PR #238 — A Dispatch Habit, Not a Bad Default: Warning on Redundant --context-mode full
The Broader Goal at the End of the Previous PR
PR #236 (Measurement Ledger Hardening Phase 1, still open at the time) closed the loop on what a cost row says about itself — every write now carries explicit provenance. Reviewing the roadmap for what would strengthen that same "trust the numbers" theme turned up a smaller, adjacent waste: synlynk dispatch callers (including this project's own PM workflow) kept passing --context-mode full out of habit, even when a task prompt was already fully self-contained — exact code, exact file paths, exact commit message. Full-context dispatches inject the whole roadmap/memory/todo snapshot into the implementer's prompt, burning tokens the task didn't need.
Strategic Shifts in This PR (if any)
None — this and #202 (next post) were explicitly sequenced by the user ("take up #237 first, then #202") as two independent, additive pieces of the same v0.12.0 "Trust & Cost-Aware Routing" release theme, not a change in direction.
Worth noting what this PR is not: it's not fixing a bad CLI default. synlynk cli.py already defaults --context-mode to "task" — confirmed before writing any code. This is purely about catching an explicit, unnecessary override.
What This PR Shipped
_context_mode_hint(context_mode, task) — a pure, unit-testable heuristic in synlynk/dispatch.py. It only fires when context_mode == "full", and only when the task text shows two independent self-containedness signals at once: a code fence (```) and something that looks like an exact commit message (re.search(r"commit\s+-m|commit message", task, re.IGNORECASE)). Both signals together are a reasonable proxy for "this prompt doesn't need project background" — either alone (code with no commit message, or a commit message with no code) isn't a strong enough signal to warn on.
Non-blocking by design. The hint prints to stdout right after dispatch_agent() resolves context_mode:
hint = _context_mode_hint(context_mode, task)
if hint:
print(f" {hint}")
It never changes behavior or dispatch outcome — a caller who wants full context anyway (e.g. deliberately, for an agent that benefits from broader project awareness even on a spec'd task) is not stopped.
Tests target the pure function, not the side-effecting one. dispatch_agent() has heavy side effects (git worktree creation, subprocess spawning, preflight checks), so the six new tests in tests/test_dispatch_context_mode_hint.py exercise _context_mode_hint() directly: all four combinations of code-fence/commit-message presence at context_mode="full", plus confirmation the hint never fires at "task" or "none" even when both signals are present.
Brainstorm Visuals Used
None — pure logic fix, no visual/UI decisions.
Practicing What It Preaches
The dispatch prompt used to implement this fix (written to the scratchpad, then handed to codex via synlynk dispatch codex --context-mode task) was itself fully self-contained: exact diff blocks, exact test cases, exact commit message template. --context-mode task was chosen deliberately over full — the same discipline the fix now nudges other callers toward. Cost: 26,016 in / 6,504 out tokens (~$0.18), 161s, first-pass with no rework.
What This Achieved on the Path to Autonomy
A small, compounding token-efficiency win: every dispatch this hint successfully steers away from full context saves the implementer agent a roadmap/memory/todo injection it didn't need, on every future dispatch that pattern-matches. It's a much smaller piece of the "every number and every dispatch is deliberate, not habitual" theme than PR #236, but it's the same theme.
Strategic Note: The Goal at the End of This PR
Merged into main directly (squash merge, explicit user authorization). Next up in the same sequence: #202 (stall-killer status bug), covered in the next post. Both feed the still-open v0.12.0 "Trust & Cost-Aware Routing" release, which also depends on PR #236 landing.