gossipcat 0.4.25 → 0.4.27

package/docs/HANDBOOK.md

@@ -55,6 +55,27 @@ The agent prompt in item 2 must see **nothing** about the surrounding orchestrat
 
 This is a **security boundary**, not a style preference. See prior session 2026-04-08 and consensus `ff598432`.
 
+**Optional elision protocol (`prompt_format: 'elided'`).** All three dispatch tools — `gossip_dispatch`, `gossip_run`, `gossip_collect` — accept an opt-in `prompt_format` parameter. The default `'inline'` is byte-identical to the description above. When the caller passes `'elided'`:
+
+- The server writes the prompt body to `.gossip/dispatch-prompts/<taskId>.txt` (atomic temp-rename, `SAFE_NAME`-validated taskId).
+- Item 1 contains a marker line of the form `[skills section elided: see <abspath>, <N> bytes — READ this file and pass its CONTENTS verbatim as the Agent(prompt: ...) value. Do NOT pass the path string.]`.
+- **Item 2 is OMITTED entirely** — no skeleton, no placeholder. The orchestrator MUST `Read(<abspath>)` and forward the file contents verbatim to `Agent(prompt: ...)`. A missing file is a hard failure; never substitute the marker text for the prompt.
+
+The on-disk prompt file carries ONLY the agent-facing prompt (identity + instructions + skills + task). It NEVER carries `relay_token`, `task_id`, the `AGENT_PROMPT:` tag prefix, or any other orchestration metadata — those stay in Item 1 only. Persistence: the absolute path is stored alongside the task in `.gossip/native-tasks.json` so `/mcp` reconnect can prune orphan files for tasks that no longer exist. Eviction is mtime-based (default 1h) plus aggregate eldest-eviction at 100 MB.
+
+Example flow inside the consensus protocol when `prompt_format: 'elided'` is passed to `gossip_collect`:
+
+```
+1. gossip_collect(task_ids: [...], consensus: true, prompt_format: 'elided')
+2. → server writes .gossip/dispatch-prompts/<consensusId>__<agentId>.txt per cross-review participant
+3. → returns ⚠️ EXECUTE NOW payload with one marker line per agent; PROMPTS section is ABSENT
+4. → orchestrator Reads each cited file and passes contents verbatim to Agent(prompt: ...)
+5. → gossip_relay_cross_review(consensus_id, agent_id, result) per agent
+6. → gossip_collect(consensus: true) for final synthesized output
+```
+
+Spec: `docs/specs/2026-05-18-native-dispatch-skill-handle-pattern.md` (Option B — server-side prompt elision).
+
 ### 5. Scoped agents write files, orchestrator commits
 
 Scoped write mode (`write_mode: "scoped"`) lets agents write files within a directory scope but **not** run `git commit` or `shell_exec` (except read-only git commands). The orchestrator validates the agent's output and commits on their behalf. Worktree agents have full git access within their isolated branch.
@@ -207,6 +228,12 @@ If `format_compliance` signals a sudden drop for an agent that was fine last rou
 
 When blocked, the error message shows age + remaining cooldown + override instruction. Pass `force: true` to bypass; every override is appended to `.gossip/forced-skill-develops.jsonl` for auditability. Chronic override patterns on an agent+category pair are a signal that the skill prompt is ineffective or `MIN_EVIDENCE` is miscalibrated for that category — investigate before reflexively forcing.
 
+### Tech-stack override
+
+Drop a `.gossip/tech-stack.md` at the project root to bypass auto-detection on `gossip_skills(action: "develop")`. Content (max 2000 chars after trim) is injected verbatim into the skill-develop prompt's `<tech_stack>` block, replacing the auto-detected description. Useful for non-Node host projects (Solidity, Rust, Move, audit workspaces) where the LLM hallucinates a Node.js stack from thin npm dep signal (issue #410, PR #411 floor + this override). Cache is session-stable — restart the MCP server to pick up edits. Empty file or read errors fall through to auto-detect (with stderr warning on errors). Files over 2 KB are clamped with a stderr warning.
+
+**Tech-stack auto-detection.** When no override is present and the npm dep count is below `TECH_STACK_MIN_DEPS=3` OR the project is non-Node, `detectTechStack` scans the project root for known manifests (Cargo.toml, pyproject.toml, requirements.txt, go.mod, foundry.toml, Move.toml, Gemfile, composer.json), the README first 30 lines / 2 KB, and a shallow file-extension census (root only, excluding `node_modules`/`.git`/`.gossip`/`dist`/`build`/`out`/`coverage`, capped at 10 extension types; config/docs extensions such as `.json`, `.md`, `.yaml`, `.toml`, `.lock` are excluded from the census since they are already captured by other signals). Any non-Node signal — manifest match, README content, or extension census — bypasses the `MIN_DEPS=3` floor so polyglot projects don't need the override file. Workspace-level manifests (e.g., `packages/contracts/foundry.toml`) are NOT scanned in this MVP; place the manifest at root or use `.gossip/tech-stack.md` for those cases.
+
 ### Verifying UNVERIFIED findings
 
 When a consensus report has `UNVERIFIED` findings (cross-reviewer couldn't check), **you must verify them yourself before presenting results**. UNVERIFIED means "the peer didn't have the tools or context to check" — you do. Read the cited files, grep for the identifiers, confirm or reject. Do not show raw consensus output with unexamined UNVERIFIED findings.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "gossipcat",
-  "version": "0.4.25",
+  "version": "0.4.27",
   "description": "Multi-agent orchestration for Claude Code — parallel review, consensus, adaptive dispatch",
   "mcpName": "io.github.ataberk-xyz/gossipcat",
   "repository": {