           { created: false },
         );
       }
@@ -144,7 +144,7 @@ export default function autoresearchExtension(pi: ExtensionAPI): void {
           `Memory bundle: ${memoryPaths(root).dir} ${fresh ? "(new)" : "(existing)"} — profile ${config.memoryProfile}.`,
           `Optimizing ${config.metricName} (${config.direction} is better).`,
           "",
-          "Next: gather literature (knowledge-gather skill) → remember_claim candidates → verify_claim → recall_memory to pick an idea → run_experiment → log_experiment.",
+          "Next: gather literature (autoresearch-vkf-knowledge-gather skill) → remember_claim candidates → verify_claim → recall_memory to pick an idea → vkf_run_experiment → vkf_log_experiment.",
         ].join("\n"),
         { created: true },
       );
@@ -491,7 +491,7 @@ export default function autoresearchExtension(pi: ExtensionAPI): void {
 
       if (ideas.length === 0) {
         return textResult(
-          "No untested ideas to score. Gather literature (knowledge-gather) and remember_claim some candidates first.",
+          "No untested ideas to score. Gather literature (autoresearch-vkf-knowledge-gather) and remember_claim some candidates first.",
           { ranked: 0 },
         );
       }
@@ -620,7 +620,7 @@ export default function autoresearchExtension(pi: ExtensionAPI): void {
     },
   });
 
-  // ── run_experiment ─────────────────────────────────────────────────────────
+  // ── vkf_run_experiment ───────────────────────────────────────────────────────
   const RunParams = Type.Object({
     command: Type.Optional(Type.String({ description: "Command to run (via `bash -lc`). Defaults to the session's configured command." })),
     claim_id: Type.Optional(Type.String({ description: "The claim/idea this run is testing, for logging." })),
@@ -629,10 +629,10 @@ export default function autoresearchExtension(pi: ExtensionAPI): void {
   });
 
   pi.registerTool({
-    name: "run_experiment",
+    name: "vkf_run_experiment",
     label: "Run experiment",
     description:
-      "Run the measurement command and capture its output and any `METRIC name=number` lines. Does not judge or record an outcome — read the metric, then record it with log_experiment.",
+      "Run the measurement command and capture its output and any `METRIC name=number` lines. Does not judge or record an outcome — read the metric, then record it with vkf_log_experiment.",
     parameters: RunParams,
     async execute(_id, params: Static<typeof RunParams>, signal, _onUpdate, ctx): Promise<AgentToolResult<{ code: number; metrics: Record<string, number> }>> {
       const root = resolveRoot(ctx);
@@ -667,7 +667,7 @@ export default function autoresearchExtension(pi: ExtensionAPI): void {
     },
   });
 
-  // ── log_experiment ─────────────────────────────────────────────────────────
+  // ── vkf_log_experiment ───────────────────────────────────────────────────────
   const LogParams = Type.Object({
     description: Type.String({ description: "What was changed in this experiment, in words." }),
     value: Type.Number({ description: "The metric value obtained." }),
@@ -681,7 +681,7 @@ export default function autoresearchExtension(pi: ExtensionAPI): void {
   });
 
   pi.registerTool({
-    name: "log_experiment",
+    name: "vkf_log_experiment",
     label: "Log experiment",
     description:
       "Record an experiment's result. Appends to the session log AND writes an experiment card back to the VKF memory (a win OR a loss is durable knowledge), updating the tested claim's belief and lifecycle. This write-back is what lets future runs avoid repeating work.",

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-autoresearch-vkf",
-  "version": "0.5.0",
+  "version": "0.5.2",
   "type": "module",
   "description": "Autoresearch with verifiable long-term scientific memory. A pi extension that gathers literature, stores it as VKF claims, runs experiments, and writes verified results back to a git-native knowledge bundle so future runs build on what was learned instead of rediscovering it.",
   "keywords": [

package/skills/{autoresearch-create → autoresearch-vkf}/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: autoresearch-create
+name: autoresearch-vkf
 description: Run an autoresearch loop with verifiable long-term memory. Use when asked to optimize/improve a measurable target (test speed, bundle size, model loss, build time, Lighthouse score, …) by drawing on the research literature and remembering what was learned across runs. Orchestrates init → gather literature → extract & verify claims → recall → experiment → write results back to VKF memory → report.
 ---
 
