You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Ollama structured-output ingestion. The Ollama JSON Schema dialect now shares the grammar subset with Anthropic, rewriting the single-element allOf-with-$ref and oneOf enum forms schemars emits into shapes llama.cpp's grammar builder compiles. The extraction, triage, and relation stages no longer dead-letter with a parse error against local models. (#195)
Pipeline responses wrapped in a Markdown code fence now parse. The extraction, triage, and relation stages deserialise strictly first and retry once with the fence stripped only when that fails, so a provider that ignores the structured-output schema and fences its JSON no longer dead-letters, while a valid payload, including one whose strings contain backticks, is never altered. (#195)
Strict-harness interop for the MCP tools. tribal_discover now treats an empty pagination cursor as the first page rather than rejecting it, and tool error results no longer carry structured content, so a harness that validates structured content against a tool's success output schema surfaces the real error message instead of a generic schema mismatch. (#200)
Install tribal 0.3.1
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/tribal-memory/tribal/releases/download/v0.3.1/tribal-installer.sh | sh
This release makes the embedding geometry configurable and adds zero-downtime reindexing of the embedding space. It is a breaking release: the initial database schema is revised for the embedding-profile model and the embedding configuration keys have moved, so there is no in-place upgrade from 0.2.x. Provision a fresh database and update the configuration (see Changed).
Added
Configurable embedding geometry. The embedding model, output dimension, and provider endpoint are config-driven rather than fixed at 768 dimensions. Each activation is recorded in an append-only log of embedding profiles; the active profile is derived from that log and is the live identity for every read and write. Embeddings are stored as halfvec, and a database trigger rejects any vector whose dimension does not match its profile.
tribal reindex, a zero-downtime migration of the embedding space to a new model, dimension, or endpoint. The run is a background, single-flight, crash-safe catch-up that embeds the corpus into a new profile while reads and writes continue against the active one, then cuts over atomically; an unchanged target is a no-op. The CLI exposes run (--provider and --model, optional --dimensions and --base-url, with --dry-run to estimate the item and tag counts before spending), cancel, and prune (supersede the non-active and failed profiles and reclaim their storage). tribal check now reports the active embedding profile.
The same three reindex operations as operator MCP tools (tribal_reindex, tribal_reindex_cancel, tribal_reindex_prune), so an authorised agent can drive a migration without shell access.
A narrow tribal.embedding:execute OAuth scope that gates the reindex tools, and a repeatable --scope flag on tribal token create to mint scoped tokens. Read and write scopes plus tribal.embedding:execute are mintable; root and uncatalogued execute scopes are refused. The local stdio principal is granted the scope automatically, so bootstrap and the tribal token create defaults are unchanged (full read and write).
Changed
The embedding configuration has moved. The flat [embedding] section is replaced by init.embedding, a genesis seed applied only when a corpus is first created (once the corpus exists the active profile is the live identity, so later edits to init are inert and tribal check reports any divergence as informational state), and by a credentials catalogue that binds a named (provider, base_url) endpoint to an API key so a migrated corpus keeps its key reachable. Environment overrides are renamed to match: TRIBAL_EMBEDDING__* becomes TRIBAL_INIT__EMBEDDING__*, and a catalogue key is set with TRIBAL_CREDENTIALS__<NAME>__API_KEY. The Docker Compose and .env.example samples are updated accordingly.
tribal_discover results now carry embedding_profile_id, the active profile that produced them, and tribal_feedback accepts it so the local retrieval-feedback log records the producing profile. A discover pagination cursor is bound to the embedding profile that issued it.
The worker pipeline prompts are reframed around tacit knowledge (the reasoning, the rejected alternatives, and the bounding constraints behind a decision) rather than a generic knowledge base, and the model-facing vocabulary is unified on "claim". The few-shot examples and the structured-output guards are unchanged.
The Docker Compose Postgres image is pinned to pgvector/pgvector:0.8.2-pg17, the minimum that provides the halfvec operations the embedding store now relies on.
Install tribal 0.3.0
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/tribal-memory/tribal/releases/download/v0.3.0/tribal-installer.sh | sh
OAuth 2.1 authentication for the HTTP and SSE transports, in a new tribal-auth crate. Tribal now runs as an OAuth authorisation server: Dynamic Client Registration (RFC 7591), PKCE, an authorisation-code flow with an explicit consent step, the RFC 8414 and RFC 9728 discovery metadata endpoints, and audience-bound bearer tokens. An OAuth-capable harness registers and authenticates itself on first connect, so the loopback wire-up carries no token to copy.
