TestReadFileRefusesNonRegular: read_file on a FIFO would block forever in open(2) waiting for a writer, leaking the tool goroutine past Ctrl+C (which

TestReadFileTruncatesOversizeContent: ReadFile obeys the same Truncate head+tail cap as every other tool. Oversize content must carry the marker so th

TestReasoningChunksAreEmitted: reasoning models stream chain-of-thought in `delta.reasoning`. The decoder must surface these as EventReasoning (else t

TestRedactSlashHidesHamrpassKey: with `logging: true`, every prompt (including `/hamrpass <key>`) is written to .codehamr/log.txt. The log is meant to

TestRedundantResizeIsNoOp: a WindowSizeMsg with unchanged dimensions (some terminals send these on focus) must not fire the hardening path; clearing t

TestReexecGuardOverridesPreexistingValue pins the loop-guard env semantics maybeSelfUpdate relies on: the re-exec'd child must see CODEHAMR_NO_UPDATE_

TestRepeatedFailureNudgeDifferentTargetResetsStreak: switching targets between failures resets the streak: exploration across distinct operations must

TestRepeatedFailureNudgeFiresOnceAfterFiveSameTargetFailures drives the repeated-failure backstop end-to-end. Five consecutive failures of the SAME ta

TestRepeatedFailureNudgeSuccessResetsStreak: a single success in the middle of a run of failures resets the streak, so the nudge never fires.

TestResetClearsChips: Reset empties store and spans, leaving only new text.

TestResizeKeepsPromptInsideTerminal: across a resize sequence View must never claim more rows than the terminal, else the textarea pushes the status b

TestResizeSettleReplaysScrollbackAtNewWidth: a matching settle returns a tea.Sequence that wipes viewport + scrollback and re-emits splash and the ful

TestResolvedKeyExpandsEnvVar: `key: ${MY_KEY}` in config.yaml must expand the env var at read time while the raw reference round-trips on Save. This i

TestResolvedKeyLiteralDollarSurvives: expansion applies ONLY when the whole key is a ${VAR} reference. A literal proxy key containing '$' (llama.cpp -

TestResolvedKeyPassesLiteralThrough: a plaintext key with no $-references must pass through unchanged so existing configs keep working.

TestResolvedKeyUnsetEnvYieldsEmpty: ${VAR} with VAR unset expands to "", matching the "no key" branch (keyless local Ollama) without a panic.

TestResponseReserveScales pins the reserve curve: floored until ctxSize/8 crosses 8k, then linear. A divisor tweak lands here loud.

TestRightArrowAtChipStartJumpsToEnd: → skips the whole chip in one keystroke.

TestRunRawSurfacesTruncatedToolArgs: the _parse_error sentinel (set by llm.resolve when the server truncates an oversized tool call mid-JSON) must sur

TestRunSlashPicksUpExternalConfigEdits: runSlash re-reads .codehamr/config.yaml before dispatching, so a profile a user hand-added to the file mid-ses

TestRunSlashWarnsOnBrokenConfig: a typo in config.yaml must not lock the user out of slash commands. The reload prints a one-line warning and keeps th

TestRunToolCallHonorsBashTimeoutBeyondLegacyCap: bash's own timeout is the only ceiling: runToolCall must not wrap the parent context in a shorter cap

TestRunawayNudgeFiresOnceAtMaxToolRounds: the per-turn tool-call counter trips exactly one soft system note when it reaches maxToolRounds, and never b

TestRunawayNudgeFiresWhenCounterSkipsCap: a multi-tool-call round increments toolRounds per call but the nudge is consulted only when the pending queu

TestSaveIsAtomicAndLeavesNoTemp: writeYAML writes a sibling temp then renames it over config.yaml so a torn write can't brick the next launch. Pin tha

TestSaveTightensPreexistingLoosePerms covers the upgrade path fresh-bootstrap misses: a config.yaml from an older world-readable codehamr (or a hand-e

TestSendEventUnblocksOnCancel pins sendEvent's anti-wedge invariant: once the parent context is cancelled, a send to an undrained channel must abort v

TestSessionTokensAccumulateAcrossTurns: the session counter is separate from the per-turn counter. finalizeTurn resets turnTokens; sessionTokens keeps

TestSessionTokensSurviveFinalizeTurn: finalizeTurn clears turnTokens but must NOT touch sessionTokens. turnStart must be set or finalizeTurn no-ops.