@@ -21,8 +21,8 @@ You are the spine. Delegate the specialized work to the sub-skills below.
 - `score_ideas` — rank untested ideas by priority (EV × feasibility × evidence × novelty × info_gain ÷ cost).
 - `find_contradictions` — mine memory for tensions that seed novel hypotheses.
 - `find_transfers` — cross-domain mechanism search for surprising analogies.
-- `run_experiment` — run the measurement command, capture `METRIC name=value`.
-- `log_experiment` — record a result and write it back to memory (updates belief & lifecycle).
+- `vkf_run_experiment` — run the measurement command, capture `METRIC name=value`.
+- `vkf_log_experiment` — record a result and write it back to memory (updates belief & lifecycle).
 - `research_status` — show session + memory state.
 
 ## The two layers
@@ -51,23 +51,23 @@ transaction record — promotion is an explicit, audited step.
    gathering anything. If prior runs already learned something, build on it and
    skip rediscovery.
 
-4. **Gather literature** → use the **knowledge-gather** skill to find candidate
+4. **Gather literature** → use the **autoresearch-vkf-knowledge-gather** skill to find candidate
    techniques (via `WebSearch`/`WebFetch` against free databases — arXiv, Semantic
-   Scholar, OpenAlex), then **claim-extract** to turn them into structured claims
-   via `remember_claim`. Then **claim-verify** to check citations and codebase fit.
+   Scholar, OpenAlex), then **autoresearch-vkf-claim-extract** to turn them into structured claims
+   via `remember_claim`. Then **autoresearch-vkf-claim-verify** to check citations and codebase fit.
 
 4b. **Synthesize new ideas** (optional but high-value) → mine memory for novelty
-   instead of only retrieving it: **contradiction-miner** (tensions →
-   hypotheses), **cross-domain-transfer** (import a mechanism from another field).
-   When many ideas compete for budget, run the **idea-tournament** skill to pick
+   instead of only retrieving it: **autoresearch-vkf-contradiction-miner** (tensions →
+   hypotheses), **autoresearch-vkf-cross-domain-transfer** (import a mechanism from another field).
+   When many ideas compete for budget, run the **autoresearch-vkf-idea-tournament** skill to pick
    the 2–3 worth testing.
 
-5. **Loop** → use the **hypothesis-loop** skill: `recall_memory` → pick the
+5. **Loop** → use the **autoresearch-vkf-hypothesis-loop** skill: `recall_memory` → pick the
    highest-value, sufficiently-novel idea → implement the smallest falsifying
-   change → `run_experiment` → `log_experiment` → repeat. Keep wins, revert
+   change → `vkf_run_experiment` → `vkf_log_experiment` → repeat. Keep wins, revert
    regressions; either way the result is now in memory.
 
-6. **Report** → use the **research-report** skill to produce the lineage report
+6. **Report** → use the **autoresearch-vkf-research-report** skill to produce the lineage report
    (paper → claim → hypothesis → patch → metric Δ → status → memory update).
 
 Keep `.autoresearch-vkf/session/prompt.md` current so a fresh agent can continue. The loop is

package/skills/{claim-extract → autoresearch-vkf-claim-extract}/SKILL.md

@@ -1,6 +1,6 @@
 ---
-name: claim-extract
-description: Convert gathered literature into structured, reusable VKF claim cards (research atoms). Use after knowledge-gather to stage candidate claims in memory with remember_claim. Turns noisy papers into small, checkable, reusable assertions.
+name: autoresearch-vkf-claim-extract
+description: Convert gathered literature into structured, reusable VKF claim cards (research atoms). Use after autoresearch-vkf-knowledge-gather to stage candidate claims in memory with remember_claim. Turns noisy papers into small, checkable, reusable assertions.
 ---
 
 # Extract claims from literature
@@ -19,7 +19,7 @@ Call `remember_claim` with:
   "Replacing static gradient clipping with EMA-based adaptive clipping lowers
   early-training validation loss for small transformers."
 - **mechanism** — *why* it should work. This is the most valuable field: it's
-  what later lets the hypothesis-loop transfer the idea across domains.
+  what later lets the autoresearch-vkf-hypothesis-loop transfer the idea across domains.
 - **context** — where it applies (architecture, scale, dataset regime).
 - **implementation_recipe** — concretely how to apply it in this codebase.
 - **failure_modes** — known/suspected ways it breaks or interacts badly.
@@ -41,4 +41,4 @@ Call `remember_claim` with:
   theoretical, or anecdotal in the confidence/reliability you assign.
 
