Configuration

Where these live. .tl-telar-thresholds.json sits at the project root and is committed. .tl-telar/external-tools.yaml lives under the git-ignored .tl-telar/ state directory. Missing files fall back to safe defaults — nothing hard-fails on absence.

Session model & cost. /tl-telar:orchestrate is a conductor that runs in your main session. Spawned agents carry their own model tier (frontmatter model:) and reviewers are pinned to Opus, but the orchestrator's own turns use whatever model /model is set to — agent frontmatter cannot override the session model. Running a long orchestrate on the most expensive model is where most of the cost comes from. Recommendation: run orchestrate on Sonnet or Opus, not the top-tier model.

`.tl-telar-thresholds.json`

The quality-gate contract. Framework-aware defaults are written at setup based on the detected stack (React Native / Expo / Flutter).

coverage

Key	Default	Meaning
`lines`	80	Minimum line-coverage %.
`branches`	75	Minimum branch-coverage %.
`functions`	80	Minimum function-coverage %.
`statements`	80	Minimum statement-coverage %.

performance

Key	Default	Meaning
`min_fps`	60	Minimum acceptable frame rate.
`max_cold_start_ms`	3000	Maximum cold-start time in milliseconds.

size

Key	Default	Meaning
`max_apk_mb`	50	Maximum Android APK size (MB).
`max_ipa_mb`	60	Maximum iOS IPA size (MB).

accessibility

Key	Default	Meaning
`required_audit_pass`	false	When `true`, an accessibility audit must pass before a task completes.

autonomy

Key	Default	Meaning
`cycle`	`"interactive"`	`interactive`: the orchestrator may pause at `checkpoint: true` WUs and at self-reflect. `unattended`: a single human gate at plan-readiness — scope + approved UI drafts + all inputs collected up front — then the WU cycle runs to PR-ready with zero pauses. Unattended never skips a decision; it makes it earlier.

execution

Key	Default	Meaning
`max_parallel_wus`	3	How many Work Units run concurrently. A WU is dispatched only when its `deps` are COMPLETE and its `file_scope` is free (or, under worktree isolation, in its own checkout). Parallel agent runs cost roughly 15× chat tokens — keep it modest. Absent/invalid → 3.

enforcement

What blocks, and the commands the gates run. *_command pairs with a *_strict flag — advisory (log only) when false, blocking when true.

Key	Default	Meaning
`coverage_command`	framework-detected	e.g. `npx jest --coverage`, `flutter test --coverage`.
`coverage_strict`	true when a runner is detected	Block on coverage below threshold.
`perf_command`	`bash scripts/perf-smoke.sh`	Performance smoke command (advisory stub by default).
`perf_strict`	false	Block on performance regression.
`size_command`	`bash scripts/size-check.sh`	APK/IPA size check (advisory stub by default).
`size_strict`	false	Block on size over budget.
`a11y_command`	echo stub	Accessibility audit command.
`a11y_strict`	false	Block on accessibility failures.
`blockPRCreation`	true	Fail gates prevent PR creation.
`blockTaskCompletion`	true	Fail gates prevent marking a task complete.
`self_reflect_per_wu`	false	When `true`, fire `/tl-telar:self-reflect` after every WU (adds a user-approval step per WU) instead of once pre-PR.
`detected_framework`	at setup	Recorded framework (react-native / expo / flutter / node).

`.tl-telar/external-tools.yaml`

External AI adapters (Codex/GPT-5.6-sol, Gemini, Kimi K3 via a native kimi CLI or a generic Anthropic-compatible compat adapter), the config-driven model roster (routing.roles + models_registry), cross-model review, budget, and opt-in Claude Code native features. Git-ignored; created only by /tl-telar:setup-orchestration. Adapters are disabled by default. Onboarding a new model is a YAML edit — no code change; resolve any role with tl-telar-external-tools.sh resolve-role <role>.

Codex plugin vs Codex adapter. Installing tl-telar@telar with codex plugin marketplace add zekiyugnak/telar-framework --ref develop loads Telar into Codex. Enabling adapters.codex here is different: it lets Telar delegate selected orchestration work to the Codex CLI, and remains opt-in.

adapters

Key	Type	Meaning
`enabled`	bool	Activate the adapter. The CLI must be installed and `auth_env_var` set first.
`type`	string	Omit for a bespoke `<name>.sh` adapter (codex/gemini/kimi). `"compat"` → the generic `compat.sh` that drives `claude` against any Anthropic-compatible endpoint.
`model`	string	Blank = adapter's own default (Codex: `~/.codex/config.toml`). e.g. `"gpt-5.6-sol"`, `"kimi-code/k3"`, `"pro"`.
`reasoning_effort`	string (codex)	gpt-5.6-sol ladder: `low｜medium｜high｜xhigh｜max｜ultra`. Sol's default is `low` — set it explicitly. (gpt-5.6-sol needs codex-cli ≥ 0.144.)
`base_url` / `api_style`	string (compat)	Compat adapter endpoint, e.g. `https://api.kimi.com/coding/`; only `api_style: anthropic` is supported.
`pricing`	obj	`{input_per_1m, output_per_1m}` USD; model-agnostic per adapter. Omitting it is fail-closed (conservative per-task cap recorded, never a silent $0). Set `0` for a flat-subscription model.
`timeout_seconds`	number	Adapter invocation timeout.
`auth_env_var`	string	Env var the dispatcher reads for auth (e.g. `OPENAI_API_KEY`, `GEMINI_API_KEY`, `KIMI_API_KEY`). Native `kimi`/Codex use their own CLI login instead.
`sandbox`	string	`"none"` — no sandbox isolation.