TestSetActivePersists: SetActive flips Active and writes config.yaml.

TestSetActiveRejectsUnknown: SetActive returns an error for an unknown name.

TestSetActiveRevertsOnSaveFailure guards in-memory/on-disk drift on Save failure. If SetActive mutates Active before a failed Save, ActiveProfile() re

TestSetCursorRuneOffsetCrossesSoftWrappedLine: walking to a target row must traverse a soft-wrapped logical line in between. bubbles' CursorUp inside

TestSetValueClearsChips: SetValue installs plain text and drops prior chips, used by slash-popover Tab-completion, where chips can't be part of a repl

TestSlashClearSurvivesBrokenConfig: with config.yaml unparseable, /clear must still wipe history. The reload warning gets wiped with the rest of scrol

TestSlashClearWipesTerminalScrollback: tea.ClearScreen (\x1b[2J) clears only the visible region; the saved-lines buffer needs the DECSED 3 sequence fr

TestSlashModelRejectsUnknown: unknown name is a quiet warn, not a switch.

TestSlashModelSwitchClearsStaleBudget: after a hamrpass turn leaves m.budget set, switching to a profile that emits no X-Budget-* headers (local Ollam

TestSlashModelSwitchDropsStickyFallbackState: llm.Client's noReasoningEffort flag ("this server 400'd on tools+reasoning_effort, stop sending it") is

TestSlashModelSwitchesActive: /models <name> sets Active and rebuilds the llm client's base URL / token / model to the new profile's values.

TestSmallPasteStaysInline: a paste below threshold stays raw, no chip.

TestSplashEmittedOnFirstSize: the splash prints once into scrollback on the first WindowSizeMsg, then scrolls up as content arrives. We verify the lin

TestStalePingForOldBackendDoesNotOverwriteConnectedFlag: a ping launched against the old profile's URL can land after the user /models'd to a new reac

TestStaleProbeDoesNotPrintActivationBannerForNonActiveProfile: a probe for a profile the user has /models'd away from must not print "✓ active: <profi

TestStaleProbeForOldProfileDoesNotOverwriteConnectedFlag mirrors the pingMsg guard for probeMsg: a probe for a no-longer-active profile must not mutat

TestStaleResizeSettleIsDiscarded: a settle whose gen no longer matches m.resizeGen (a newer resize bumped it after scheduling) must be a no-op, else t

TestStaleStreamCloseDoesNotKillLiveTurn: the prior turn's channel closing after a fresh submit must NOT run handleStreamClosed against the live turn.

TestStaleStreamEventDoesNotMutateLiveTurn: after Ctrl+C kills turn 1 and the user submits turn 2, turn 1's readEvent Cmd can still fire (buffered chan

TestStaleToolResultDoesNotEnterLiveHistory: a bash result from cancelled turn N must not append to turn N+1's history (its unmatched tool_call_id woul

TestStatusBarLiveTimerAndFrozenSummary: during an active turn the bar shows the phase label plus a ticking wall-clock; at idle after a clean finish it

TestStatusBarOmitsBudgetWithoutHeaders: endpoint sends no budget header, no pass segment appears.

TestStatusBarShowsBudgetFromHeaders: the pass segment renders whenever X-Budget-Remaining arrives: the header is the only signal, no profile gating. T

TestStatusBarShowsSessionTokens: once the counter is non-zero it appears in the status bar; at zero the bar stays quiet.

TestStatusBarShowsSpinnerWhenWaiting verifies the micro-animation text appears in the bottom bar while a request is in flight, and disappears when not

TestStreamContentShowsLiveInViewport: a mid-turn content event populates the streaming buffer, promotes phase thinking→streaming, and is visible in Vi

TestStreamingShownInLiveView: the streaming buffer is rendered above the prompt by View() while content is in flight, so the user sees tokens immediat

TestStrictYAMLRejectsUnknownKey: unknown top-level keys in config.yaml must fail loud, not be silently ignored: surfaces typos immediately.