 Everything you stage here is a **candidate** (status `draft`) with a transaction
-record — nothing is trusted yet. Hand off to **claim-verify**.
+record — nothing is trusted yet. Hand off to **autoresearch-vkf-claim-verify**.

package/skills/{claim-verify → autoresearch-vkf-claim-verify}/SKILL.md

@@ -1,6 +1,6 @@
 ---
-name: claim-verify
-description: Verify staged candidate claims before the loop builds on them — check that the cited source really says it, classify the evidence, and confirm codebase relevance. Use after claim-extract to promote or downgrade claims with verify_claim. This is the trust layer that prevents memory poisoning.
+name: autoresearch-vkf-claim-verify
+description: Verify staged candidate claims before the loop builds on them — check that the cited source really says it, classify the evidence, and confirm codebase relevance. Use after autoresearch-vkf-claim-extract to promote or downgrade claims with verify_claim. This is the trust layer that prevents memory poisoning.
 ---
 
 # Verify claims
@@ -32,7 +32,7 @@ Check, in order:
   `conflicts_with` the other card's id.
 - `deprecated` — true but stale/superseded/not applicable here.
 - `rejected` — misread, hallucinated, or unsupported.
-- (`locally_tested` / `replicated` are normally set by `log_experiment`, not here.)
+- (`locally_tested` / `replicated` are normally set by `vkf_log_experiment`, not here.)
 
 Always give a **reason** — it becomes part of the audit trail (a VKF
 transaction). After each call, the tool reports `vkf validate` so you can see the
@@ -46,5 +46,5 @@ bundle stays governed.
   the source.
 - A claim's truth in a paper ≠ its usefulness for our goal. Keep those separate.
 
-When the trusted set is healthy, hand back to **autoresearch-create** for the
-**hypothesis-loop**.
+When the trusted set is healthy, hand back to **autoresearch-vkf** for the
+**autoresearch-vkf-hypothesis-loop**.

package/skills/{contradiction-miner → autoresearch-vkf-contradiction-miner}/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: contradiction-miner
+name: autoresearch-vkf-contradiction-miner
 description: Generate novel hypotheses from tensions already in memory — conflicting claims, ideas that won in one place and lost in another, and different mechanisms aimed at the same goal. Use when the loop needs fresh, non-obvious ideas rather than more literature.
 ---
 
@@ -44,8 +44,8 @@ Record it with `remember_claim`, setting:
 - a `mechanism` (required — a hypothesis with no mechanism is just noise),
 - an honest `confidence` (these are speculative; start low–medium).
 
-It enters memory as a **candidate** like any other idea — then `claim-verify` and
-the `hypothesis-loop` (via `score_ideas`) decide whether it's worth testing.
+It enters memory as a **candidate** like any other idea — then `autoresearch-vkf-claim-verify` and
+the `autoresearch-vkf-hypothesis-loop` (via `score_ideas`) decide whether it's worth testing.
 
 ## Discipline
 

package/skills/{cross-domain-transfer → autoresearch-vkf-cross-domain-transfer}/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: cross-domain-transfer
+name: autoresearch-vkf-cross-domain-transfer
 description: Generate novel ideas by importing a mechanism from another field into the current problem. Use when you want surprising analogies that keyword search misses — search by mechanism, not keywords.
 ---
 
@@ -47,11 +47,11 @@ Record the best candidate with `remember_claim`:
 - `failure_modes` — note where the analogy might break (the assumptions the source
   domain has that yours doesn't).
 
-Then let `claim-verify` and `score_ideas` decide if it earns an experiment.
+Then let `autoresearch-vkf-claim-verify` and `score_ideas` decide if it earns an experiment.
 
