- Use is_some_and() instead of map_or(false, ...) for Option check
- Use inline format args with {previous_id} in format! macro

Kimi-K2.7 is the current frontier Kimi release. Add it as the newest-first
candidate in PROVIDER_DEFAULT_MODELS for the kimi-for-coding provider,
mirroring how k2.6/k2.5 are already handled (resolved by the provider at
runtime via the stable kimi-for-coding virtual id).

The legacy master branch was deleted (2026-06-28); main is now the sole
long-lived branch. Update the branching section to reflect that master no
longer exists and must not be recreated or pushed to.

…aphim-grep

KG-driven search configuration specific to pi_agent_rust so terraphim-grep
works offline (no server) with project-aware concept boosting.

- .terraphim/kg/: 13 project concepts (provider, session, extension, tool,
  sse, model-registry, acp, interactive-tui, conformance, hashline-edit,
  auth, asupersync, pi-terraphim-router), each with description + synonyms::
- .terraphim/thesaurus.json: 180 compiled entries (synonym -> {id, nterm})
- .terraphim/config.json: Pi Agent Rust Engineer role -> project KG + Ripgrep
  haystack

terraphim-grep loads the thesaurus (180 entries), fff-search scans the repo,
and matched concepts (e.g. 'extension','hostcall') boost relevant chunks
above the baseline score. Requires terraphim-grep built with the code-search
feature (fff-search); without it search_code is compiled out.

Structurally-derived identifiers (pub struct/enum/trait + impl Provider)
mined via ast-grep now back each concept, so terraphim-grep matches real
code tokens (ReadTool, AnthropicProvider, ModelRegistry, AuthStorage, ...)
instead of prose only. KG-boosted chunks now rank the actual definition
files (e.g. providers/anthropic.rs, tools.rs) above baseline.

- kg/*.md: appended curated, name-selective code anchors per concept
  (provider/session/extension/model-registry/sse/acp/auth/interactive-tui/
  tool/hashline-edit). Thesaurus grew from 180 -> 409 entries.
- scripts/refresh-anchors.sh: repeatable, idempotent refresh — re-runs the
  ast-grep extraction, appends new anchors, regenerates thesaurus.json.
  Run it whenever the code changes to keep the KG aligned.

Anchors are selective (name patterns, not file-wide dumps) to keep the
thesaurus signal-rich. Idempotent: a second run appends 0.

…tch zai

zai-coding-plan routing defaults were stuck at context_window=128_000,
max_tokens=16_384 -- stale values that silently capped every response
from models like glm-5.2 at a fraction of their real capacity. The
sibling zai provider already registers glm-5.2 with context_window=
1_000_000, max_tokens=131_072 (same underlying model, same real API
limits). Aligned zai-coding-plan to context_window=262_144 (matching
zai/glm-5.1, a safe uniform default for the provider) and max_tokens=
131_072.

Verified empirically: a long-generation stress test that previously
hit stopReason=length at exactly 4096 tokens on every attempt now
completes with zero length-truncations across 500+ turns.

Also required rebuilding and redeploying the local pi-rust binary,
which was 5 weeks stale (v0.1.16 vs source v0.1.20) and missing an
already-fixed max_tokens registry-seeding bug from an earlier commit.

…o MessageEnd events (terraphim-agents#80, D4)

Companion to terraphim-agents PR #106 (merged 2026-08-14 as f46c9c4). Adds EXIT_CODE_EMPTY_OUTPUT = 4 + cooperative upstream signal. Round-1 author self-review (3/5); round-2 independent review (5/5) confirmed correct race handling and retry semantics. P1 follow-up for orchestrator-side recognition of exit_code == 4 filed as separate terraphim-agents task.