routing & budget

Key	Default	Meaning
`routing.roles`	obj	Model per pipeline role (`architect｜moderator｜developer｜reviewer｜tester`); each role takes `model`/`models`, optional `escalation`, `panel`, `options`. Single source of truth for the roster; resolved via `resolve-role`.
`routing.models_registry`	obj	How each model runs: `exec: claude` (native tier) or `exec: adapter` (`tool`), plus `tier`/`effort`. Onboard a model = one entry here + reference it from a role.
`routing.default_implementer`	`"cheapest-available"`	Also `round-robin`, `claude`, or a specific tool name.
`routing.escalation_order`	`[codex, gemini, claude]`	`--tool auto` tries adapters in this order (first enabled+healthy wins).
`routing.fallback_by_health`	`true`	A role's model whose adapter is unhealthy falls to its `escalation`.
`routing.health_recheck_every_n_tasks`	10	Re-run health checks periodically.
`budget.per_task_usd`	1.00	Max USD per adapter invocation; over → fall back to Claude.
`budget.per_session_usd`	10.00	Session-wide adapter spend cap.
`budget.ledger_file`	JSONL path	Records every invocation for audit + `budget-status`.
`budget.circuit_breaker_message`	string	Printed when the budget trips.

cross_model_review (Phase γ)

An additive second review: Review 1 (the Claude roster) always runs; when enabled, Review 2 runs on top with a model distinct from both the writer and Claude. Both must PASS.

Key	Default	Meaning
`enabled`	false	`true` → Review 2 is required on every WU.
`on_unavailable`	`"block"`	`block`: stop if no distinct second model is ready. `warn_and_proceed`: skip Review 2 on Review 1 alone, logged loudly.
`matrix`	writer→reviewers	Which models may review each writer's diff. A model never reviews its own code.

cc_features — Claude Code native capabilities

Enabled is intent, not capability. A feature activates only when enabled: true and a runtime probe confirms the capability is really present. Otherwise Telar falls back to the current-behavior path (fail-closed), logs one line, and never hard-fails. Defaults are true because absence degrades safely — older Claude Code runs unchanged.

The gating is one tested code path — scripts/tl-telar-cc-features.sh — that the plan-review gate and the orchestrator call, rather than re-deciding it in prose.

cc_features:
  dynamic_workflows:
    enabled: true
    on_unavailable: "warn_and_proceed"
  worktree_isolation:
    enabled: true
    on_unavailable: "warn_and_proceed"

Key	Default	Meaning
`dynamic_workflows.enabled`	true	Run the plan-review gate as a deterministic Workflow script (`parallel()` fan-out + schema-validated verdicts) when the `Workflow` tool is available. Both paths emit the identical aggregated verdict, so downstream is oblivious to which ran.
`worktree_isolation.enabled`	true	Run each parallel WU in its own git worktree so WUs with overlapping `file_scope` run concurrently (relaxes the disjoint constraint). Fallback: disjoint-scope serialization — slower, never wrong.
`<feature>.on_unavailable`	`"warn_and_proceed"`	When the capability isn't confirmed: `warn_and_proceed` runs the fallback (logged); `block` stops at preflight (use when the feature is a hard requirement).

Capability probe signals. dynamic_workflows → the Workflow tool is present in the top-level session. worktree_isolation → isolation: worktree support is positively confirmed; because older Claude Code silently ignores that frontmatter, an unconfirmed probe keeps the disjoint constraint even when enabled: true — trusting the probe, never the flag.

Values you can't configure

Fixed in source, not tunable via either file:

Aspect	Fixed value
Plan-review reviewers	3 (parallel)
Design-review reviewers	6 (PM, Architect, Designer, Security-Design, CTO, Mobile-Platform)
Per-WU adversarial reviewers (always-on)	2 (Code + Mobile Security)
Per-WU adversarial reviewers (conditional)	+2 (A11y, Performance — by file_scope)
Max gate retries	3, then human escalation
Commit & push	Never automatic — the orchestrator emits COMMIT-READY and leaves it to you

Auto-managed state (don't hand-edit)

Written and maintained by the orchestrator under .tl-telar/; git-ignored working files, not durable artifacts:

File	Role
`.tl-telar/plans/active-plan.md`	Plan + WU list; the resume sentinel.
`.tl-telar/context/execution-state.md`	Per-WU phase, retry count, validation results; doubles as the resume journal.
`.tl-telar/context/project-context.md`	Cross-WU coherence context read by later WUs.
`.tl-telar/project-profile.json`	Setup sentinel — framework + plugin version captured once at setup.
`.tl-telar/context/external-tools-budget.jsonl`	External-adapter spend ledger.
`.worktreeinclude`	Files copied into each WU worktree when isolation is active.

See also. The Orchestration deep-dive explains how these settings drive the pipeline, and how the cc_features gate chooses its path at runtime.