TestSubmitRedactsHamrpassKeyFromHistoryAndScroll: redactSlash keeps the bearer token out of the debug log, but submit must also keep it out of scrollb

TestSubmitResetsFailureStreak: a fresh user submission is a new goal, so a stale streak from the previous goal must not carry over and trip the nudge

TestSystemPromptFitsFixedSystemReservation pins ctx.FixedSystem against the embedded prompt. Grow the prompt past the reservation and Pack over-alloca

TestSystemPromptIncludesWorkingDirAndInvestigateRule: the system prompt must (a) tell the model to investigate files itself rather than ask the user t

TestToWireAlwaysSendsContent: silent tool results (empty stdout) must still serialize "content":"". Ollama's /v1 shim 400s if the field is absent or n

TestToWireParseErrorArgsStayValidJSON: when resolve() stamps _parse_error for a truncated tool call and that assistant message round-trips into the ne

TestToolArgsStreamBumpsEstimateAndPhase: a tool-call argument fragment (a file streaming into write_file) ticks the live token estimate AND flips the

TestToolCallFlushesStreamedContent: a tool-call event ends the content phase. Whatever streamed before it is rendered and committed *now*, so the user

TestToolCallLeakWarningDetectsStrandedXML: a turn ending with leaked tool-call XML stranded in the newest assistant message warns the user; clean text

TestToolResultFailed pins the failure classifier the nudge keys on: a "(cancelled)" result (user Ctrl+C) is never a failure; write/edit fail iff the t

TestToolTargetKey pins the identity the repeated-failure nudge keys on: file tools key on tool+path, bash on tool + the trimmed first command line, el

TestTruncateSnapsToRuneBoundary: Truncate's byte-offset cut must not slice multi-byte runes mid-sequence: output stays valid UTF-8.

TestTurnEndsWhenAssistantEmitsNoToolCalls pins the turn-end contract: a final assistant message with no tool calls returns to phaseIdle and hands cont

TestTwoChipsTrackedIndependently: two pastes make two chips; deleting the first leaves the second intact with its span shifted left.

TestTypingIntoChipInteriorKeepsContent: typing a rune while the cursor sits inside a chip label, when two chips share an identical label, must not spl

TestTypingOverflowKeepsCursorVisible: once content overflows maxTextareaHeight, pre-grow must NOT pin the viewport to the top: that would scroll the c

TestTypingPreservesEarlyContent: a marker at the START of a line that wraps to many rows must stay visible in the textarea View. Guards a scroll bug:

TestTypingShiftsSpans: typing before a chip shifts its span right by one; reconcile tracks it automatically.

TestURLOverrideDoesNotPersist: a CODEHAMR_URL override lives in cfg.URLOverride and ActiveURL reflects it, but Save writes only the stored URL, so re-

TestUnqueueOnlyWhenTextareaEmpty: with a draft already in the textarea, Backspace must delete a character as usual and leave the queued slot untouched

TestUnqueueRestoresToTextarea: Backspace on an empty prompt while a turn runs pulls the queued prompt back into the textarea and clears the slot, a re

TestValueExpandsAllChips: Value() interleaves surrounding text with full paste contents in order.

TestVerboseLogCapturesTurnRecords drives a realistic two-round turn (reasoning → bash tool call → final answer) with logging on, and asserts the verbo

TestVerifyNudgeEndToEndRePromptsThenFinishes drives a full turn that does real work (8 bash rounds) then "finishes" with a confident summary. The fini

TestVerifyNudgeFiresOnceAtMinRounds: the finish re-grounding nudge trips one soft system note only once a turn has done real work (toolRounds >= verif

TestVerifyNudgeRePromptsSubstantialCleanFinish: a substantial turn ending with a clean, non-empty summary must be re-prompted once (phase back to thin

TestVerifyNudgeSkipsHonestUnverifiedFinish: a substantial turn whose summary already marks something `unverified` has done the honest self-assessment

TestVerifyNudgeSkipsTrivialTurn: a turn that did little work (toolRounds below the gate) must finish normally (finalized, no re-prompt, no system note

TestVerifyNudgeYieldsToEmptyReply: an empty newest assistant message is the empty-reply nudge's domain even on a substantial turn: the verify nudge mu

TestViewHandlesZeroWidth reproduces the "shrunk UI" startup flash: before WindowSizeMsg arrives, View must not panic or emit garbled layout.

TestVisualPromptLinesSumsAcrossLogicalLines: visualPromptLines splits on newlines and sums wrapRows for each segment: the prompt field grows to hold t

TestWidenResizeAlsoSuppresses: widening changes the splash layout (text→art at the wordmark threshold) and leaves narrow rows misplaced, so it takes t

TestWidthChangeSuppressesView: any width change (narrow OR widen) flips suppressView so nothing commits between SIGWINCH and the settle. View() return