 ## Discipline
 
 - **Require a mechanistic reason for transfer**, not just surface similarity. "Both
   use matrices" is not a transfer.
 - If you gathered claims only from your own domain, there's nothing to transfer
-  *from* — use `knowledge-gather` to pull in adjacent fields first.
+  *from* — use `autoresearch-vkf-knowledge-gather` to pull in adjacent fields first.

package/skills/{hypothesis-loop → autoresearch-vkf-hypothesis-loop}/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: hypothesis-loop
+name: autoresearch-vkf-hypothesis-loop
 description: The core experiment loop — recall memory, pick the highest-value sufficiently-novel idea, run the smallest falsifying experiment, and write the result back to memory. Use to drive iterations of an autoresearch loop after claims have been gathered and verified.
 ---
 
@@ -33,9 +33,9 @@ ideas deliberately and you never repeat settled work.
      regress), *risk* (what could break), *novelty basis* (why it's not a repeat).
 
 4. **Run the smallest falsifying experiment.** Make the minimal change in scope,
-   then `run_experiment`. Read the `METRIC` line — don't eyeball logs.
+   then `vkf_run_experiment`. Read the `METRIC` line — don't eyeball logs.
 
-5. **Judge honestly, then `log_experiment`.** Record the value, the tested
+5. **Judge honestly, then `vkf_log_experiment`.** Record the value, the tested
    `claim_id`, whether you `kept` it, conditions, and notes. The tool:
    - derives win/loss/inconclusive vs the baseline,
    - writes an **experiment card back to memory** (a loss is durable knowledge),
@@ -54,4 +54,4 @@ ideas deliberately and you never repeat settled work.
 - **One variable at a time** so the result attributes cleanly to the hypothesis.
 
 When you've made meaningful progress or exhausted promising ideas, hand to
-**research-report**.
+**autoresearch-vkf-research-report**.

package/skills/{idea-tournament → autoresearch-vkf-idea-tournament}/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: idea-tournament
+name: autoresearch-vkf-idea-tournament
 description: Run a structured multi-perspective tournament over candidate ideas to pick the 2-3 worth testing. Use when there are many candidate hypotheses competing for limited experiment budget.
 ---
 
@@ -7,7 +7,7 @@ description: Run a structured multi-perspective tournament over candidate ideas
 
 When many ideas compete for a limited experiment budget, don't just take the top
 of one ranking. Run a tournament: judge each idea from several perspectives, then
-advance only the best 2–3 to the `hypothesis-loop`.
+advance only the best 2–3 to the `autoresearch-vkf-hypothesis-loop`.
 
 ## Assemble the field
 
@@ -45,7 +45,7 @@ the numbers miss — especially the Skeptic's failure-mode and gaming checks.
   `verify_claim`) so the tournament's reasoning is remembered and they aren't
   re-litigated next round.
 
-Hand the 2–3 winners to the **hypothesis-loop**.
+Hand the 2–3 winners to the **autoresearch-vkf-hypothesis-loop**.
 
 ## Discipline
 

package/skills/{knowledge-gather → autoresearch-vkf-knowledge-gather}/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: knowledge-gather
+name: autoresearch-vkf-knowledge-gather
 description: Gather frontier knowledge relevant to a research goal — search papers, repos, docs, and benchmarks for candidate techniques. Use as the discovery step of an autoresearch loop, before extracting claims. Collects candidate knowledge; it does not invent ideas or run experiments.
 ---
 
@@ -63,5 +63,5 @@ For each candidate, capture enough to become a claim later:
 - **Look for contradictions and gaps** between sources — they're the richest
   seeds for novel hypotheses later.
 
-Hand the collected candidates to **claim-extract**, which writes them into memory
+Hand the collected candidates to **autoresearch-vkf-claim-extract**, which writes them into memory
 with `remember_claim`.

package/skills/{research-report → autoresearch-vkf-research-report}/SKILL.md

@@ -1,5 +1,5 @@
 ---
-name: research-report
+name: autoresearch-vkf-research-report
 description: Produce the autoresearch report with full idea lineage — paper → claim → hypothesis → patch → metric change → status → memory update. Use to summarize an autoresearch run into an auditable, human-readable report at .autoresearch-vkf/session/report.md.
 ---