Changed
tribal bootstrap and tribal mcp-config choose the wire-up shape from the deployment's onboarding mode rather than always emitting a token. A loopback deployment with dynamic registration enabled advertises a URL-only OAuth snippet with nothing to copy; every other surface (reachable beyond loopback, or with registration disabled) embeds the persisted static token. Pass --static-token to force that token for a harness that authenticates with a bearer header only.
tribal check's token check follows the same onboarding mode: it skips on a URL-only surface, where clients authenticate over OAuth, and fails when a surface that depends on a static token has none.
A missing bearer token on a network request now logs at DEBUG rather than WARN, so a steady-state healthcheck cycle no longer emits misleading authentication warnings.
Security
The unauthenticated Dynamic Client Registration endpoint is refused whenever the OAuth surface is reachable beyond loopback. With no explicit advertised URL, a wildcard bind (0.0.0.0 or [::]) is treated as routable and fails closed; a loopback server.public_mcp_url is the trusted-exposure override for the container host-port-mapping shape. server.public_mcp_url is validated at load as an http(s) endpoint with a host and no fragment, and the same check guards the non-validating tribal mcp-config renderer.
Install tribal 0.2.5
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/tribal-memory/tribal/releases/download/v0.2.5/tribal-installer.sh | sh
Pipeline response schemas are now sent in each provider's accepted JSON Schema dialect with hard schema enforcement. OpenAI requests carry the strict subset (closed all-required objects, nullable optionals, internally-tagged enums emitted as anyOf of closed branches) with strict: true; Anthropic requests rewrite oneOf to anyOf and close every object so its grammar enforcement applies. The previous advisory mode allowed weaker cloud models, including the recommended gpt-5.4-mini, to return malformed shapes that dead-lettered ingest jobs after exhausting retries. Structured output is now guaranteed by the provider rather than hoped for.
Changed
Pipeline response-schema field descriptions are tightened across the extraction, triage, and relation stages to give the model a clearer instruction surface.
MCP tool descriptions are rewritten to remove implementation jargon, presenting each tool through its user-facing contract rather than its internal mechanism.
The triage stage's relation-direction explanation in the prompt is aligned with the legend, with explicit source/target binding, so the model classifies similar-item relations more consistently.
Install tribal 0.2.4
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/tribal-memory/tribal/releases/download/v0.2.4/tribal-installer.sh | sh
Triage now references similar items by position index rather than by their identifier, resolving them server-side. This removes a failure mode in which weaker models that could not reproduce a knowledge-item identifier failed triage and exhausted retries. An unresolvable reference is handled gracefully: an out-of-range duplicate match is treated as a novel item rather than failing the job.
Install tribal 0.2.3
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/tribal-memory/tribal/releases/download/v0.2.3/tribal-installer.sh | sh
Support for reasoning and adaptive-sampling models. Tribal resolves, per provider and model, which request fields a target accepts and shapes each request to match: OpenAI reasoning models (the o-series and the GPT-5 reasoning line) send the output cap as max_completion_tokens and have caller sampling parameters dropped, while Anthropic adaptive models (Opus 4.7) have temperature dropped but keep their required max_tokens. Both tribal check --providers and ingest now succeed for these models. An unrecognised model keeps sending every parameter, so ordinary models are unaffected.
Per-stage temperature and max_tokens from configuration now reach the model. Previously they were silently dropped and only recorded in the system fingerprint.
Configuration validation for sampling parameters: when set, temperature must be within [0.0, 2.0] and max_tokens at least 1. Model IDs may no longer be empty or whitespace-only.
Changed
Sampling parameters are now optional in configuration. An unset temperature or max_tokens means "use the provider default" rather than a built-in number, and the four built-in per-stage sampling defaults have been removed.
The system fingerprint records the effective post-reconcile sampling parameters, so it reflects what is actually sent. Fingerprints change for any model whose request shape the capability layer adjusts.
Rewrote the README around the shipped onboarding surface (tribal bootstrap, tribal check, tribal mcp-config, Docker Compose, and the skills), and added CONTRIBUTING.md.
Fixed
Reasoning models are no longer rejected by the provider readiness probe (tribal check --providers), which previously forced temperature and max_tokens values these models reject.
Install tribal 0.2.2
Install prebuilt binaries via shell script
curl --proto '=https' --tlsv1.2 -LsSf https://github.com/tribal-memory/tribal/releases/download/v0.2.2/tribal-installer.sh | sh
