@witness-ai/opencode 0.7.5 → 0.7.7

package/package.json

@@ -1,7 +1,7 @@
 {
   "name": "@witness-ai/opencode",
   "mcpName": "io.github.IngTian/witness",
-  "version": "0.7.5",
+  "version": "0.7.7",
   "description": "MCP memory / second brain for Claude Code & OpenCode — captures your coding sessions and distills how your knowledge and habits evolve over time, queryable by your agent.",
   "type": "module",
   "main": "./witness.js",
@@ -21,10 +21,10 @@
     "README.md"
   ],
   "optionalDependencies": {
-    "@witness-ai/opencode-darwin-arm64": "0.7.5",
-    "@witness-ai/opencode-linux-x64": "0.7.5",
-    "@witness-ai/opencode-win32-x64": "0.7.5",
-    "@witness-ai/opencode-win32-arm64": "0.7.5"
+    "@witness-ai/opencode-darwin-arm64": "0.7.7",
+    "@witness-ai/opencode-linux-x64": "0.7.7",
+    "@witness-ai/opencode-win32-x64": "0.7.7",
+    "@witness-ai/opencode-win32-arm64": "0.7.7"
   },
   "keywords": [
     "mcp",

package/prompts/lens/corpus-example/README.md

@@ -0,0 +1,77 @@
+# corpus-example — a lens for a NON-person corpus
+
+Every other lens witness ships watches a *person*. This one watches a **subject that is not you**:
+market commentary, tracking how the regime state changes over time.
+
+Copy it as the starting point for any document corpus — research notes, a knowledge base, incident
+reports, meeting logs. The shape transfers; only the dimensions and the prose change.
+
+```sh
+cp -R "$CLAUDE_PLUGIN_ROOT/prompts/lens/corpus-example" ./regime   # edit for your domain
+witness lens register regime ./regime
+witness lens enable regime
+```
+
+If your archive is *only* a corpus (no person tracking), start it with no person lens at all:
+
+```sh
+WITNESS_NO_DEFAULT_LENS=1 witness doctor    # first open decides; the built-in lens is never seeded
+```
+
+Then feed it records and read the result:
+
+```sh
+witness ingest --file news.ndjson     # {"text":"…","id":"n1","session":"2026-01","ts":"2026-01-05"}
+witness facets regime                 # the current state, with confidence + provenance
+witness profile regime                # the narrative brief
+```
+
+## What makes this lens different from the person lenses
+
+**`extract.md` names the subject explicitly** — "notice claims about *the market regime* … not about
+any person, author, or writing style." Without that, a model reading market prose will happily start
+characterizing the *analyst*, because that is the more natural reading of "notice things about the
+subject of this text."
+
+**The dimensions are states, not traits.** `rates`, `inflation`, `growth`, `risk_appetite`,
+`positioning` are things that have a *current value which changes*. That is what makes the
+bi-temporal layer earn its keep: when the regime flips, the old value closes with a date and the new
+one opens, so you get a history rather than a replacement.
+
+**`review.md` is explicit about what counts as a change.** The instruction is to set
+`contradicts_prior` only for a *sustained* change of state, not a single noisy print — and that
+absence of evidence is not contradiction. Getting this wrong in either direction is the main failure
+mode: too eager and every data release rewrites history, too shy and a real regime change never
+records.
+
+**`key` must be a stable slug.** `policy_path` reviewed in January and again in June is one facet with
+two dated versions. A key that drifts (`policy_path_jan`, `jan_policy`) produces two unrelated facets
+and no history at all.
+
+## Verified working
+
+This lens was run end to end before shipping: 7 market-news records → 22 observations → 13 facets,
+in an archive with no person lens enabled. Then a hawkish→dovish regime flip closed **five** facets
+and opened their replacements, e.g.:
+
+```
+inflation/core_trend
+  "Disinflation has broken rather than paused: core CPI reaccelerating at 3.8%…"  valid_to 2026-06-…
+  "Core is disinflating persistently rather than reaccelerating — 2.1% y/y…"      (current)
+```
+
+The generated brief also flagged that its own earlier sequencing thesis had been falsified by the new
+data — which is the point of keeping the history instead of the latest snapshot.
+
+## Also worth knowing
+
+The **summary prompt** is person-shaped by default too. If the narrative reads like it is describing a
+personality rather than your subject, override it — one file, no registration:
+
+```sh
+mkdir -p "$(witness doctor | awk '/data root/{print $3}')/summarize"
+$EDITOR "$(witness doctor | awk '/data root/{print $3}')/summarize/unified.md"
+```
+
+An archive with no override keeps receiving improved defaults on upgrade; once you override, your file
+is never touched.

package/prompts/lens/corpus-example/extract.md

@@ -0,0 +1,17 @@
+You are reading one batch of market commentary. Notice claims about **the market regime** — what
+the prevailing state is and what is changing — not about any person, author, or writing style.
+
+Tag each observation with exactly one dimension:
+- rates — the path and expectations for policy rates
+- inflation — level, direction, and composition of price pressure
+- growth — activity, labour, demand
+- risk_appetite — how willing capital is to take risk
+- positioning — how participants are actually positioned vs. what they say
+
+Anchor every observation in the text. Prefer a claim that could later be shown WRONG over a vague
+one. Note the direction of change when the text implies one.
+
+Return ONLY a JSON array. Each element:
+[{ "dimension": "rates", "observation": "…", "evidence": "…", "poignancy": 6 }]
+
+poignancy 1-10: how much this observation would change a reader's model of the regime.

package/prompts/lens/corpus-example/lens.json

@@ -0,0 +1,4 @@
+{
+  "name": "regime",
+  "dimensions": ["rates", "inflation", "growth", "risk_appetite", "positioning"]
+}

package/prompts/lens/corpus-example/review.md

@@ -0,0 +1,20 @@
+You are maintaining a REGIME MODEL from accumulated observations about market commentary.
+
+You are given observations and the current facets (each an attribute with a confidence 0-1). Decide
+what the regime state is NOW, and mark a facet as contradicting its prior value ONLY when the
+observations show a sustained change of state — not a single noisy print.
+
+Return ONLY a JSON array:
+[{
+  "dimension": "rates",
+  "key": "policy_path",
+  "value": "…the current state, one sentence…",
+  "confidence": 0.7,
+  "because_of": ["obs_id", "…"],
+  "contradicts_prior": false
+}]
+
+Rules:
+- `key` is a stable slug for the thing being tracked, so the same key across reviews forms a history.
+- Set contradicts_prior=true only for a real regime change; that is what writes a dated change record.
+- Absence of evidence is not contradiction — say nothing rather than assert a reversal.

package/witness.js

@@ -35,7 +35,7 @@ function warnIfInert() {
 function spawnWitness(args, payload) {
   if (!WITNESS_BIN || process.env.WITNESS_WORKER === "1") return
   try {
-    const env = { ...process.env, WITNESS_OPENCODE_PLUGIN: "1" }
+    const env = { ...process.env }
     env.WITNESS_ASSETS ||= modelDir(PACKAGE_ROOT)
     // Point the binary at THIS package's bundled prompts. The binary lives in a
     // separate per-platform package, so its exe-relative probe can't find them —